Special Issue of ICAMC 2026 | International Journal of Science, Engineering and Technology

ICAMC Papers

Yield Enhancement And Prediction Of Crops Using Deep Learning

Authors: Abhishek Sharma, V K Jain, Vivek Kumar

Abstract: Agriculture all over the globe is becoming more challenging due to increasing populations, erratic climate conditions, and a scarcity of cultivable land. The ability to predict crop yields accurately is crucial for the development of management strategies for agriculture and for the guarantee of food security. Traditional statistical models oftentimes do not reflect the complex, nonlinear interactions among the factors of crop, soil, and environment. A very effective approach for precision agriculture is deep learning (DL), a branch of artificial intelligence that can model high-dimensional and unstructured data. A comprehensive review of the application of DL models in agriculture as a tool for the prediction and enhancement of production is the focus of this paper. The research indicates improved accuracy of forecasts for multiple crops by evaluation of various architectures, such as CNN, LSTM, and hybrid CNN-LSTM models with attention mechanisms. Moreover, the integration of climate factors, soil properties, and remote sensing data gives valuable insights into the application of yield improvement strategies. There is also discussion of future prospects and challenges, such as transparency, data limitation, and scalable implementation.

DOI: http://doi.org/10.5281/zenodo.21700618

Analyzing Decadal Cyber Crime Trends In India: A Data Driven Approach Using Clustering And Predictive Modeling

Authors: Manisha M. More, Mayuri More, Varsha Gidde, Sheetal Sandeep Patil

Abstract: The proposed study explores the NCRB i.e. National Crime Record Bureau data for the period of 2012 To 2022 and assessed the trends cybercrime across the states, the metropolitan cities, and the sectors such as IT and IPC. The analysis is conducted using Power BI for data visualization and statistical methods in Python looking for any trends or correlations. In addition, clustering techniques were utilized in grouping states and sector data observation into three groups, i.e., low risk, medium risk and high risk. The analysis showed that most states were classified as low risk and only a few instead showed an extremely high intensity of Cyber Crime suggesting a degree of geographical concentration for Cyber Crime. The authors also used two types of machine learning model (Linear Regression and Random Forest) to identify any trends based on time. However, both machine learning models displayed limited success in their predictability meaning that Cyber Crime occurrences cannot simply be explained by reference to time alone. The overall findings of this study then align with the conclusion that while some Cyber Crime occurrences are similar and concentrated to a particular location or point in time, they are also dissimilar based upon when they occur and where they occur. As a whole the study illustrates a methodical analytical technique of using both data visualization through statistical methods and explorative AI methods to provide better insight into Cyber Crime trends.

DOI: http://doi.org/10.5281/zenodo.21700719

AI-Powered Insurance Ecosystems: Risk Scorecards, Operational Efficiency, Financial Reporting And Compliance

Authors: Chetan Prakash Ratnawat

Abstract: The insurance sector is experiencing a considerable change that is Artificial intelligence (AI), data analytics, and automation. Past AI deployments mainly targeted separate insurance functions like underwriting, claims processing, and regulatory compliance that have resulted in hardly any benefit for the overall organization. This paper suggests an AI-driven insurance model that interconnects risk assessment, operational activities, financial reporting, and regulatory adherence into one well-organized multi-layered structure. After studying a variety of sources including peer-reviewed articles, industry standards, and regulatory documents, the findings indicated that AI has the biggest impact when it comes to low-risk, high-exposure tasks such as claims triaging, document processing, and workflow automation. On the other hand, tasks that require human expertise like the underwriting decision-making should not be done by AI alone. The results also show that AI integration at an ecosystem level enhances the decision consistency, operational efficiency, and governance transparency compared to isolated implementations. Besides, the research underscored the improvements that prior studies have been able to quantify – for instance, the faster processing of claims, better fraud detection, and more stringent compliance monitoring. This research adds to the body of InsurTech literature by advocating for the creation of a compliance-focused AI governance framework, which fully adheres to legal and ethical principles. The proposed model acts as a guide, balancing practicality with retaining transparency, accountability, and trust within insurance companies.

DOI: http://doi.org/10.5281/zenodo.21700751

Architecting Automated Resume Screening Systems

Authors: Anushka Rawat, Dr.Ajay Rana, Deepshikha Bhargava, Garima Panwar

Abstract: In the project, the aim is to create an automated system to help with resume screening. Typically, resumes do not come in a standard format. They vary in structure, writing style, and presentation. Therefore, processing resumes is far from easy compared to other texts. Unlike data in a neat and organized structure, as in academic data sets, processing a resume is a challenge. In academic data sets, the data is clean and in a neat organization. In order to manage these problems, certain textual processing techniques were applied, and the steps included normalization, tokenization, and feature selection. The system was tested numerous times in order to improve its ability to execute perfectly even with imperfect data. Over time, the scope extended beyond how to improve the precision of the program to include the influence of data quality in the real world. Aside from that, there are also ethical considerations involved in the making of the system, such as the implication that biases in the data might be inadvertently reflected in automated systems, where measures are being taken to make a more transparent and interpretable system, keeping recruiters involved by introducing flexible decision thresholds to keep human judgment in charge. Overall, however, it is a system meant to assist, not replace, the recruiter. In this manner, it can help speed up the hiring process and make it more efficient by reducing human error and quickly identifying potential candidates.

DOI: http://doi.org/10.5281/zenodo.21700793

Real-time Speech-to-Speech Translation System For Enhanced Multilingual Communication

Authors: Dolly Chauhan, Kritika Sharma

Abstract: This paper presents the development of a real-time speech-to-speech translator and its analysis in order to overcome the language barriers between two languages. Based on advanced deep learning models, i.e. the Whisper-Speech-to-Text (STT) and MMS-Text-to-Speech (TTS) APIs of OpenAI and Facebook, respectively, the system takes the audio input, identifies the language spoken, converts it to English text and thereby transforms this text to English speech. The proposed architecture will take into account powerful audio preprocessing, such as Voice Activity Detection (VAD) and noise elimination, to improve the accuracy and reliability of the translation pipeline. A dataset of 200 samples from the Mozilla Common Voice corpus was the one used for the system's validation, which clearly showed its 16.6% latency decrease and superior noise robustness over the standard cascaded API baselines. The accuracy of translation, system time, and the quality of synthesized speech performance metrics are tested in a variety of languages, which proves the efficiency of the system and the possibility of its advancement into the practical side of using the system in various communication settings. The prototyped system presents the effectiveness of composing AI models in providing a smooth on-the-fly language conversion.

DOI: http://doi.org/10.5281/zenodo.21700844

A Comprehensive Analysis On Ocean Observation Systems And AI Integrated Technological Implementations In Ocean Science

Authors: Abhendra Pratap Singh, Nandini Sharma, Arpit Dwivedi, Vansh Garg, Saksham Aggarwal, Aakriti Sharma, Vanshika Dua

Abstract: The integration of ocean observing networks and artificial intelligence is changing how and who access marine information. This review looks at technological and recent innovations along with uses from the Argo and Biogeochemical-Argo programs, satellite remote sensing, autonomous underwater vehicles, and cloud-native data architectures. This review assesses the integration of these systems with Artificial Intelligence, particularly large language models enhanced with vector-database retrieval systems and retrieval-augmented generation (RAG). This paper discusses the major hindrances to efficient ocean data utilization like data fragmentation, also throwing light on the intricate formats like NetCDF, scalability, and policy barriers. It highlights modern practices such as practical lakehouse architectures, cloud processing, and AI systems with semantic retrieval controls. Furthermore, the review describes and discusses AI prototypes and systems built from case studies designed to assist in disaster response and mitigation, monitor climate, and manage fisheries by transforming data on spectra, profiles, and images into concise answers and easy-to-understand visual response graphics. It also talks about the persisting questions of trust, explainability, provenance, multilingual support, and inclusive engagement at the community level as the foundation of new research. These systems anticipate the provision of standard metadata, uncertainty framing, AI systems, and ethically aligned deployment pathways. Integrating resilient observing systems with human centred AI and scalable infrastructure is essential to realize a transparent, operational, and community-driven ocean-observing ecosystem promoting further research in this domain.

DOI: http://doi.org/10.5281/zenodo.21701083

A Localized Framework For AI Implementation In Rural Indian Government Schools: Addressing Gaps In The National Education Policy (NEP) 2020

Authors: Jyoti, Akshita Tiwari, Divyansh Kajla, Aman Tomar

Abstract: Artificial Intelligence (AI) plays a major role in educational transformation in India, according to the National Education Policy 2020 (NEP 2020). Nonetheless, the accelerated rush to implement computing thinking and AI since Grade 3 poses a threat to increasing the rural-urban digital divide because of poor infrastructure, language differences, and inequities in rural government schools. Although most urban schools are enjoying the use of cloud-based and generative AI, many rural schools are still without reliable electricity, consistent internet connections, adequate hardware, and well-trained teachers. The paper suggests a Localized AI Implementation Framework to overcome the policy gaps in last-mile in rural Indian classrooms. The research follows a three-stage process via a Sequential Exploratory Mixed-Methods design (QUAL-quan) based on the introduction of socio-technological adoption theories and the national data used in the study, including ASER and UDISE+. Phase I will entail qualitative interviews with teachers, administrators, and stakeholders in the community in linguistically diverse districts to discover contextual barriers. Phase II measures these insights using big data surveys using UTAUT measures of technology acceptance. Phase III tests an offline-first Edge AI model on systems based on Raspberry Pi to test the infrastructural resilience, vernacular AI performance, teacher workload reduction, and student learning outcomes. It has three pillars, namely: (1) Infrastructure Resilience, through edge computing that makes it possible to use AI without constant internet connectivity; (2) Localized Multilingual Pedagogy, in which AI automates administrative work without reducing pedagogical authority of teachers; and (3) Human-Centric Teacher Empowerment, by which AI automates administrative work and does not decrease the pedagogical authority of teachers. The study hypothesizes that culturally resonant, offline AI systems will lead to better STEM learning, use fewer resources than cloud-dependent models, and that teacher time resources can be used to mentor students. It is anticipated to result in higher levels of digital equity, teacher efficiency, and quantifiable student engagement and performance in the rural setting. On the whole, the framework proposes decentralized and context-sensitive AI ecosystems that can match the ambitions of NEP 2020 with the context of underserved classrooms in rural areas. It provides policy-consistent and scalable ways of ensuring fair adoption of AI across the country.

DOI: http://doi.org/10.5281/zenodo.21701251

Football AI Analyzer: A Deep Learning System For Real-Time Football Match Analysis

Authors: Prajakta Musale, Nachiket Girnar, Palash Nikam, Pradip Pardhi, Prathmesh Parase, Naman Jain

Abstract: Football analytics has become an increasingly important tool for understanding team strategies and evaluating individual player performances. Traditional football analysis systems rely on expensive multi-camera setups and complex hardware requirements, limiting the ability of many analysts to perform detailed tactical analysis. This study presented the Football AI Analyzer (Version 2.0), an enhanced object detection system capable of analysing and producing tactical information from a single broadcast camera feed using YOLOv8. The methodology combined YOLOv8 for detecting players and the ball with ByteTrack and a Kalman filter for robust multi-object tracking across frames. Player positions were mapped onto a top-down view of the pitch using homography transformation. Voronoi tessellation was applied to measure spatial dominance, and Gaussian smoothing was used to generate individual player movement heatmaps. The system architecture was implemented using FastAPI for backend processing and WebSockets for real-time data streaming to a React-based dashboard. Results demonstrated that the system achieved a processing rate of 32 FPS and an IDF1 tracking score of 89.5% over a test dataset of 12,500 annotated frames. The Football AI Analyzer presents a realistic and cost-effective alternative to expensive proprietary tracking systems.

DOI: http://doi.org/10.5281/zenodo.21701314

Hybrid Machine Learning Based Integrated Network Scanning And Anomaly Framework

Authors: Sarthak Jain, Suyash, Upendra Kumar, Utsav Chauhan, Ashish Kumar

Abstract: Modern networks produced enormous volumes of data which made them increasingly vulnerable to advanced cyber threats. Traditional scanning tools and standalone anomaly detection systems fell short in identifying evolving or zero day attacks. This study proposed a hybrid framework that integrated active network scanning with machine learning based anomaly detection to deliver an adaptive and automated solution. The framework combined Nmap based host scanning tcpdump based traffic capture and Zeek based feature extraction with a Random Forest classifier and an autoencoder trained on normal traffic to flag deviations. Recent advancements in supervised and unsupervised anomaly detection together with integrated intrusion detection systems published between 2020 and 2025 were reviewed and synthesized to position the proposed approach within the broader field. Experimental comparison across five learning models showed that the Convolutional Neural Network achieved the highest accuracy of 93 percent followed by Long Short Term Memory at 91 percent while Random Forest balanced accuracy and interpretability at 89 percent. The hybrid correlation mechanism that combined scan derived signals with model predictions reduced false alarms and strengthened real time threat visibility compared with single method systems. The study concluded that combining active scanning with machine learning significantly improved detection accuracy reduced false positives and enabled actionable reporting for security analysts. Future directions identified included adaptive learning methods lightweight models suited for edge devices and secure distributed detection through federated learning.

DOI: http://doi.org/10.5281/zenodo.21701384

DEPLANT-Deep Learning For Plant Disease Prediction

Authors: Aditya Kumar Gupta, Pratishtha Srivastava, Swathy R

Abstract: Plant diseases significantly reduced agricultural productivity and directly affected farmers' income, while manual inspection methods for disease detection remained laborious, subjective, and difficult to scale. This paper presented DEPLANT, a deep learning based system developed to detect plant diseases from images of leaves. Three deep learning architectures, namely Convolutional Neural Network CNN, VGG16, and Residual Network ResNet, were implemented and compared for the classification of sixteen categories of healthy and diseased plant leaves using the PlantVillage dataset. The dataset consisted of 22079 RGB images, split into 80 percent for training and 20 percent for validation. Experimental results showed that the ResNet model achieved the highest validation accuracy of 87.34 percent, followed by VGG16 with 85.47 percent, while the baseline CNN model reached 92.11 percent training accuracy but only 63.62 percent validation accuracy, indicating overfitting. These results demonstrated that transfer learning based architectures generalized considerably better than a shallow CNN trained from scratch on a limited dataset. The DEPLANT system was designed for real time deployment through a Streamlit based web application, allowing farmers and agricultural experts to upload leaf images and obtain instant predictions. The proposed framework supports smart farming and precision agriculture by enabling faster, more accurate, and more accessible plant disease diagnosis compared with traditional manual inspection methods.

DOI: http://doi.org/10.5281/zenodo.21701434

Adaptive Sound Profiles: Using Machine Learning

Authors: Kiran Srivastava, Shreyansh Singh, Syed Shahzeb Ali, Uddhav Garg

Abstract: The personal audio devices often fall short because they can’t keep up with the constantly changing condition, we use them in. People end up adjusting the volume over and over-whether the background noise suddenly gets louder, the listening environment shifts, or their own habits change. All of this interrupts the experience and makes listening less enjoyable. This study explores how to fix that by introducing a fully autonomous, personalised audio system called adaptive sound profiles. At the core of this approach is a blend of machine-learning-based personalisation, real-time sound monitoring, and a hands-free gesture control system. It uses two microphones: one listens to what’s coming out of the speaker, and the other listens to what’s happening around the user. By comparing these two audio signals, the system can automatically adjust the volume so that the sound stays clear and comfortable, no matter what’s going on in the environment. The framework doesn’t stop at sound alone. It also considers factors like time of day and predicted user behaviour. A built-in Memory Component keeps track of each person’s listening preferences and applies them on its own, allowing the system to grow and adapt with the user over time. Instead of just reacting when the noise around you suddenly changes the system learns from what is happening in the environment and prepares in advance. The result is a smart, privacy-friendly and personalized way to control audio making listening feel smoother and much more enjoyable.

DOI: http://doi.org/10.5281/zenodo.21701470

Weakly Supervised Urban Heat Island Segmentation Using U-Net And Physics Guided Labels

Authors: Krishna Agarwal, Aman Shaw, Shwet Kashyap

Abstract: Urban Heat Island (UHI) is an established environmental phenomenon where urban environments experience substantially higher temperatures than their surrounding rural environments, which is directly attributed to the high density of infrastructure and minimal vegetation, along with anthropogenic heat emissions. Accurate identification of UHI regions is crucial for effective urban planning. Conventional UHI identification methods based on threshold value analysis are ineffective and do not account for spatial continuity. This study presented an effective and efficient framework that combined satellite data with physical models and deep learning to identify UHI regions. Landsat-8 thermal and optical data were employed to derive Land Surface Temperature (LST) and Normalized Difference Vegetation Index (NDVI) values. A rule-based approach was initially employed to generate weak UHI labels, which were then refined using a U-Net convolutional neural network (CNN). Experimental results across five diverse global cities demonstrate that the proposed method is highly effective for UHI identification. By leveraging physical regularization, it resolves the noise overfitting of traditional neural baselines—outperforming them in classification calibration (ROC-AUC and AUPRC) and cross-city generalizability—while entirely eliminating the need for manual pixel-level annotation.

DOI: http://doi.org/10.5281/zenodo.21701657

Blockchain Technology: Evaluation Across Security, Scalability, Privacy And Governance

Authors: Rudra Rai, Krishnakant Tripathi, Priya Pandey

Abstract: Blockchain is becoming an approachable data platform with several stakeholders having a shared history of transactions without being owned by an individual. The fundamental concepts of cryptographic hashing, peer-to-peer communication, and agreement protocols are sufficiently documented, and a lot of current research focuses on performance, security, privacy, and governance individually rather than as interacting dimensions. Recent blockchain research publications and deployments were structured based on a four-axis perspective that follows the dynamics of a system in terms of security, scalability, privacy, and governance. This lens was applied to consent mechanisms including proof-of-work, proof-of-stake, and practical Byzantine fault tolerance, and to techniques such as sharding and layer-2 that are designed to enhance throughput. Basic throughput calculations using reported block sizes, transaction sizes, and block intervals indicate that configuration limits are usually much larger than the transaction rates achieved in practice, implying that protocol overheads and network behaviour require a major share of the budget. A survey of attacks and defences indicates that increases in speed and programmability often expand the attack surface at the consensus and smart contract layers, which motivates the development of better analysis and monitoring tools. The results are applied to draw design insights for domains of finance, supply chains, healthcare, identity, and smart city platforms, and to highlight remaining problems in benchmarking and cross-chain coordination. A practical mapping of major blockchain platforms, namely Bitcoin, Ethereum, and Hyperledger Fabric, onto the four-axis framework was also provided to demonstrate its utility for platform comparison and selection.

DOI: http://doi.org/10.5281/zenodo.21701701

SEAScale: Predictive Serverless Autoscaling Architecture For High Demand Ride Sharing Platforms

Authors: Pradeep, Husain Zaidi, Sakshi Singh

DOI: http://doi.org/10.5281/zenodo.21701718

AI-Driven Crop Yield Forecasting And Adaptive Sowing Recommendations Using Temporal Fusion Transformers

Authors: Abinaya M, Kanchana B, Santhiya D, Gayathri M

Abstract: Agriculture is a key area of focus, as it is important for the global food supply, and millions of farmers around the world live off their agricultural products. Rainfall has become increasingly unpredictable. Temperature patterns have become increasingly variable. Climate patterns have shifted. Also, farmers often lacked access to modern climate prediction and advisory services. Erratic rainfall, temperature variations, and seasonality have caused climatic risk, a major source of uncertainty for all farmers, especially smallholders. Existing models have relied on static patterns of historical climatic data and failed to account for events over time. In addition to low productivity, revenue loss also occurred. In contrast to previous models that suggested strict recommendations based on past data and crop conditions, the proposed model provided data-driven and optimal recommendations, such as the best sowing time and the best crop for precision agriculture. Additionally, it was a scalable and reliable crop failure risk model and could improve productivity and sustainable climate-resilient agricultural practices.

DOI: http://doi.org/10.5281/zenodo.21701812

Transformer-Based Phishing URL Detection With Lightweight Context-Aware URL Encoding

Authors: Swati, Disha Vishwakarma, Anuradha Singh, Anusha Rathore

Abstract: The cybersecurity menace that is phishing attacks continues to thrive although online criminals are always intent on developing deceitful links that are exact replicas of genuine websites. Conventional machine learning (ML) methods were found to require use of human-coded lexical features which often fail in detecting complex subtleties of such forms of deception as, for example, homoglyph attacks, coded URLs or dynamically changing redirections. The authors of the research presented a new lightweight technique based on transformers for detecting phishing URLs, which uses context-sensitive URL tokenization and the self-attention methods, enabling detecting both structural features and semantic meaning of the URLs without the need for manual feature extraction. The authors employed the following methods: subword tokenization, transformer-based learning of the meanings of the words in context and some changes relying on adversarial thinking which helped achieve all of its goals, including better performance in capturing previously unknown forms of phishing. The authors evaluated their system on the publicly available datasets, and their results were comparable to those produced by people on regular systems, which indicates the high efficiency of the proposed technology. The conducted research proves that new transformer techniques are better at overcoming modern obfuscation techniques than conventional methods.

DOI: http://doi.org/10.5281/zenodo.21701841

A Review On The Role Of Artificial Intelligence In Enhancing EFL Learning : Perspectives From Higher Education

Authors: Abhendra Pratap Singh, Nandini Sharma, Saksham Goyal, Yatharth, Bhavya Pratap

Abstract: The rapid development of artificial intelligence (AI) into English as a foreign language (EFL) education has sparked a complete shift towards personalized student driven education. This review paper gives a systematic view of the evolving landscape of AI working tools like Intelligent Tutoring Systems (ITS) and AI powered writing to evaluate generative AI and speech recognition with their specific roles for enhancing language ability. By looking at the latest research and technological development together, the study evaluates how AI was improving language learning by giving personalized instructions and instant feedback and reducing anxiety. The paper analyses that technologies make learning more practical and relevant and also accessible for everyone. The paper also reviewed some critical concerns regarding algorithmic bias, unavailability to access digital technologies due to internet shortage and the necessity of maintaining human-centric design. Hence the paper shows that AI is future oriented which emphasizes the teaching model which is humanly friendly. Overall, AI is very helpful in language learning, but only if it is perfectly designed and is used ethically. It must be used carefully to ensure a better understanding and a positive experience.

DOI: http://doi.org/10.5281/zenodo.21701904

Adaptive Learning System Using Machine Learning For Personalized Education

Authors: Azmi, Ayush Vashistha, Chouhan Kumar Rath

Abstract: Online learning platforms have expanded rapidly in recent years; however, most existing systems still follow a uniform content delivery approach that fails to address individ ual learning differ-ences. Variations in learning pace, topic-wise understanding, and response behavior often result in reduced en gagement and ineffective learning outcomes. This paper presents an AI-based adaptive learning framework that personalizes educational resources by analyzing learner assessment data. The system integrates unsupervised clustering and supervised classification methods to catego-rize learners into three proficiency levels—Beginner, Intermediate, and Advanced—using features such as accuracy, response time, topic-wise performance, attempt frequency, and improvement rate. Ensemble learning techniques are employed to improve the accuracy of predictions and cap-ture complex learning patterns. Based on predicted proficiency levels and identified weak areas, a hybrid recommendation mechanism dynamically adjusts question difficulty and provides personal ized learning materials, including instructional videos, reading content, and adaptive practice ex-ercises. Experimental evalu ation using standard classification metrics indicates increased learner engagement and measurable performance improvement following adaptive recommendations. The proposed framework offers a scalable and data-driven solution for personalized online education through constant observation and adaptive updates to support effective knowledge acquisition and long-term retention.

DOI: http://doi.org/10.5281/zenodo.21701942

Solar Power Generation Prediction And Optimal Site Selection Using Machine Learning And Geospatial Data

Authors: P. Vibin Sri Balaji, Srinivasan R

Abstract: The global transition towards renewable energy lacks intelligent frameworks for solar plant locations optimization and long-term energy generation forecasting. This research presents a comprehensive AI based methodology for identifying optimal SOLAR plant installation sites and forecasting energy output for a period of 10-year. The proposed methodology integrates satellite-derived meteorological data from NASA POWER database, photovoltaic performance modelling, multi-parameter feature engineering, Random Forest regression, and suitability-based spatial ranking. Annual energy generation is computed in MWh using physical PV system parameters including panel area, efficiency, and performance ratio. A generalized and geographically adaptable framework is developed to enable scalable renewable energy planning. The initial Experimentation projects high predictive accuracy (R² ≈ 0.91) and low error margins, validating the effectiveness of the approach. The framework supports strategic energy infrastructure planning, investment decision-making, and sustainable policy development.

DOI: http://doi.org/10.5281/zenodo.21702488

Ohmega: A Composable Miniature AI Framework For Drag And Drop Model Assembly

Authors: Gurpreet Kaur, Savita, Nishu, Pooja, Aradhay Gupta

Abstract: We propose Ohmega, a composable AI framework that enables developers and end users to assemble pre-built miniature AI modules (micro-agents) via a drag-and-drop interface to construct larger, task-specific systems. These miniature AIs are small, focused components that each implement a single capability, such as text classification, entity extraction, arithmetic computation, API invocation, or simple policy decision-making. By treating such modules as first-class, reusable building blocks, Ohmega allows users to configure and reconfigure workflows without retraining large, monolithic models. The framework supports using modules in isolation or composing them in parallel or in sequence, with optional arbitration and ensemble strategies to aggregate or select outputs when multiple modules contribute to a decision. A central orchestrator manages data flow, enforces composition rules, and records execution traces, promoting transparency, debuggability, and safer behaviour. This design is intended to reduce the time, cost, and specialised expertise required to build reliable AI applications, while also making system behaviour more interpretable through explicit, inspectable graphs of interconnected micro-agents. Ohmega is particularly suited to multimodal AI, multi-agent systems, and LLM-centric applications in no-code or low-code settings, where domain experts may wish to prototype, adapt, and govern AI pipelines directly. By combining modularity, visual composition, and principled orchestration, the framework aims to accelerate prototyping and deployment of practical AI systems, encourage reuse of well-tested components, and provide a foundation for future work on automated composition, safety guarantees, and marketplace ecosystems for miniature AIs.

DOI: http://doi.org/10.5281/zenodo.21702525

The Role Of Natural Language Processing In Transforming Education: A Theoretical Perspective

Authors: Heena Yadav, Rakesh Verma, Seema

Abstract: The past decade witnessed remarkable growth in Artificial Intelligence (AI) research, and Natural Language Processing (NLP) emerged as a particularly prominent branch, focused on interpreting human communication in text and speech. NLP brought educational technology new possibilities, including the ability to grade student essays, provide feedback on mistakes, and guide the learning process through intelligent tutoring in real time. The present paper examined these possibilities from a theoretical standpoint and emphasised the importance of meaningful integration of NLP into contemporary educational systems. The study focused on developments in computational linguistics, deep learning, and transformer-based language models. It explored their applications in automated assessment, intelligent tutoring, sentiment analysis of learner feedback, and personalised instruction. A conceptual framework for the pipeline of linguistic processing, machine learning, and pedagogical feedback was introduced to explain how learner-generated text moves through the pipeline to foster adaptive learning. Despite the promising potential of these technologies, certain challenges were identified that require researchers' attention, namely the fairness of algorithmic assessment, the multilingual adaptability of systems, and the ethics of working with student data. Overall, the study's findings indicate the great promise of NLP as an integral part of future intelligent educational systems, provided its development adheres to rigorous technical and educational principles.

DOI: http://doi.org/10.5281/zenodo.21702569

A Comprehensive Study On Machine Learning Driven Smart Water Conservation And Management Device

Authors: Abhendra Pratap Singh, Arpit Dwivedi, Shree Bhagwan, Akash Yadav, Shashank Pandey, Nandini Sharma

Abstract: Water scarcity has become one of the leading global problems because of rapid urbanization, population growth, climate change, and increasing need of water for industry purposes. The traditional water management systems usually cannot ensure continuous monitoring, timely detection of leaks, and optimal use of water resources. Therefore, considerable water losses may occur. This review aimed at considering the development of intelligent water management systems which use the technologies of IoT and machine learning. The recent researches were considered and their results were used to estimate the application of the technologies in the areas such as water monitoring, water leak detection, irrigation management, water demand forecasting, water quality assessment, and resource optimization in the context of domestic, agricultural and urban sectors. According to the reviewed literature, the IoT technologies helped to improve real-time data collection and system monitoring while machine learning increased the level of predictive analysis, anomaly detection, automated decision making, and operational efficiency. Moreover, the integration of the mentioned technologies made it possible to decrease water losses and increase sustainability of water resource management. Nevertheless, there are still some limitations in the widespread implementation of those technologies. They include such factors as high cost of installation, insufficient infrastructure, poor data quality, cybersecurity risks, lack of interoperability, and scalability difficulties. Generally, the current review demonstrated the significance of development of the efficient, reliable and affordable smart water management systems capable of providing real-time monitoring, predictive analysis and automation. Solving all the mentioned technical, economical and infrastructural problems will be crucial for water conservation.

DOI: http://doi.org/10.5281/zenodo.21702807

Analysis of Deep Learning Architectures for Smart Plant Disease Detection and Organic Remedy Recommendation

Authors: Pradeep, Usha Dhankar, Aditya Kumar

Abstract: Plant diseases pose a major threat to global agricultural productivity, food security, and sustainability. Traditional disease diagnosis methods are often time-consuming, subjective, and require expert intervention. Recent advancements in artificial intelligence, particularly deep learning, have revolutionized agricultural disease detection through automated image-based systems. This study presents a comprehensive analysis of deep learning architectures such as Convolutional Neural Networks (CNNs), Transfer Learning Models (ResNet, VGG, Inception), and Vision Transformers (ViTs) for smart plant disease detection and integrates an organic remedy recommendation framework. The research employs publicly available plant disease datasets to train and compare model performances in terms of accuracy, precision, and recall. Furthermore, an ontology-based organic treatment recommender system is developed to suggest sustainable, eco-friendly remedies. The experimental results demonstrate that deep learning models achieve over 97% detection accuracy, while the proposed organic recommendation layer enhances decision-making for sustainable agriculture.

DOI: http://doi.org/10.5281/zenodo.21702860

Towards Sustainable Agriculture: A Hybrid Deep Learning With Super Resolution For Grape Leaf Disease Detection From Low Resolution Image

Authors: Rachit Khandelwal, Paresh Jain

Abstract: Grapes are small, sweet, and versatile fruits that come in various colors and are not only enjoyed fresh but also used for making wine, raisins, and a variety of culinary products. A healthy grape harvest requires prompt identification of grape disease and taking appropriate measures to stop it from spreading. Accurate disease identification is now possible with the help of recent research advancements. However, these techniques yield low accuracy and perform poorly with low-resolution images. A novel framework incorporating TSRNet model based super resolution followed by DataLiteViT classification model is suggested as a solution to this problem. For scale factors of ×2, ×4, and ×6, respectively, the utilised lightweight TSRNet model with 2.25 M parameters can reach 30.12, 28.38, and 27.57dB Peak Signal Noise Ratio (PSNR) values and Structural Similarity Index Measure (SSIM) value of 0.913, 0.841, and 0.756. The proposed DataLiteViT model achieved accuracies of 97.90%, 97.12%, and 96.69% for low resolution images with scale factors of × 2, × 4, and × 6, respectively. Furthermore, the employed method utilizing TSRNet model for super-resolution improved DataLiteViT’s classification accuracies of 99%, 98.71%, and 98.28% for LR images with scale factors of ×2, ×4, and ×6, demonstrating an efficient and highly effective solution.

DOI: http://doi.org/10.5281/zenodo.21702891

Cognitive Offloading And The Atrophy Of Generative Thinking In Frequent LLM Users

Authors: Akshat Kandpal, Gurpreet Kaur, Janhvi Gupta, Satyam Kumar Roy, Mukund Keshav

Abstract: The use of large language models became widespread to complete high-level generative work including brainstorming, drafting, and constructing arguments. This paper introduced a theoretical model called Generative Atrophy through Offloading (GATO) which posited that using these types of generative tools for repeated periods of time caused a gradual decrease in one's ability to generate independently. GATO drew from existing research within cognitive offloading theory, neuroimaging research into how cognition works while engaging in creative activities with technology assistance, and controlled research studies examining the effects of technology on creativity. Empirical evidence supporting GATO included four mechanisms: decreased ambiguity tolerance, atrophy of structural synthesis processes, contraction of originality, and decreased generative self-efficacy. Three large datasets were analyzed demonstrating that generative offloading occurred frequently and consistently across multiple platforms. A cognitive resilience framework was also developed based on empirical research.

DOI: http://doi.org/10.5281/zenodo.21702935

Smart Waste Segregation Monitoring System For Urban Local Bodies

Authors: Mamta, Roshan Kumar, Bhavishya, Peassa Prithvi

Abstract: Rapid urbanization has made the amount of municipal solid waste much larger, which makes it much harder to sort and manage. Traditional methods heavily depend on manual processes, which are slow and often make mistakes. This paper describes a Smart Waste Segregation Monitoring System that uses the Internet of Things (IoT) to automatically sort waste into dry and wet categories while keeping an eye on the fill level of bins in real time. The system combines sensors, microcontrollers and a cloud-based dashboard to make it possible to automatically sort items and keep an eye on them from a web interface. The system performance is analyzed in terms of sensor limitations, communication reliability, and scalability, highlighting challenges in handling composite waste. The proposed system helps with urban city projects and makes urban waste management more efficient by improving the waste in sorted order and how it is picked up.

DOI: http://doi.org/10.5281/zenodo.21703101

A Secure Real Time Cryptocurrency Portfolio Management System Using Blockchain And Predictive Analytics

Authors: Gurpreet Kaur, Mayank Malik, Harshita, Rachit Sharma

Abstract: The rapid growth of cryptocurrency markets has increased the demand for secure and intelligent portfolio management solutions. Most of the current systems present in the market mainly provide basic functions such as tracking assets purchased, monitoring prices of the assets and displaying simple price charts, but they lack deeper analytical capabilities and user – centric design. This study proposed an improved and user – friendly cryptocurrency portfolio management system that incorporates real – time data handling, secure authentication methods for the users and intelligent decision support for overall user-friendly experience. The system leveraged modern web technologies and API integration to access ongoing market data and generate accurate portfolio analysis. Additionally, the proposed model included predictive analysis and risk evaluation features to support better investment decisions. Special attention was given to making the system easy to use for beginners through a simplified interface and well – guided functionalities. Overall, the proposed system overcame limitations of modern tools by delivering a secure, scalable and intelligent platform for managing digital assets effectively.

DOI: http://doi.org/10.5281/zenodo.21703238

Blockchain-Enhanced LSTM Based Machine Learning For Real-Time IoT Security: Anomaly Detection And Trust Management At The Edge

Authors: Harisankar R Nair, Duane Chembakassery, Prassanna J

Abstract: The rapid proliferation of Internet of Things (IoT) devices has introduced significant security challenges, including unauthorized access, data breaches, and distributed denial-of-service (DDoS) attacks. Traditional security mechanisms often fall short due to the resource constraints of IoT devices and the evolving nature of cyber threats. This paper presents a novel Machine Learning (ML)-powered anomaly detection system, integrated with Blockchain-Based Trust Management, to enhance IoT security. The proposed system utilizes a Raspberry Pi 4 as an edge device to monitor real-time network traffic, extract relevant features, and employ an Autoencoder/LSTM-based ML model for anomaly detection. When an anomaly is detected, the system triggers an alert, mitigates threats using firewall rules, and logs security events on a private blockchain. A trust score mechanism is implemented via smart contracts to dynamically assess device behavior, enabling automated device quarantine and adaptive security measures. Experimental evaluations demonstrate that the LSTM model achieves an accuracy of 98.85%, with high precision and recall, effectively identifying network anomalies. The blockchain-based trust management framework ensures tamper-proof security event logging while dynamically adjusting device trust scores based on anomaly detection results. The results indicate that this integrated approach enhances both detection accuracy and accountability in IoT networks, making it a viable solution for real-time, decentralized IoT security.

DOI: http://doi.org/10.5281/zenodo.21703358

AI Driven Scrap Classification And Price Prediction Using CNN And Random Forest Regression For Circular Economy Applications

Authors: Gurpreet Kaur, Firoz Ansari, Angad Gupta

Abstract: The increasing generation of household waste and the informal nature of scrap trading have created challenges related to price transparency, material classification, and efficient reuse of recyclable resources. This study proposed an artificial intelligence-driven digital platform that integrated scrap trading, second-hand product resale, automated scrap classification, price prediction, and dealer matching within a unified ecosystem. A Convolutional Neural Network was employed to classify scrap images into plastic, metal, paper, glass, and electronic waste categories. Random Forest Regression was applied to predict scrap prices using material type, weight, volume, market demand, and prevailing market rates. A location-based matching mechanism was also incorporated to connect users with suitable nearby scrap dealers. The experimental evaluation showed that the proposed classification model achieved an accuracy of 94.2 percent, while the price prediction model obtained a Mean Absolute Error of 2.1, a Root Mean Square Error of 3.4, and an R squared score of 0.91. The findings indicated that the integrated approach improved scrap identification, pricing transparency, and transaction efficiency. The proposed platform provides a scalable digital approach that supports material reuse, sustainable waste management, and circular economy practices.

DOI: http://doi.org/10.5281/zenodo.21703475

Heart Disease Diagnosis Using Agentic AI

Authors: Aditi Chauhan, Shruti Mishra, Mr. Neeraj Tantubay

Abstract: Every year, a significant number of deaths worldwide are caused by heart disease. This study proposed an independent multi-agent system for early diagnosis that can work efficiently even when patient data arrive slowly and are usually incomplete in real clinical setting. The proposed system does not rely at all on language learning tools, unlike traditional machine learning models which require access to all (pre-trained) data or large language model tools and their inherent trustworthiness and reproducibility issues. The system depicted diagnosis as a series of small, adaptive decisions by six specialised agents. Using a conformal prediction reliable quantification of uncertainity and an active feature selection approach based on the short term bandit framework that balanced expected uncertainty reduction vs. cost per clinical test. Agents iteratively retrieved only the missing features most relevant to the task, terminated once sufficient confidence was achieved, and produced intelligible rationales with clinical suggestions. This research tested the system on UCI Cleveland heart disease dataset. It had an AUROC of 0.979, accuracy of 90.0%, precision of 100%, recall of 83.3% and F1-score ratio of 0.909 whilst using on average far fewer features than static baselines. We demonstrated successful predictive performance, decreased the cost of diagnostics, and improved explainability. This work proposes a simple, transparent and cost-sensitive framework that is transferable for real-world healthcare deployment in low-resourced settings.

DOI: http://doi.org/10.5281/zenodo.21703581

Real Time Risk Aware Route Optimization For Safe And Intelligent Navigation Systems

Authors: Mamta, Shaivya Grover, Palak Chugh, Mishthi Jain

Abstract: Modern navigation systems primarily optimize routes based on travel time and distance, often neglecting safety related parameters such as road condition, weather, visibility, and traffic hazards. This paper proposes a real time risk aware route optimization framework for safe and intelligent navigation systems that integrates multi source data including GPS, weather APIs, traffic information, and user generated hazard reports. A weighted risk evaluation model is employed to compute a composite safety score for each candidate route, and a modified Dijkstra’s algorithm is utilized to identify the safest route rather than merely the shortest path. The system dynamically updates route recommendations in response to changing environmental and traffic conditions, thereby enhancing adaptability and reliability. The proposed approach aims to improve road safety, reduce accident risks, and contribute toward the development of intelligent transportation systems.

DOI: http://doi.org/10.5281/zenodo.21714118

Comparative Analysis Of Explainable AI Techniques For Interpreting Machine Learning Models

Authors: Richa Sharma, Archita Kumari, Archana Mongia

Abstract: Breast cancer diagnostic machine learning models need to be not only accurate but also interpretable to be deployed clinically with any reasonable degree of confidence. Although Random Forest, XGBoost and Neural Networks generate accurate predictions, their respective "black box" nature of operation can undermine the clinical community's willingness to trust the output. Explainable AI supplement methods such as SHAP and LIME have been developed in part to address this issue. Herein, we compare the SHAP & LIME measured explainability consistency of three different classifying algorithms (Random Forest, XGBoost, and Neural Network) as they were trained on the Breast Cancer Wisconsin dataset. Each algorithm's accuracy, F1, and Receiver Operating Characteristic Area Under Curve (ROC-AUC) were calculated to evaluate model classifying performance. Spearman rank correlation was calculated to evaluate explainability consistency as assessed by SHAP & LIME. Neural Networks exhibited the highest classifying accuracy (96.49%) and F1-score (0.9718); however, XGBoost provided the greatest consistency between SHAP & LIME (ρ = 0.9534), followed by Neural Network (ρ = 0.9049) and Random Forest (ρ = 0.8889). XGBoost provides the best overall balance of predictive accuracy with explainability consistency.

DOI: http://doi.org/10.5281/zenodo.21715616

Well-Architected Framework For Agentic AI

Authors: Sudhir Kumar Kakumanu, Mukesh Shukla, Srushti Anil Patil

Abstract: AI agents are changing how enterprises solve problems. While traditional AI models addressed specific tasks like classification or prediction, AI agents offer goal-driven, autonomous capabilities—sensing, reasoning, planning, and acting. These data-driven autonomous systems depend on high-quality, well-governed data to reason, ground decisions, and deliver reliable outcomes. However, deploying AI agents in complex enterprise environments is difficult. This paper presents the Well-Architected Framework (WAF) for AI Agents and an executable WAF Auditor Agent. The framework defines nine pillars—spanning data foundation, security, governance, and beyond—from Agent Complexity to Sustainability, while the Auditor Agent turns these guidelines into an automated assessment workflow. This allows organizations to evaluate architectural designs and receive recommendations before and during development, providing a clear path to production readiness.

DOI: http://doi.org/10.5281/zenodo.21715658

Deep Learning-Based Automated Tuberculosis Detection Using Chest X-Ray Images

Authors: Nithin, Akhil Kumar, Shivani Sharma, Lalit Verma

Abstract: Despite ongoing advances in medicine, Tuberculosis remains a critical health problem, especially in developing nations where the lack of appropriate diagnostic aids is prevalent [15]. Timely detection of the condition is key to successful management; however, visual analysis of chest radiographs can be laborious and subjective among specialists. Thus, this research proposes a machine learning framework for identifying TB from chest X-rays. The methodology entails applying transfer learning on a pre-trained Convolutional Neural Network (CNN) to categorize the input images into TB-positive and healthy subjects [1]. Public datasets will serve as the source of data for training and testing purposes, combined with preprocessing and data augmentation strategies to boost predictive power [10]. The efficacy of the algorithm will be evaluated by accuracy, precision, recall, and F1-score measurements. Moreover, the interpretability of the algorithm is enhanced using Grad-CAM visualization to pinpoint significant lung areas affecting the classifier's decision-making process [7].

DOI: http://doi.org/10.5281/zenodo.21715691

AI In Creative Writing: Partner Or Rival Of Human Creativity

Authors: Seema Gupta, Gauri Kathait, Rachit Sharma, Udhav Bhardwaj

Abstract: Artificial Intelligence (AI) is transforming the landscape of creative writing, offering both opportunities and challenges for writers, educators, and readers alike. This paper explores the dual role of AI in creative writing—as a collaborative partner that enhances human creativity and as a potential rival that challenges traditional notions of originality and authorship. Through a systematic literature review combined with qualitative analysis of case studies and educational implementations, this study examines how AI influences storytelling, poetry, and narrative design. The research identifies a critical gap in understanding the practical dynamics of human-AI collaboration in creative contexts and proposes a partnership-rivalry framework to characterize AI's multifaceted role. Furthermore, the paper discusses ethical considerations, societal implications, and future directions for AI-human collaboration in the literary world. Findings suggest that while AI cannot replace the nuanced emotional depth and imaginative capacity of human writers, it serves as a powerful augmentation tool to inspire, refine, and expand the creative process. The study concludes that the relationship between AI and human creativity is not adversarial but complementary, pointing toward a future where collaboration between humans and machines defines new forms of literary expression.

DOI: http://doi.org/10.5281/zenodo.21715729

Comparative Analysis Of Machine Learning Models For Heart Disease Prediction Using Clinical Data

Authors: Lalit Verma, Manjot Kaur, Shivani Sharma, Siya Thakur

Abstract: Cardiovascular disease is one of the leading causes of mortality worldwide, which makes early and accurate prediction essential for timely clinical intervention. This study conducted a comparative analysis of six machine learning models, namely Random Forest, Gradient Boosting, Naive Bayes, Support Vector Machine (SVM), K-Nearest Neighbors (KNN), and Light Gradient Boosting Machine (LightGBM), for the prediction of heart disease from clinical data. Each model was trained and evaluated on a clinical dataset comprising established cardiovascular risk factors, and performance was assessed using accuracy, precision, recall, F1-score, and confusion matrix analysis. Naive Bayes achieved the highest accuracy at 90.74 percent, followed by SVM at 88.88 percent and LightGBM at 83.33 percent. Random Forest, KNN, and Gradient Boosting achieved accuracies of 79.62 percent, 81.48 percent, and 77.77 percent respectively. The results show that probabilistic classifiers performed particularly well on the relatively small clinical dataset used in this study, whereas ensemble tree based methods required larger training samples to fully exploit their variance reduction capability. Naive Bayes also produced the best recall and F1-score, while SVM achieved the highest precision. These findings indicate that model selection for heart disease prediction should be guided by dataset characteristics and by the clinical priorities of the intended application rather than by model complexity alone. The comparative framework developed in this study provides a practical basis for selecting machine learning models suited to cardiovascular decision support systems and highlights directions for future work involving larger datasets and deep learning approaches.

DOI: http://doi.org/10.5281/zenodo.21715760

A Hybrid RSA And ML-KEM Framework For Secure And Quantum-Resilient Communication

Authors: Karishma Dobhal, Y.P. Raiwani

Abstract: Classical public-key cryptography, mainly RSA, is at risk from the emergence of quantum computers. A complete shift to Post-Quantum Cryptography (PQC) faces vulnerabilities due to implementation and trust issues, despite NIST's approval of standards such as ML-KEM (Kyber). This paper presented a performance and bandwidth trade-off study of a proposed Hybrid Key Encapsulation Mechanism that fused the NIST-standardized ML-KEM-768 (FIPS 203) with the traditional RSA-3072. Two separate secret keys S1 and S2 were produced and encapsulated utilizing both classical and post-quantum public keys in our dual-layered model. In order to create a strong session key that ensures trust transition and backward compatibility with current network infrastructures, these secrets were then combined using a Hybrid Key Derivation Function (HKDF). In comparison to standalone systems, the study assessed the computational cost and communication latency brought about by this hybrid overhead. According to experimental data, the hybrid architecture offers a crucial security fail-safe against both classical flaws and future quantum threats without prohibitively high performance, even though it increases ciphertext size.

DOI: http://doi.org/10.5281/zenodo.21715795

A Review Of Artificial Intelligence Applications In Vehicle Safety And Driver Assistance Systems

Authors: Uma Gautam, Ankush Rana, Abhendra Pratap Singh, Vijayalaxmi, Vansh Garg, Charvi

Abstract: This paper summarizes advances in intelligent technology to enhance road safety with an emphasis on increased use of systems to detect car accidents. It reviews previous studies involving image processing, sensor-based systems and the combination of artificial intelligence and deep learning (AI with DL). It was proven that the latest systems use combination of sensors (like accelerometer, GPS, video, etc.) and Advanced Drivers Assistance Systems (ADAS) for better detection and less errors. This study shows that AI and DL have the ability to improve the speed and quality of processing data in accident scenarios. It also addresses data, system reliability, and practical aspects to improve the effectiveness and safety of intelligent accident detection systems in evolving road transport technologies. It also identifies research gaps and provides the basis for subsequent studies.

DOI: http://doi.org/10.5281/zenodo.21715838

Does Cognitive Offloading To AI Systems Influence Metacognitive Awareness And Critical Thinking Ability

Authors: Medha Yadav, Yash Kapoor, Aditya Arora, Ankush Rana, Priyanka Yadav

Abstract: This study explores how dependence on Artificial Intelligence (AI) affects critical thinking, with AI Dependency and Metacognition acting as key mediating factors. In today’s world AI tools are becoming more common in educational and professional environments such as teaching and in offices as well , leading individuals to depend on AI for a greater number of cognitive tasks. Although this trend simplifies many angles of life, it also raises concerns about a possible decline in our ability to think independently. The paper reviews existing research to assess how AI dependence impacts cognitive engagement and decision-making processes. While AI can enhance productivity and learning, a very excessive reliance on it may lead to reduced cognitive demands and hinder the development of critical thinking abilities. It is also suggested that metacognition plays a vital role in critical thinking because of its function in overseeing, assessing, monitoring and regulating cognitive processes such as problem solving, learning etc. A conceptual model is presented that includes Metacognition, Cognitive Offloading, AI Dependency, and Critical Thinking. This framework demonstrates both direct and indirect connections between Cognitive offloading and Critical thinking. It indicates that metacognition has a positive effect on reflective and analytical thinking. These findings highlight the importance of purposeful usage of AI and the cultivation of metacognitive skills to improve critical thinking in this digital age.

DOI: http://doi.org/10.5281/zenodo.21715894

SybilBelief: A Semi-Supervised Markov Random Field Approach For Structure-Based Sybil Detection

Authors: Rishu Kumari, Ridhi Jain, Sanidhya Pal, Nitesh Kaushik, Meenakshi Sharma

Abstract: Online social networks face persistent threats from Sybil attacks in which a single adversary creates many fake identities to manipulate trust ratings, spread misinformation, or claim advertising rewards. This study presents SybilBelief, a semi-supervised framework that treats Sybil detection as probabilistic inference over a Markov random field built from the network graph. The method propagates belief through loopy belief propagation while learning a single homophily parameter θ from a small set of labeled seed nodes through pseudo-likelihood maximization. To prevent numerical underflow during high-degree message products, all message updates are performed in the log domain using the log-sum-exp trick. The edge parameter is optimized via gradient ascent on a pseudo-likelihood objective, implemented through PyTorch’s automatic differentiation applied to an unrolled 50-iteration belief propagation graph, with gradient checkpointing employed to manage memory. Rather than relying on a fixed global threshold for classification, thresholds are tuned per dataset via nested validation, and the Area Under the Precision-Recall Curve (AUPRC) is reported as the primary threshold-invariant metric. Experiments on four datasets — a synthetic Barabási-Albert graph, Epinions, Slashdot Zoo, and a Twitter sample — demonstrate that SybilBelief significantly outperforms SybilRank and SybilLimit under severe label scarcity (0.5% seeds; p < 0.01 in all cases), degrades more gradually under infiltration attacks (macro-F1 remains above 0.7 up to approximately 7–8 infiltration edges per Sybil node), and achieves a crossover with GCN performance at approximately 5% labeled seeds. Targeted seed contamination degrades macro-F1 by approximately twice the margin of equivalent random noise, confirming that adversarial label corruption represents a distinct and more severe threat than random label noise.

DOI: http://doi.org/10.5281/zenodo.21716025

Machine Learning–Driven Ultrasonic Characterization Of Cement Mortar HydrationAcross Extended Temperature Ranges

Authors: Vansh Garg, Saksham Aggarwal, Anup Paul, Padmaja Panda, Chinmayee Tripathy

Abstract: Backscattering study of ultrasonic waves is a non-destructive technique used in characterization of cement mortar during hydration. Experimentally measured Attenuation, acoustic impedance, amplitude and velocity of ultrasonic waves at 25C, 32C, and 42C temperatures are considered for analysis. The present study explores different Machine learning models and expands the experimental data recorded at 25C, 32C and 42C to a range of 20C to 45C. Three machine learning models: linear regression, backpropagation-based neural network and linear discriminant analysis (LDA) are used for extrapolation. The critical phase transitions during hydration of cement mortar was detected. For this the angular relationship between attenuation and impedance vectors, was analysed by the Sliding Window Principal Component Analysis (SW-PCA) approach. Further volatility angle was analyzed to distinguish stable, transitional, and decoupled microstructural states of mortar. The SW-PCA approach again used to evaluate the extended datasets generated by various machine learning models. From ultrasonic wave velocity and impedance, density of mortar is estimated for both experimental and predicted datasets. This study was incorporated with angular volatility trends to obtain threshold temperatures for effective setting in mortar. The present study signifies that 25C supports a controlled and consistent solidification. 32C corresponds to a kinetic transition state exhibiting an elevated instability in microstructure and 42C leads to accelerated but structurally weaker densification. Further in the chosen machine learning models, linear regression captures overall trends, LDA and neural network models more appropriately represent non-uniform hydration behaviour and effective capture of crossover effects. The proposed SW-PCA analysis along with the ML framework captures an improved insight into temperature-dependent microstructural evolution during hydration.

DOI: http://doi.org/10.5281/zenodo.21716202

AI-Powered CCTV Surveillance For Intelligent Examination Monitoring And Automated Alert System Using YOLOv8

Authors: Rishabh Dubey, Saransh Kumar, Sanjana, Manpreet Singh

Abstract: Institutional security faces significant challenges in academic environments due to the logistical volatility of high-stakes examinations. This research developed an autonomous surveillance framework specifically designed for proctoring and premise security. The system utilized the YOLOv8 architecture for real-time detection, integrated with a Dynamic Identity Injection module to provision session-specific biometric datasets. This configuration allowed for contextually resetting authorization logic for every individual examination event. Spatio-temporal behavior analysis was facilitated to identify anomalies such as unauthorized item usage and suspicious gaze deviations. To enhance accountability, Vision-Language Models were integrated to generate natural language descriptions of flagged incidents for review by authorities. The architecture employed edge-centric processing and a rolling evidence cache to ensure compliance with the Digital Personal Data Protection Act. By combining probabilistic risk scoring with transient data management, this study provided a scalable and legally resilient solution for maintaining academic integrity in modern educational infrastructures. Experimental evaluation demonstrated a mean Average Precision of 0.92 at IoU 0.5, a processing latency of under 15 ms per frame at 30 FPS, and a 98.2% biometric validation accuracy, confirming the viability of the framework for real-world deployment.

DOI: http://doi.org/10.5281/zenodo.21716279

RescueRoute: An AI-Driven Real-Time Adaptive Traffic Management System For Emergency Vehicle Prioritization

Authors: Abhishek, Arsh Ahmed, Piyush Tiwari, Mr. Karmbir

Abstract: Urban traffic congestion significantly impacts emergency response systems, often causing delays that may result in loss of life and property. Traditional traffic signal systems rely on fixed timing or localized adaptive control, which lack coordination and predictive intelligence. This paper proposes RescueRoute, an AI-driven real-time adaptive traffic management system designed to prioritize emergency vehicles through intelligent multi-intersection coordination. The system integrates computer vision for vehicle detection, deep reinforcement learning for adaptive signal control, and vehicle-to-infrastructure communication for real-time data exchange. Unlike existing approaches, the proposed system introduces end-to-end route optimization, predictive traffic analysis, and dynamic green corridor generation. A hybrid architecture combining centralized intelligence with decentralized control ensures scalability and efficiency. Simulation using SUMO demonstrates significant performance improvements compared to traditional traffic systems. Experimental results show approximately 40–45% reduction in emergency response time, 35–45% reduction in average vehicle waiting time, and nearly 50% improvement in traffic throughput across multiple scenarios. These results validate the effectiveness of RescueRoute in improving emergency response efficiency while maintaining overall traffic flow. Furthermore, the proposed RescueRoute framework emphasizes key characteristics such as reliability, scalability, and adaptability, which are essential for modern intelligent transportation systems. The system is designed to function efficiently in highly dynamic urban environments. where traffic patterns change rapidly. By leveraging real-time data processing and AI-driven decision-making, the framework ensures consistent performance even under unpredictable conditions. These features make RescueRoute highly suitable for deployment in future smart city infrastructures, where automated and intelligent traffic control plays a critical role in improving emergency response efficiency and overall urban mobility.

DOI: http://doi.org/10.5281/zenodo.21719253

A Review On Leveraging Machine Learning And Big Data For Personalized Healthcare Systems

Authors: Abhendra Pratap Singh, Arpit Dwivedi, Shree Bhagwan, Akash Yadav, Shashank Pandey, Nandini Sharma

Abstract: Recent advancements in technology, along with the availability of large volumes of healthcare data, offer an opportunity to adopt innovative technologies such as artificial intelligence (AI), machine learning, and big data in healthcare for better health service delivery. The use of innovative technologies such as artificial intelligence, machine learning, and big data enables efficient decision-making, disease detection, and personalized treatment. This paper reviews machine learning and big data in personalized medicine, presenting details about various tools that can be utilized within the context of healthcare, such as predictive modeling, data mining, and healthcare analytics. Furthermore, emerging technologies in personalized medicine have been discussed, including federated learning, blockchain technology, and real-world data. In addition, the paper also discusses existing developments in intelligent healthcare systems, such as patient monitoring, adaptive learning models, and using healthcare analytics for decision-making processes. Additionally, the paper highlights existing key challenges related to applying machine learning and big data for personalized healthcare, including data heterogeneity, lack of high-quality training data, algorithmic biases, difficulty in model interpretation, issues with security and data privacy, and technical barriers. Finally, the paper highlights the research gaps, examines the existing ways of addressing the problem, and provides recommendations regarding the future of personalized medicine using AI technology.

DOI: http://doi.org/10.5281/zenodo.21719327

A Comprehensive Study On AI Driven Predictive Energy Optimization Frameworks For Electric Vehicle

Authors: Abhendra Pratap Singh, Nandini Sharma, Rachit Sharma, Gauri Kathait, Daksh Bhatia

Abstract: The fast growing use of electric vehicles creates a larger need for intelligent systems that can adaptively optimize energy use in real time. This paper provides an overview of how AI driven predictive optimization frameworks for EV’s are changing the way energy management will be executed. It reviewed the impact of artificial intelligence, machine learning, and deep learning on the limitations of conventional methods for managing and optimizing energy. Advanced algorithms such as Long short term Memory networks, XGBoost, deep reinforcement learning, fuzzy logic, and genetic algorithms allow the predictive modelling of battery performance, power usage, and energy requirements based on an individual route. Utilizing real time data, such as traffic volume, road slope, weather information, and driver behaviour, these systems create a dynamic and context based approach to determining the most effective way to use energy, ultimately resulting in greater distance travelled between recharges and increased battery life. They also explore how Vehicle To Grid Communication (V2G) and model predictive control will work together to optimize energy usage in an operational environment that includes issues, such as sensor reliability and communication dropouts. Some conclusions from this literature review render AI driven energy management systems have excellent performance in controlled traditional environments, but challenges remain in developing standard testing protocols, ensuring data safety from cyberattacks, and the need for onboard real time systems to develop standardised protocols for evaluation and testing. Therefore, in order to use a holistic and multi technology AI structure for predictive analytics, a significant amount of work remains to be done to develop standard operating protocols for energy management of electric vehicles and their associated energy networks.

DOI: http://doi.org/10.5281/zenodo.21719385

A Review On Big Data Analytics Server Machines For Cloud Computing

Authors: Abhendra Pratap Singh, Kanika, Nehal Sharma, Nandini Sharma

Abstract: In today’s era, data is the most important element, and data generation is increasing day by day. Due to this increase in data, traditional servers cannot handle large amounts of information efficiently. Therefore, the present generation requires more advanced server infrastructure. In traditional systems, limited storage, low processing speed and lack of scalability are common issues. In big data systems, the 3V model (Volume, Velocity, Variety) is followed. Big data analytics server machines have high-performance processors. They also include parallel processing support, distributed architecture and cluster-based systems. The main role includes on-demand resources, virtualization, resource pooling and elastic scalability. Therefore, big data analytics server machines are a fundamental component of modern computing environments. Big data analytics plays a vital role in server machines and serves as a fundamental component of modern computing environments. This paper represents how this technology has transformed. This paper discusses big data features such as volume, velocity, and variety, along with evolving technologies such as Apache Hadoop, MapReduce, Spark, SaaS, PaaS, and IaaS. It also explains the differences between big data analytics, big data storage, and big data warehousing. Furthermore, it discusses cloud computing and software such as Kafka and its importance in cybersecurity. Security is a crucial component of cloud-based big data analytics, as it protects sensitive data. Cloud companies ensure protection through data encryption, access control, and compliance with security standards.

DOI: http://doi.org/10.5281/zenodo.21719710

A Study On AI–IoT Integrated Smart Library System For Real-Time Resource Tracking And Efficient Book Retrieval In Higher Education

Authors: Abhendra Pratap Singh, Juned Ahmad, Ananya Singh, Meenu, Nandini Sharma

Abstract: This paper examines how AI and IoT are converting university libraries from book based systems into advanced digital ecosystems with real-time services and improved user experience. It discusses digital resources management, smart libraries, user services, and technology adoption. It discusses AI technologies like ML, NLP, predictive analytics, and conversational systems, which enable efficient cataloging, smart information retrieval, personalized recommendations, and automation, which improve decision-making and overall service quality in academic libraries. It explains how IoT supports tracking, smart systems, and automation with features like smart shelving, space management, and safety. AIoT creates a connected system that improves performance and resource utilization and discusses implementation affected by infrastructure, support, user acceptance, and staff experience. It highlights challenges such as data privacy concerns, value based issues, high costs, and opposition to change. It stresses the importance of future planning, training, and policy development for successful implementation. Overall, AI and IoT are transforming libraries into intelligent, efficient, and user centered systems. It also focuses on making smart library technologies secure, ethical, and sustainable.

DOI: http://doi.org/10.5281/zenodo.21719768

EyeWitness: Dual-Level Iris Geometry And Gradient Fusion For Deepfake Forensics

Authors: Chetan Khalal, Vaibhav Chavle, Prof. Amol Bhadange, Diya Gundecha, Shivam Dabade

Abstract: Deepfakes generated using Generative Adversarial Networks (GANS) have become increasingly realistic, creating risks related to misinformation, privacy loss, and digital security. Existing deepfake detection systems often rely on deep leaming black-box models that lack interpretability and struggle to generalize across different datasets or newer GAN architectures. To address these limitations, this paper presents an interpretable and physiologically grounded deepfake detection method based on iris and pupil analysis-features that GANs still fail to reproduce accurately. The proposed system uses a two-level detection framework. The first level evaluates pupil shape consistency through segmentation, contour extraction, and ellipse fitting. The Boundary Intersection over Union (BloU) score identifies irregular or distorted pupil shapes commonly found in GAN-generated faces. The second level performs iris gradient similarity analysis by generating Sobel-based gradient maps of both irises. The similarity score between the left and right iris gradient maps captures mismatched textures, contours, and reflections- physiological cues that remain consistent in real eyes but not in synthetic ones. Evaluations conducted on FFHQ realimages and StyleGAN3 fake images show strong results. While iris gradient analysis alone achieves an AUC of 0.94, combining both levels yields 97.9% accuracy, 96.0% precision, and an AUC of 0.979. This demonstrates that integrating biometric features provides a reliable, explainable, and computationally. efficient solution for deepfake detection.

DOI: http://doi.org/10.5281/zenodo.21719822

A Machine Learning-Based Adaptive Learning Platform For Class 10 NCERT Using Python And Web Technologies

Authors: S. Swetha, Cr. Sravanthi, M. Thanu Sree Reddy, S. Aarthi

Abstract: The traditional approach to education is not capable of meeting the different capabilities and learning requirements of the students, which leads to different learning achievements. In this paper, a ML-based Adaptive Learning Platform for Class 10 NCERT students is proposed. The performance of the students is determined based on quiz marks, accuracy, time, number of right and wrong responses. The obtained performance data is processed using Decision Tree classification model to classify the learners into one of the three levels, namely Easy, Medium, and Hard. Depending on the learner level predicted by the model, suitable quizzes, study materials, video tutorials and AI assistance are recommended to the students. The proposed system was implemented as a web application and tested using learner performance data. The experimental results revealed that the classification accuracy reached 86%, which proves the efficiency of the Decision Tree model in predicting the learner levels. With the help of the adaptive recommendation approach, the involvement of the learners in the learning process increased due to provision of personalized learning content.

DOI: http://doi.org/10.5281/zenodo.21719864

From Managers To Machinists: Mediators As Disruptors Of Traditional Leadership

Authors: Uma Gautam, Ankush Rana, Nikhil Kumar, Arvind Singh, Sneha

Abstract: Artificial Intelligence is revolutionizing both organizations and society by enhancing operational efficiency and enabling better decision-making processes. However, its rapid development also raises critical ethical issues such as data privacy, inherent biases, and accountability challenges. Addressing these concerns responsibly is essential to ensure that AI technologies are used sustainably and ethically, fostering trust and positive impact across various sectors and communities. The landscape of leadership is evolving to incorporate digital skills and artificial intelligence management, highlighting the importance of maintaining human-centred values. Modern leaders are focusing on ethical practices, transparency, and fairness to foster trust among stakeholders. This shift aims to create a harmonious balance between technological advancement and societal well-being, ensuring that innovation benefits everyone while upholding core principles that prioritize the rights and needs of individuals and communities. The future of leadership relies heavily on effective collaboration between humans and artificial intelligence, with a focus on fostering inclusivity, promoting sustainability, and encouraging ongoing learning and development. Leaders who prioritize ethics and demonstrate adaptability are essential for ensuring responsible AI implementation. Such leaders not only drive organizational success but also contribute positively to societal progress, creating a balanced and innovative environment for all stakeholders.

DOI: http://doi.org/10.5281/zenodo.21719915

Impact Of AI On Work-Life Balance, Job Security And Job Satisfaction

Authors: Rachna Chandra, KM Reshama

Abstract: Artificial intelligence has affected all aspects of our life from education to our jobs, entertainment and our health. This study has been done by considering four attributes – work-life balance, job satisfaction, job security, and artificial intelligence. In this study, we will qualitatively explore the impact of artificial intelligence on work-life balance, job satisfaction and job security. On the one hand, artificial intelligence helps in creating work-life balance by reducing work load. On the other hand, it has created concern about job security among employees working in various industries. In this study, we have found that people have mixed responses on how AI affected the job market. There are some people who think that AI will enhance job opportunities in the future while others consider it a threat for future job conditions.

DOI: http://doi.org/10.5281/zenodo.21720665

FocusGuard: An Emotion-Aware Anti-Procrastination Browser Extension Using Sentiment Analysis

Authors: Shubham Sharma, Devayush, Daksh Arora, Karmbir

Abstract: Procrastination presented a pervasive challenge that hindered productivity and negatively impacted mental well-being, largely because traditional intervention tools like website blockers failed to address the underlying emotional states driving this behavior. This study introduced FocusGuard, a novel browser extension that utilized sentiment analysis to evaluate user emotions and deliver highly personalized, context-aware interventions. Unlike conventional rule-based systems, FocusGuard integrated Natural Language Processing to dynamically monitor digital behavior and optional text inputs, enabling the system to understand whether a user was experiencing stress, boredom, or a lack of motivation. The methodology involved a multi-stage approach where data was collected, preprocessed, and analyzed to trigger an intelligent decision engine. This engine subsequently provided emotionally intelligent responses, such as customized motivational messages and focus-enhancing audio, tailored specifically to the user's current affective state. The results indicated that combining artificial intelligence with human-centered design created a more supportive environment that effectively mitigated procrastination compared to static blocking mechanisms. Ultimately, the system fostered long-term productivity improvements by addressing emotional dysregulation at the core of task avoidance.

DOI: http://doi.org/10.5281/zenodo.21720792

Navigating Adolescence In A Digital Age: Social Media And Mental Health

Authors: Ms. Ritu Jangra, Mayuresh Yadav, Riya Rani

Abstract: The proliferation of social media portals has substantially impacted the daily lives of adolescents, escalating issues about their consequences on mental health. The surveys analysed that extensive use of social media by teenagers exploits their health by increasing anxiety, depression, sleep disturbances and low self-esteem. On the other side, these portals formulate a broad amount of user-generated data, building new frontiers for enhanced technological analysis. The latest evolution in machine learning, deep learning and cloud computing gave researchers access to evaluate massive amounts of data from social media that identify the behavioural patterns associated with mental health conditions. Different methodologies, such as natural language processing, sentiment analysis, and predictive modelling, are gaining attention for their ability to detect early signs of depression and anxiety by observing online activities, posts, and user interactions. This review examines the correlation between social media usage and adolescent mental health while assessing the significance of cloud-based infrastructure and artificial intelligence (AI) in investigating such data. Cloud computing provides scalable storage and processing capabilities that are essential for managing big data sets which are generated by social media platforms, enabling immediate overviews and analysis. Nevertheless, in spite of the positives of these technologies in early detection and intervention, numerous concerns remain the same, such as data privacy, ethical issues, algorithm bias and limitations in accurately interpreting human emotions. The analysis presents that augmented usage of social media generates a higher risk of anxiety and depression among adolescents; the inclusion of machine learning and cloud computing offers solutions for early detection and mental health support. Continued research should focus on enhancing ethical, transparent and reliable systems that manage technological advancements with user privacy and well-being.

DOI: http://doi.org/10.5281/zenodo.21720841

A Dual-Stage Diagnostic Framework For Diabetic Retinopathy Grading And Lesion Segmentation Using EfficientNetV2 And NnU-Net

Authors: Md. Anisur Rahman, Rahul Dahiya, Naveen Sharma, Vinay Kumar Nassa, Pradeep

Abstract: Diabetic Retinopathy (DR) remained a leading cause of preventable vision loss worldwide, with its prevalence driven by the global diabetes epidemic. This study aimed to develop an accurate and scalable automated screening solution to overcome the interpretability limitations of existing black-box classifiers. A dual-stage diagnostic framework was proposed that integrated global severity grading with pixel-level lesion segmentation. In the first stage, an EfficientNetV2-S architecture was optimised with Focal Loss (gamma = 2.0) and five-pass Test-Time Augmentation to categorise fundus images into five DR severity grades (0 to 4), achieving a peak training accuracy of 94.43% and a best validation accuracy of 61.17%. In the second stage, a self-configuring 9-stage nnU-Net performed semantic segmentation of pathological biomarkers using high-resolution patches (1024 x 1536 pixels), achieving a Dice Similarity Coefficient of 0.81 for Hard Exudate detection, which surpassed all existing IDRiD dataset benchmarks. An adaptive inference mechanism was implemented to ensure that computationally intensive segmentation was triggered only when pathology was confirmed by Stage 1, thereby preserving efficiency on standard hardware. The complete framework was deployed as a portable, cross-platform desktop application via the Eel library and PyInstaller, enabling clinical use without internet connectivity or specialist dependency management. Results demonstrated that combining weighted classification with deep segmentation significantly enhanced both the interpretability and the reliability of automated DR screening.

DOI: http://doi.org/10.5281/zenodo.21721320

Technological Evolution In Software Development Life Cycle: The Implications Of Artificial Intelligence

Authors: Ms. Ritu Jangra, Riya Rani, Mayuresh Yadav

Abstract: The Software Development Lifecycle (SDLC) is a pathway for software designing, modification, evaluation and optimisation. Baseline models like Waterfall and Agile go through problematic situations like inflexibility, delayed error detection, scope creep and obstacles in the management of complex projects. To tackle these anomalies, Artificial Intelligence (AI) took the spotlight by intensifying various phases through automation, predictive analytics, and smart decision-making. This review addresses the outcomes of AI in SDLC, outlining the improvement in required analysis, system designing, code generation and bugs detection as well as testing and deployment. AI technologies like GitHub Copilot and ChatGPT are reinforced to be the real-world applications that alter the developer productivity and software quality. AI serves various purposes like increased efficiency, reduced development times, higher accuracy and early bug detection. Despite these benefits, the study illustrates various demerits like data dependency and lack of transparency as well as ethical concerns. But it enhances the requirement of standardised frameworks, enriched explainability and optimised cooperation among humans and AI. Ultimately, AI can considerably boost the SDLC, but a balanced automation with human expertise is important to shape the future of software engineering by promoting more efficiency and steady and knowledgeable development in the processes.

DOI: http://doi.org/10.5281/zenodo.21734603

Mitigating Algorithmic Bias Through Disparate Impact Ratio: A Pre-processing Approach To Fair Machine Learning

Authors: Arzu, Assistant Professor, Nikhil Kumar, Dhruv Upreti, Durgesh Rajpurohit, Vanshika thakur

Abstract: Artificial Intelligence (AI) systems are increasingly used in high-stakes decision-making domains such as recruitment, healthcare, finance, and education, where fairness and accountability are critical. Although these systems are often perceived as objective, they can inherit and amplify biases present in historical data, leading to discriminatory outcomes for certain demographic groups. This study investigates algorithmic bias in machine learning models, with a primary focus on measuring and mitigating unfairness using the Disparate Impact Ratio (DIR) and Equal Opportunity Difference (EOD).Using the Adult Income dataset, this research evaluates bias in income prediction tasks where gender is treated as a protected attribute. Two machine learning models Logistic Regression and Decision Tree are implemented to analyze both predictive performance and fairness. The study follows a two-stage experimental design: a baseline model trained on original data, and a bias-mitigated model using the Disparate Impact Remover, a preprocessing technique that adjusts feature distributions to reduce dependency on sensitive attributes.Results demonstrate that bias mitigation techniques can significantly improve fairness metrics, particularly by increasing the Disparate Impact Ratio toward acceptable thresholds and reducing disparities in true positive rates across groups. However, these improvements often come with a trade-off in predictive accuracy, highlighting the inherent tension between fairness and performance. The findings emphasize that no single bias mitigation strategy is universally optimal; instead, effectiveness varies depending on the dataset, model, and application context.This study contributes to the growing field of fairness-aware machine learning by providing a comparative evaluation of bias detection and mitigation techniques in practical settings. It underscores the importance of integrating fairness measures into the AI development lifecycle and offers insights for designing more ethical, transparent, and inclusive AI systems

DOI: http://doi.org/10.5281/zenodo.21735295

The ML based H2 blended fueled internal combustion engines: Optimization, challenges, and future directions

Authors: Gautam Panwar, Adarsh Kumar Sharma, Kuber Bassi, Piyushi Suyal, Bhumi Sehrawat, Praveen Kumar

Abstract: Engines running on hydrogen (H₂ICE) offer strong potential for reducing traditional fossil fuels while leveraging existing engine architectures. Hydrogen introduces key advantages, like high combustion efficiency, broad ignition ranges, and no carbon output. Yet problems such as the creation of NOₓ, combustion instability, and storage limitations halt its widespread rollout. In this context, machine learning (ML) techniques provide a practical way for handling the complex and nonlinear behavior of hydrogen combustion systems. Complex setups, covering artificial neural networks (ANN), support vector regression (SVR), random forest (RF), and gradient boosting algorithms such as XGBoost, demonstrate strong capability at forecasting how the engine runs, optimizing operating conditions, and controlling emissions. This paper presents a comprehensive assessment of ML-assisted optimization of hydrogen engine systems, aimed at boosting output and lowering pollution. By combining hydrogen fuel with data-driven strategies, real-world gains in engine efficiency, stable operation, and reduced emissions can be achieved, supporting the transition to sustainable energy systems.

DOI: http://doi.org/10.5281/zenodo.21768087

NUTRIVEDA: An AI-Driven Framework For Personalized Nutrition Integrating Machine Learning, Large Language Models, And Ayurvedic Principles

Authors: Tarun, Dr. Naveen Sharma, Mr. Rahul Dahiya, Ms. Sakshi Rajwar, Mr. Tarun Singh Negi

Abstract: Modern lifestyle trends, particularly the rise of sedentary behaviors coupled with the proliferation of ultra-processed diets, pose an unprecedented challenge to global public health. The epidemiological shift toward non-communicable diseases, such as obesity, type 2 diabetes mellitus, cardiovascular pathologies, and complex gastrointestinal disorders, has demonstrated the inadequacy of generalized dietary guidelines. This paper presents the concept of "NUTRIVEDA," a comprehensive framework encompassing AI-driven recommendation systems, clinical intervention frameworks, and the integration of artificial intelligence with traditional Ayurvedic medicine. The core objective is the generation of tailored weekly meal plans that map precisely to the nutritional needs, biometric markers, and preferences of the individual user. By synthesizing disparate data streams, these AI-driven platforms represent a paradigm shift in preventative medicine.

DOI: http://doi.org/10.5281/zenodo.21768139

Analyzing Model Generalization: A Study Of Overfitting, Underfitting, And Bias–Variance Trade-Off

Authors: Sudhanshu, Nitin Gupta, Dev Button, Jyoti Bansal, Vinayak Sharma

Abstract: Machine learning models usually work better or worse depending on how well they can use what they learned from the training data to make predictions about new data. Overfitting and underfitting are the two biggest problems in this area. Both can have a big impact on how reliable and accurate the model is at making predictions. This is a meta-analysis of overfitting, underfitting, and generalization that looks at a lot of different machine learning models and datasets. The meta-analysis integrates theoretical frameworks and empirical experiments to examine the training and testing performances across various model complexities. The generalization gap is the main way to measure how well a model works, and the bias-variance tradeoff is the main idea behind the whole paper. The research studies used standard benchmark datasets and a range of algorithms, such as linear models, decision trees, ensemble methods, and neural networks. Error curves, comparative metrics, and statistical summaries were all used to show the patterns in model performance.

DOI: http://doi.org/10.5281/zenodo.21768178

A Comprehensive Review On Role Of Linear Algebra In Artificial Intelligence And Various Engineering Applications

Authors: Manish Kumar Jain, Dharmendra Singh Sengar, Abhendra Pratap Singh, Yash Kumar, Tarun Kumar Mishra, Nandini Sharma

Abstract: AI is growing much faster in every field including engineering and linear algebra has provided a strong base for it. Linear Algebra is a core branch of mathematics which is used in mostly all engineering fields. This review paper carries out its applications on five major fields of engineering, i.e., civil engineering, electrical engineering, mechanical engineering, computer science engineering, and chemical engineering. Vectors, matrices, system of linear equations, linear transformation, and eigenvalues have been discussed here. Linear algebra can be seen in civil engineering through structural analysis with the help of finite element methods and hydraulic modeling. In electrical engineering, it has a big part in circuit analysis, signal processing, control systems as well as communication; while in mechanical engineering it is applied for vibration studies robotics and fluid dynamics. In chemical engineering modeling processes thermodynamics and transport phenomena take place with the help of linear algebra. In computer science linear algebra helps machine learning data analysis computer graphics and cryptography to advance further. This paper illustrates the critical role that linear algebra plays as an underlying structure supporting AI and ML systems by their use of vector and matrix-based data representation as well as model computation. It reveals how different techniques such as neural networks dimensionality reduction optimization are based on fundamental principles from linear algebra to handle large datasets in modern engineering fields. Hence Linear algebra is important for new ideas and tech progress in engineering.

DOI: http://doi.org/10.5281/zenodo.21768308

Sustainable And Progressive Societal Development Asserts Artificial Intelligence Measures In Creating And Restructuring Smart Cities

Authors: Abhinav Singh, Arya, Annapurna, Dr. Suresh Kumar

Abstract: Smart cities require artificial intelligence (AI) based technologies for planning of land divisions and mapping, sewer and drainage systems, road and traffic management. New York city, USA and Dubai city, UAE; London city, UK and Melbourne city, Australia; Singapore city, Singapore and Tel Aviv, Israel considered and reviewed to understand and analyze these AI based technologies already used in above specified areas. Planning, re-structuring and implementation of AI based technologies for smart cities is inevitable and Indian cities needs implementation of these technologies.

DOI: http://doi.org/10.5281/zenodo.21768343

Ai Chatbot With Real Time Voice Communication

Authors: Vibhu Tyagi, Utkarsh Bhardwaj, Rupesh Kumar

Abstract: Virtual Assistant is significant for people to interact with computers and perform tasks on different environment in the today’s world. Old chatbot systems are not much efficient at handling many users simultaneously and lack secure mechanisms. In this paper, we proposed Assistant AI Chatbot that can understand the feeling of human and response the answer quickly. AI Chatbot’s Process Flowcharts and Diagrams have been designed. We have conducted comprehensive analyse of Technology Stack, Security Hardening and User Interface Implementation. In the Evaluation and result we have examined the result of Performance Analysis, System Efficiency Evaluation Security and Reliability Analysis. My proposed AI Chatbot performance is better than Tradition AI Chat Bot.

DOI: http://doi.org/10.5281/zenodo.21768411

Adversarial Attacks On AI-Based Cyber Security Systems

Authors: Pragati Bhandari, Aditya Patil, Prince Khalane, Mayuri Pawar, Kiran Salunke

Abstract: Artificial Intelligence (AI) and Machine Learning (ML) have completely changed the landscape of cybersecurity and added advanced capabilities for building these intelligent intrusion detection systems, malware classifiers, or anomaly detection frameworks. However, this integration has brought a new and particularly severe category of threats – adversarial attacks, which exploit the inherent vulnerabilities present in various AI and ML models to get round security mechanisms. This review paper represents systematic and comprehensive analysis of adversarial attacks against AI based cybersecurity systems. We conducted a survey of the taxonomy of adversarial attacks such as evasion, poisoning, model extraction, model inversion, and backdoor attacks and take a look at their methodologies, threat models and their real-life implications. We analysed attack algorithms starting from the Fast Gradient Sign Method (FGSM), Projected Gradient Descent, Carlini and Wagner (C&W) attacks, DeepFool, and new black-box methods including MI-FGSM and AutoAttack. Furthermore, we systematically evaluated state-of-the-art defence mechanisms also known as certified and adversarial training, input preprocessing, randomized smoothing and ensemble defences. We presented an empirical comparison of attack success rates from 6 AI-based intrusion detection systems and measure the success of defences under a variety of attack paradigms. Our analysis showed that while adversarial robustness has made great progress, there still exists a large attack capability-defensive mechanism discrepancy. This paper concludes with identification of open research challenges for the future and future directions that are crucial for creating trustworthy AI-based security systems.

DOI: http://doi.org/10.5281/zenodo.21768465

A Comprehensive Timeline Review Of Artificial Intelligence, Machine Learning, Deep Learning And Explainable AI Based Models And Methods

Authors: Sandeep Dalal, Dr Jyoti Chaudhary

Abstract: We are witnessing transitioning from rule-based systems to complex deep neural architectures in the way Artificial Intelligence (AI), Machine Learning (ML), Deep Learning (DL) and Explainable AI (XAI) that have evolved over the past decades and keep on growing. This paper presents a comprehensive and detailed literature review in a timeline sequence while highlighting key models and the important methods that have been developed from year 1950 to year 2025. The study categorizes all the major developments done into AI, ML, DL and XAI providing a detailed comparative and logical analysis of their working principles and their limitations. Here in this proposed review paper it identifies the research gaps and the major emerging trends such as hybrid XAI systems along with human centered AI systems.

DOI: http://doi.org/10.5281/zenodo.21768659

SoulSpace: An AI-Powered Anonymous Mental Health Platform For College Students In India

Authors: Satyam Gupta, Preeti, Manav Garg, Aditya Kr Choubey, Piyush Bhandari

Abstract: Mental health disorders among college students represent a growing public health concern, yet a substantial proportion of affected individuals never receive appropriate care. In India, where the treatment gap is estimated between 60% and 83%, two predominant barriers — social stigma and the absence of culturally appropriate digital infrastructure — prevent students from seeking help even when support is theoretically available. This paper presents SoulSpace, an AI-powered campus mental well-being platform with three core technical contributions: (1) an anonymous identity architecture requiring no personally identifiable information for core platform access; (2) Aura, a custom fine-tuned chatbot built on TinyLlama 1.1B trained on India-specific mental health dialogue using Low-Rank Adaptation (LoRA); and (3) a composite Wellness Score Engine integrating PHQ-9 and GAD-7 clinical instruments with real-time NLP sentiment analysis and behavioural signals into a 0–100 score driving a five-band Action Ladder. The platform further incorporates an in-chat SOS detection system combining a rule-based crisis lexicon of over 800 expressions with a fine-tuned BERT classifier achieving ROC-AUC of 0.88 at latency below 50 ms. All components are validated on synthetic data (N = 4,200); prospective clinical validation is required.

DOI: http://doi.org/10.5281/zenodo.21768769

Learnova: A Scalable and Secure MERN-Based E-Learning Platform for Modern Web Applications

Authors: Saba Zaidi, Tejaswa Rajput, Ayushi kumari, Aryan, Rahul Bhandari

Abstract: To enhance online learning experiences, a new educational platform called Learnova is being designed within this project. As stated in the objectives of this paper, the primary objective is developing a website that is user-friendly, efficient, and safe. The proposed website should facilitate collaboration between students and teachers and provide a convenient way for organizing educational processes. Learnova will leverage contemporary technologies that will enable more convenient adaptation of learning activities and management of course development and administration. It utilizes a MERN stack for implementation purposes. It is a group of technologies such as MongoDB, Express.js, React, and Node.js that enable websites to be fast and highly scalable. The platform design is easy to perceive and intuitive. Furthermore, Learnova utilizes reusable and maintainable code, as it relies on contemporary web technologies that include MongoDB, Express.js, React, and Node.js among others. Moreover, the platform utilizes RESTful APIs to ensure efficient interaction between the system modules. Learnova features several essential functions. For instance, users will be able to develop, administer and control courses. The system also provides user login and access control. One can monitor the progress of the course completion. Learnova users can also communicate with one another in real-time. The site was developed with the use of HTML, CSS, and JavaScript technologies that create an engaging interface for Learnova users. Additionally, it features integrated payments and cloud storage for media files to provide improved reliability. Learnova utilizes appropriate authentication methods and encryption technology to ensure user data protection. The proposed platform is beneficial for many reasons. First, users will be provided with feedback and suggestions regarding their current achievements and goals to pursue. Secondly, the platform works on multiple devices that ensures convenience of usage and high mobility. In summary, this project emphasizes the importance of modern web technologies in designing efficient platforms. In conclusion, Learnova aims to provide a reliable and safe environment for users by implementing MERN stack technologies.

DOI: http://doi.org/10.61463/ijset.vol.14.issue4.166

A Review On Machine Learning Driven Frameworks For Employee Behaviour Prediction And Proactive Organizational Risk Management In Human Resource Management

Authors: Abhendra Pratap Singh, Nehal Sharma, Kanika, Nandini Sharma

Abstract: In today’s era, many multinational companies (MNCs) operate worldwide. In these organizations, human resource management and senior managers face several challenges, such as employees leaving their jobs suddenly, low performance, stress and burnout, and skill gaps. This paper reviews the use of machine learning, artificial intelligence (AI), and sentiment analysis to address these problems effectively. These tools help predict which employees may leave their jobs, who may underperform, potential future workforce risks, and other related issues. The system works by collecting employee data, cleaning and preparing the data, predicting risks, and assisting human resource management in taking early action. This approach provides several benefits, such as reducing employee turnover, saving company costs, and improving employee satisfaction. The paper reviews machine learning based frameworks that identifies human resource problems at an early stage to avoid such issues. It aims to transform human resource management from a reactive approach to a proactive one and improve decision-making for betterment.

DOI: http://doi.org/10.5281/zenodo.21769373

A Comprehensive Study And Analysis Of Shortest Path Algorithms: Classical To Heuristic

Authors: Sakshi, Aastha, Mansi, Prince Kumar Sharma

Abstract: The Shortest Path Problem (SPP) is an important part of graph theory with Network optimization with today’s system. The main objective of this research paper is to present a comprehensive analysis, ranging from classical algorithms to modern approaches. In this paper, we compared the Dijkstra, Bellman-Ford, Floyd-Warshall, Johnson's, and A (A-Star)* algorithms based on their theoretical complexity, time complexity, algorithms and real-world applications. Our research highlights that while Dijkstra's algorithm remains the main function for navigation systems, Bellman-Ford's algorithm works on graphs with negative weights. Additionally, we explore how the A* algorithm optimises the search space using heuristic functions, making it a better. Alternative to Artificial Intelligence and robotics. Through In this research analysis, we conclude that the efficiency of any The algorithm depends mainly on the network structure and specific constraints. This paper provides researchers and Developers with a clear roadmap for choosing the right algorithm for their shortest and fastest path.

DOI: http://doi.org/10.5281/zenodo.21772049

Music Recommendation System Using Facial Expressions

Authors: M Velmurugan, Danthaluri Tharuni, Maramreddy udaya sree

Abstract: Music has a strong influence on how we feel and on how we experience our world. As AI and ML are rapidly developing, new possibilities are opening for making technology more personalized and emotionally intelligent.The project, entitled “Music Recommendation System Using Facial Expressions,” seeks to create a music player that can recognize the emotional disposition of its listener and provide recommendations for songs that correspond to their emotional states.A webcam records the face of the user in real time. Open CV and CNN algorithms analyze various emotions of the user such as happiness, sadness, anger, surprise, or having a neutral expression. Once such an expression has been identified, the system selects music from an existing database that comprises songs for various moods, suggesting songs that fit the user’s mood at any particular moment. A simple interface has also been offered for listening to these suggested songs.This kind of technology is quite different from the usual recommendations that depend solely on past listening patterns and selections. This technology offers a more intuitive and emotional link with the selection of music. This technology can prove beneficial in the entertainment sector itself and in providing health assistance. Preliminary testing of this technology has proven quite effective in terms of emotion recognition and popularity of the recommendations.Overall, this project aims to connect human emotions with smart digital systems by using computer vision, deep learning, and affective computing for a more personalized music experience.

DOI: http://doi.org/10.5281/zenodo.21772116

AI-Powered Sign Language Converter Using Image Recognition Techniques For Smart Home Control

Authors: Ameya Katkar, Riya Suryavanshi, Harshad Kadam, Purva Kamat, Prof. M. S. Chavan

Abstract: The demand for assistive technologies and smart home automation had increased the necessity of developing a system with effective and real-time human-computer interaction. This paper proposes an AI-based sign language translator for smart home automation, where the user can use hand gestures to control home appliances. We used a computer vision-based, deep learning IoT embedded system to design a simple yet effective system for accurate and low-latency gestural identifications. Unlike raw image-based approaches, the system uses vision-based input acquired by a camera, from which MediaPipe extracts 21 hand landmarks for every frame. For identifying dynamic gestures, one approach is to use a hybrid CNN–LSTM model that learns spatial and temporal features from sequences of gesture data. Trained on gesture sequences of 30 frames per sample, the model obtains an average recognition accuracy ~95% on the test dataset. The system operates on an edge device which runs on Raspberry Pi technology to deliver its real-time function while achieving cost-effective and energy-saving results. The system operates in real-time environments because it processes each gesture with an average inference time of one to two seconds. The system uses relay modules for gesture recognition to create control commands which are sent to IoT devices to perform actions such as activating buzzers and controlling fans and switching lights on and off. The suggested method achieves precise results through landmark-based processing which decreases computational needs by 70 to 80 percent compared to standard image-based deep learning methods. The system performs well in a variety of backgrounds and lighting situations, making it appropriate for real-world implementation. The proposed method provides a flexible solution which operates with short delays and low costs to create smart home automation systems that assist disabled users to achieve independent living.

DOI: http://doi.org/10.5281/zenodo.21772166

Explainable AI Framework For Early Detection Of Mental Stress And Burnout Using Multimodal Behavioral Analysis

Authors: Kabir, Poonam, Meenakshi, Divanshi, Reenu Batra

Abstract: Next-day mental stress has important consequences for health, learning, and work performance, yet most digital phenotyping studies focus on detecting current stress rather than forecasting it. Wearables, smartphones, and in-the-moment self-reports provide continuous information about physiology, behaviour, and experience, and explainable artificial intelligence offers tools to make prediction models more transparent to users and clinicians. This narrative review synthesized recent research on multimodal, explainable next-day stress prediction, with particular attention to biophysical signals, smartphone-based behaviour, micro ecological momentary assessment (micro-EMA), and visual emotion analysis, and it outlines a conceptual framework for explainable next-day stress prediction that can guide future empirical work. The review examined studies that used physiological signals such as electrodermal activity, heart rate, and skin temperature, smartphone-derived features, EMA or micro-EMA self-reports, and facial-expression data for stress detection or short-term stress prediction. Major scientific databases were searched using combinations of terms related to stress detection, digital phenotyping, micro-EMA, next-day stress, explainable AI, and facial-expression datasets, and the identified studies were grouped thematically into four domains: biophysical stress detection, smartphone and micro-EMA-based sensing and prediction, explainable models in stress and mental-health prediction, and visual analysis of stress. Across the literature, biophysical models based on electrodermal activity, heart rate, and skin temperature provided robust stress markers, while smartphone features and micro-EMA ratings enabled forecasting of short-term and next-day perceived stress in daily life. Explainable AI methods, especially feature-attribution techniques such as SHAP, were used to highlight which behavioural or physiological variables drive model outputs, and facial-expression models trained on datasets such as RAF-DB and CK+ could separate stress-related emotions from non-stress states. However, fully integrated pipelines that combine biophysical, behavioural, self-report, and visual data for explainable next-day stress prediction remain rare. Existing work demonstrates that next-day stress can be predicted from recent physiology, behaviour, and micro-EMA, and that explainable models make these predictions more interpretable. Future research should move toward multimodal, explainable next-day stress frameworks that unify biophysical, smartphone, and visual signals, incorporate user-centred explanations, and are evaluated in ecologically valid student and worker populations.

DOI: http://doi.org/10.5281/zenodo.21772269

An Analysis And Evaluation Of The Army Design Bureau As A Digital Framework For Defence Innovation

Authors: Nikhil Rajput, Ish Pandey, Dmitri Kruglov, V.C. Pandey

Abstract: The Army Design Bureau is a platform that helps the Indian Army work with soldiers, startups, industry, and institutions to come up with new ideas. The platform helps manage the process of coming up with new ideas, from submitting them to putting them into action. The Army Design Bureau helps the Indian Army to stay ahead in technology. Makes India a leader in defence innovation. The Army Design Bureau makes it easier for people to submit and track ideas. Helps the Indian Army to find solutions to its problems and makes it easier for people to work together. It helps people to work together and helps make the process more transparent and efficient. By doing this, it helps the Indian Army achieve its goal of being a leader in technology. The platform is important for the Indian Army’s goal of being a leader in defence innovation. The Army Design Bureau helps the Indian Army work with others to come up with ideas and solutions. It makes the process of coming up with ideas more efficient and transparent. The platform helps the Indian Army stay ahead in technology. This paper examines the architectural framework and institutional integration of ADB and evaluates its impact on defence innovation management and capability development. A qualitative case study approach is adopted, incorporating framework analysis, secondary data review, and process mapping of the platform.

DOI: http://doi.org/10.5281/zenodo.21772902

WasteWise: An Incentive-Based Digital Platform For Circular Economy- Driven Waste Management And Upcycled Product Ecosystems

Authors: Ish Pandey, V. C. Pandey, Abhendra Pratap Singh, Farooque Alam, Manya Jain, Parth Dabas

Abstract: The emergence of more paper, plastic, floral, and reusable wastes presents substantial environmental problems that call for new approaches to solving them. In view of the above, this paper will examine the WasteWise platform as a potential innovation aimed at transforming the management of waste using the principles of circular economy. The application will allow citizens to donate recyclable waste products such as paper, flowers, and plastics, earning digital incentives in return for donating their waste products. The incentives earned will enable the donors to buy sustainable products on the online marketplace of the application. The distinguishing feature of WasteWise is the circular approach applied, whereby the donated waste will be converted into sustainable products. The process will make sure that the products obtained from the conversion will enter the consumer market via the WasteWise platform. It will be noted that by integrating waste collection, incentives, upcycling, and eco-commerce into a single system, the above model of the waste management process minimises waste generation, promoting responsible consumption and supporting zero-waste initiatives. The research paper will focus on WasteWise and will examine how the technology can be used to introduce circular economy concepts into practice.

DOI: http://doi.org/10.5281/zenodo.21772944

Beyond Heuristics: A Data-Driven Hybrid Architecture For Semantic Resume Analysis And Context-Aware Explainable AI

Authors: Priyanka Narang, Tathya Sharma, Aditi Kashyap, Mamta

Abstract: Traditional Applicant Tracking Systems (ATS) and open source resume analyzers rely on rigid regex rules and hardcoded penalty schemes. As a result, they are brittle, and candidates are penalized for using different but equivalent terminology. These systems often fail to reflect how skills and roles actually appear in real industry data. This paper presented a fully data-driven multi-model machine learning pipeline that replaced arbitrary scoring rules with seven coordinated models. A Multi-Stage Waterfall Extraction model combined a custom spaCy named entity recognition system with SBERT-based cosine similarity to overcome sparse and noisy annotations. An XGBoost-driven linear regression severity engine and an implicit ontology inference module revealed invisible ATS filters rooted in real hiring expectations. A deterministic Explainable AI layer then grounded all feedback in the candidate's own verified text, removing operational hallucinations. The paper reported empirical results including the Soft Skill Anomaly, which challenges common developer assumptions about soft skills in ATS design.

DOI: http://doi.org/10.5281/zenodo.21772999

Depression Detection System Using BERT: An Extensive NLP Study

Authors: Guni Paliwal, Mohini, Yashvi Gaur, Dr. Lalit Kumar Sagar

Abstract: Depression is one of the most common mental health problems, but it often goes unnoticed due to the social shame associated with the mental health issues and because employees are not regularly monitored for their psychological well-being. The recent advances in Natural Language Processing are now able to identify the emotional patterns from written text. This offers a way of tracking the changes in emotional and mental health in a way that doesn’t make the person feel watched or monitored, thus protecting their privacy. Here, we propose a method to analyze diary-style entries written by employees, which can show early signs of depression. It is a quiet way to check on people’s mental health as this system uses things that have already written, like journals, rather than asking the direct and personal questions. In this approach, we used two complementary models: LSTM network which captures that how emotions appear in the text as they change over time, and a fine-tuned BERT model which understands the deeper and more complex meaning and context of the words used. The text data goes through structured preprocessing: normalization, tokenization, and addressing class imbalance. For analysis, the text is categorized into three levels of depression: no depression, moderate depression, and severe depression. The system is evaluated based on its performance in identifying the depression, and is done by evaluating precision, recall, and F1-score, but the main focus is on recall which makes sure that the system doesn’t miss anyone who might need help. This is done to be extra careful and to avoid overlooking the employees who might be at risk. For clarity, the system uses two main techniques to explain its decisions, one is attention-based highlighting, which highlights the most important words or sentences that influenced the analysis. The other technique, called SHAP analysis, which helps to drive the model’s decisions by showing that how each part of information in text contributed to final conclusion. To protect employee privacy, the system uses ethical safeguards which make sure the data is anonymous, getting permission from user before analysis. The entire system is designed to be the most supportive tool for companies to detect early signs of stress and to create a more caring and healthy workplace for everyone.

DOI: http://doi.org/10.5281/zenodo.21786596

DL Thermal Imaging For Crop Drought Classification

Authors: Karmbir, Kunal Rawat, Akshit Pandey, Cherry

Abstract: The precise and timely identification of drought stress in plants is vital for optimizing crop yield and water management. Although preliminary results on deep learning models have indicated promising results in processing thermal images for stress identification, previous studies have faced significant challenges due to small dataset sizes, single-crop evaluation, and highly controlled and manipulated experimental conditions. This paper proposes a robust and scalable deep learning framework for multi-class drought stress identification based on thermal image analysis. To bridge the gap between current research works, we propose a large- scale dataset for thermal image analysis under natural environmental conditions. We compare our proposed framework with state-of-the-art models and propose a comprehensive framework for real-time deployment on edge platforms such as UAVs.

DOI: http://doi.org/10.5281/zenodo.21786618

Beyond Words: Emotion-Aware Automatic Speech Recognition Using Prosodic Feature Fusion

Authors: Barun Singh Bisht, Gurpreet Kaur, Gurpreet Kohli

Abstract: Automatic Speech Recognition (ASR) systems achieved significant progress in converting speech into text; however, most existing approaches focused primarily on lexical accuracy and overlooked the emotional context present in human speech. This limitation reduced the effectiveness of ASR in applications that required natural and expressive interaction. In this study, a pipeline was developed for preserving emotional information from speech through prosodic feature analysis. Raw audio was processed through extraction, enhancement, alignment, and feature analysis stages to generate a structured dataset with synchronized annotations. Two models, a Support Vector Machine (SVM) and a Bidirectional Long Short-Term Memory (BiLSTM) network, were trained on six emotion classes, namely happy, sad, neutral, angry, fear, and curious, to evaluate their performance in emotion recognition. The experimental results showed that while the SVM provided a reasonable baseline, reaching 50 percent accuracy at 1000 samples per emotion, the BiLSTM model achieved higher accuracy of 69 percent under the same conditions, owing to its ability to capture temporal dependencies in speech. These findings highlight the importance of prosodic features and sequential modelling for developing more expressive and context-aware speech recognition systems.

DOI: http://doi.org/10.5281/zenodo.21787176