Jaihind College of Engineering Artificial Intelligence and Data Science Technical Seminar (JCOEAIDSTS) | International Journal of Science, Engineering and Technology

Retrieval-Augmented Generation For Document Question Answering: A Review Of Retrieval, Embedding, Vector Search And Grounded Answer Generation

Authors: Nalawade Soham Sandip, Prof. J. M. Patade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: The rapid growth of digital documents creates a need for systems that can find relevant information and answer questions accurately. Traditional question-answering systems rely on predefined knowledge bases or keyword search and may fail to understand the meaning of complex questions, while large language models alone can produce fluent but unsupported answers. Retrieval-Augmented Generation (RAG) combines information retrieval with the generation abilities of large language models. This paper reviews RAG for document question answering. It explains sparse retrieval with BM25, dense retrieval with dual encoders and sentence embeddings, approximate nearest-neighbour vector search, re-ranking, and the integration of retrieved passages into generation, including fusion-in-decoder and self-reflective retrieval. Practical issues such as chunking, context length and the tendency of models to overlook information in the middle of long contexts are discussed, together with evaluation of faithfulness and relevance. A five-layer system is proposed in which users upload PDF documents that are extracted, cleaned and chunked, embedded with a sentence-transformer model and stored in a vector database; queries retrieve relevant chunks that are passed to an LLM to generate grounded answers through a Streamlit interface. Research gaps, objectives and a comparison of retrieval methods are presented.

DOI: https://doi.org/10.5281/zenodo.24102328

Reinforcement Learning For Energy Management In Smart Homes: A Review Of Algorithms, Environments And IoT-Integrated Control

Authors: Malwal Shreyas Ravindra, Prof. R. V. Varhade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Rising electricity demand and the spread of smart home technology create a need for intelligent, sustainable residential energy management. Traditional approaches rely on fixed schedules and predefined rules that respond poorly to changing prices, demand, weather and user requirements. Reinforcement Learning (RL) offers a different approach in which an agent learns optimal control decisions by interacting with the home environment and receiving feedback. This paper reviews RL for smart-home energy management with the goals of reducing consumption and cost while maintaining comfort. It introduces Markov decision processes and the main RL algorithms, from tabular Q-learning to deep Q-networks, proximal policy optimization and soft actor-critic, then surveys applications to demand response, appliance scheduling, heating, ventilation and air conditioning control, electric-vehicle charging, rooftop solar and battery storage. Simulation environments for training and benchmarking agents are discussed. A five-layer architecture integrating IoT sensing and smart meters, state construction with forecasting, an RL decision engine, device control and a user interface with comfort preferences is proposed, together with research gaps, objectives, a comparison of algorithms and challenges such as safety, sample efficiency, comfort modelling and transfer between homes. The review concludes that RL can adaptively optimize household energy use when combined with safety constraints and user feedback.

DOI: https://doi.org/10.5281/zenodo.24102915

Smart Tourist Safety Monitoring Using IoT And Artificial Intelligence: A Review Of Location Tracking, Risk Detection And Emergency Response

Authors: Deore Adarsh Sunil, Prof. M. A. Auti, Prof. M. C. Kolse, Prof. S. K. Said

DOI: https://doi.org/10.5281/zenodo.24103595

Multimodal AI With Wearable And IoT Biosensors: A Review Of PPG, ECG And Temperature Fusion On ESP32 Platforms For Continuous Health Monitoring

Authors: Vishwe Archana Sanjay, Prof. S. S. Bhor, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Advances in artificial intelligence, the Internet of Things and wearable technology enable continuous health monitoring instead of periodic examinations that may miss physiological changes. This paper reviews multimodal AI health monitoring based on low-cost biosensors. It examines photoplethysmography for heart rate and blood oxygen saturation, single-lead electrocardiography for cardiac rhythm, and body temperature sensing, and describes how an ESP32 microcontroller acquires these signals and transmits them over Wi-Fi to a computer or cloud service. Signal processing steps including filtering, QRS detection and feature extraction are reviewed together with public physiological databases used to develop and validate algorithms. Machine learning methods such as random forest and logistic regression for classifying normal and potentially abnormal states are compared with deep learning on raw signals, and multimodal fusion strategies, early, late and hybrid, are discussed in the context of multimodal machine learning. Dashboard design for displaying live readings, history and threshold alerts is described. A five-layer architecture comprising sensing, acquisition, communication, multimodal AI and dashboard layers is proposed, along with research gaps, objectives, a comparison of sensors and challenges including motion artefacts, sensor accuracy limitations such as bias in pulse oximetry, battery life, privacy and the need for clinical validation. The review concludes that multimodal AI with wearable biosensors supports timely observation of physiological change under professional supervision.

DOI: https://doi.org/10.5281/zenodo.24104130

Artificial Intelligence In E-Commerce: A Review Of Personalization, Dynamic Pricing, Visual Search, Fraud Detection And Agentic Commerce

Authors: Shelot Raj Satish, Prof. S. R. Borhade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Artificial intelligence is transforming e-commerce by changing how businesses attract, engage and retain customers. This paper reviews AI applications across the e-commerce value chain. It examines personalized recommendation, from item-to-item collaborative filtering to deep click-through-rate models that capture changing user interests; dynamic and algorithmic pricing and its effects on marketplaces; demand forecasting and price optimization for inventory planning; visual and voice search; conversational chatbots and assistants for customer support; and fraud detection on payments and accounts. Using machine learning, natural language processing and computer vision, these technologies analyse customer behaviour and transactions in real time, increasing conversion, order value and satisfaction while reducing operational cost and losses. The paper also discusses emerging trends such as generative AI for product content, conversational commerce, hyper-personalization and autonomous shopping agents. A five-layer e-commerce AI architecture comprising data, customer intelligence, decision, engagement and trust layers is proposed, along with research gaps, objectives, a comparison of applications and challenges including privacy, algorithmic bias, transparency, implementation cost and skills. The review concludes that long-term success depends on responsible and transparent AI that protects consumer trust.

DOI: https://doi.org/10.5281/zenodo.24104429

Hallucination In Large Language Models: A Review Of Causes, Taxonomy, Detection Methods And Mitigation Strategies

Authors: Fulawade Harsh Nitin, Prof. N. G. Rokade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Large Language Models (LLMs) generate human-like text, answer questions and summarize information, but they sometimes produce incorrect, misleading or fabricated content while presenting it as fact. This phenomenon, known as hallucination, reduces the accuracy, trustworthiness and safety of generative AI systems, particularly in education, healthcare, law and finance. This paper reviews LLM hallucination in a structured way. It explains causes arising from training data, model architecture and training objectives, decoding and missing context, and presents a taxonomy that distinguishes factuality from faithfulness errors. Benchmarks that measure truthfulness and hallucination rates are discussed. Detection methods are reviewed, including self-consistency checks across sampled answers, uncertainty measures such as semantic entropy and fine-grained atomic fact verification. Mitigation strategies are examined across the model life cycle: data curation, alignment with human feedback, retrieval-augmented generation, prompting for verification and abstention when uncertain. A layered hallucination-aware LLM application architecture is proposed, along with research gaps, objectives and a comparison of methods. The review concludes that hallucination cannot be eliminated entirely but can be detected, reduced and communicated so that users can rely on LLM outputs appropriately.

DOI: https://doi.org/10.5281/zenodo.24104945

Smart Bus Tracking And Monitoring System: A Review Of GPS–IoT Vehicle Tracking, Arrival-Time Prediction And Passenger Information Services

Authors: Khandagale Pratiksha Mangesh, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Public bus services carry large numbers of passengers in Indian cities, yet passengers often wait at stops without knowing when the next bus will arrive, and operators lack real-time visibility of their fleets. A smart bus tracking and monitoring system uses GPS, IoT communication and web or mobile applications to provide live bus locations, routes, estimated arrival times and schedules to passengers and transport authorities. This paper reviews the technologies behind such systems. It covers GPS-based vehicle tracking, onboard IoT units with cellular communication, trajectory processing and map matching, and methods for predicting bus arrival times, from historical averages and Kalman filtering to artificial neural networks and recurrent models that account for traffic and multiple routes. Smartphone-based crowdsourced tracking and research on the effect of real-time information on passengers are reviewed, together with open data standards. A five-layer architecture comprising onboard units, communication, a central server with prediction, administrator tools and passenger applications is proposed. Research gaps, objectives, a comparison of prediction methods and challenges including GPS errors in dense areas, data costs and privacy are discussed. The review concludes that integrating AI-based prediction makes bus services more reliable and attractive.

DOI: https://doi.org/10.5281/zenodo.24105259

Explainable Artificial Intelligence (XAI) In The Banking Sector: A Review Of Techniques For Transparent Credit, Fraud And Compliance Decisions

Authors: Pingale Neha Pravin, Prof. M. R. Kumbhoje, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Banks increasingly use complex machine learning models such as neural networks, random forests and gradient boosting for credit scoring, loan approval, fraud detection, risk assessment and anti-money laundering. These models process sensitive financial data and directly affect people's access to credit and other essential services, yet many operate as black boxes that cannot explain why a loan was rejected or a transaction flagged. This raises concerns about fairness, bias, accountability and compliance with regulations that require meaningful justification of automated decisions. This paper reviews Explainable Artificial Intelligence (XAI) for banking. It examines why explainability is needed, then surveys inherently interpretable models such as logistic regression and decision trees, post-hoc techniques including LIME, SHAP, accumulated local effects and feature importance, and counterfactual explanations that tell customers what would change a decision. Applications in credit risk, fraud detection and customer decision support are discussed, together with fairness metrics and regulatory frameworks including the GDPR, the EU AI Act, model risk management guidance and the Reserve Bank of India's responsible AI framework. A layered XAI architecture for banking is proposed, with research gaps, objectives and a comparison of techniques. The review concludes that explainability is essential for transparent, ethical and reliable AI-driven banking.

DOI: https://doi.org/10.5281/zenodo.24105564

Small Language Models For Edge AI: A Review Of Efficient Architectures, Compression Techniques And On-Device Deployment

Authors: Dawkhar Vaishnavi Ganpat, Prof. M. A. Auti, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Small Language Models (SLMs) bring language understanding and generation directly to resource-constrained devices such as smartphones, wearables, IoT nodes and embedded systems. Large language models demand substantial memory, energy and connectivity, which limits their use in latency-sensitive, privacy-critical and offline settings. This paper reviews SLMs for Edge AI. It examines how compute-optimal scaling and high-quality training data enable compact models to perform well, and surveys contemporary open SLM families. Compression techniques are analysed in depth, including knowledge distillation, post-training quantization to four and eight bits, activation-aware and outlier-smoothing quantization, one-shot pruning, and parameter-efficient fine-tuning with low-rank adapters. Inference optimizations such as memory-efficient attention and speculative decoding are discussed. Trade-offs among accuracy, model size, inference latency and power consumption are compared. A lightweight deployment pipeline is proposed that selects a base model, adapts it to the task, compresses it and benchmarks it on target hardware before deployment. Research gaps, objectives and challenges including reasoning limits, hallucination, evaluation and secure updates are discussed. The review concludes that SLMs make private, low-latency, offline-capable language intelligence practical at the edge.

DOI: https://doi.org/10.5281/zenodo.24105854

Smart Greenhouse Using Ai And Iot: A Review Of Climate Sensing, Automated Control And Predictive Microclimate Management

Authors: Sandbhor Akanksha Somnath, Prof. M. R. Kumbhoje, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: A smart greenhouse is an IoT-based automated agriculture system that maintains suitable conditions for plant growth while reducing water, energy and labour. Conventional greenhouses depend on continuous human monitoring and manual switching of pumps, fans and lights, which wastes resources and allows stress conditions to go unnoticed. This paper reviews AI- and IoT-enabled smart greenhouses. It examines sensing of temperature, humidity, soil moisture, light intensity and carbon dioxide; microcontroller platforms such as Arduino and ESP32 that drive irrigation pumps, exhaust fans, foggers and grow lights; and communication through Wi-Fi and MQTT to cloud dashboards and mobile applications for remote monitoring. Beyond fixed-threshold control, it reviews machine learning that predicts the greenhouse microclimate a few hours ahead, recurrent networks for time-series forecasting, and AI-based climate and irrigation strategies that have competed with expert growers. A five-layer architecture comprising sensing, edge control, connectivity, cloud intelligence and user application is proposed, along with research gaps, objectives, a comparison of control strategies and challenges such as sensor drift, power failures, energy cost and crop-specific tuning. The review concludes that predictive, AI-assisted control can make greenhouses more productive and resource-efficient than threshold-based automation alone.

DOI: https://doi.org/10.5281/zenodo.24106176

IoT-Based Smart Home Automation: A Review Of Architectures, Microcontroller Platforms, Communication Protocols And Security

Authors: Raut Riya Prakash, Prof. M. R. Kumbhoje, Prof. M. C. Kolse, Prof. S. K. Said

DOI: https://doi.org/10.5281/zenodo.24107138

AI-Based River Chemical Pollution Monitoring And Cleaning: A Review Of Sensor Networks, Water Quality Prediction And Automated Treatment

Authors: Mule Sahil Vijay, Prof. J. M. Patade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Industrial effluents, agricultural runoff, sewage and hazardous chemicals contaminate rivers, threatening human health, aquatic ecosystems and the communities that depend on river water. Conventional water-quality monitoring relies on periodic manual sampling and laboratory analysis, which is slow, expensive and cannot provide continuous real-time information. This paper reviews an AI-based approach to river chemical pollution monitoring and cleaning. It covers wireless sensor networks that measure pH, temperature, turbidity, dissolved oxygen, conductivity and specific chemical concentrations; machine learning and deep learning models that classify pollution levels, compute water quality indices, predict future contamination and detect anomalous discharge events; and robotic and automated treatment technologies, including autonomous surface vehicles and chemical treatment processes, that respond to detected pollution. Water quality indices and their limitations are discussed. A five-layer architecture comprising in-river sensing, communication, AI analytics, a treatment decision and actuation layer, and an authority dashboard is proposed, together with research gaps, objectives, a comparison of prediction models and challenges such as sensor fouling, calibration, power supply, ecological side effects of treatment and enforcement. The review concludes that AI-driven continuous monitoring can detect and respond to river pollution far earlier than manual methods.

DOI: https://doi.org/10.5281/zenodo.24107562

Smart Security Cameras Using IoT: A Review Of Raspberry Pi Video Streaming, OpenCV-Based Detection, Face Recognition And Secure Alerting

Authors: Shinde Karan Laxman, Prof. S. R. Borhade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: IoT-based smart security cameras offer intelligent, automated monitoring at low cost, making them attractive for homes and small businesses. This paper reviews smart security camera systems built on single-board computers such as the Raspberry Pi with a Pi Camera module, which stream video over the local network and to remote users through a web interface. It examines the computer vision pipeline, from classical motion detection by background subtraction and pedestrian detection with histogram-of-oriented-gradients features to deep-learning object detectors and lightweight networks for embedded devices, as well as face recognition based on learned embeddings for distinguishing known people from strangers. Edge processing for real-time video analytics, alert mechanisms such as automated e-mail and push notifications with captured snapshots, and storage of event clips are discussed. The security of the camera itself is reviewed, since poorly configured IoT devices are routinely compromised. A five-layer architecture comprising capture, edge vision, event management, secure communication and user interface layers is proposed, together with research gaps, objectives, a comparison of detection methods and challenges such as false alarms, low light, privacy and limited computing power. The review concludes that edge-based intelligent cameras can provide affordable, responsive and privacy-conscious residential security.

DOI: https://doi.org/10.5281/zenodo.24108020

The Role Of Artificial Intelligence In Building A Smart Future: A Review Of Applications, Enablers, Risks And Responsible Adoption

Authors: Some Sanjana Sanjay, Prof. S. R. Borhade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Artificial intelligence is becoming a general-purpose technology that makes systems more intelligent, efficient and capable of supporting human activities. This paper reviews the role of AI in building a smarter, safer and more sustainable future. It traces the foundations of modern AI in deep learning and transformer-based foundation models, and examines applications in smart cities and mobility, healthcare, education, agriculture, industry, scientific discovery, environmental monitoring and climate action, and robotic assistance in dangerous tasks. Evidence on AI's potential contribution to the Sustainable Development Goals is discussed, together with its possible negative effects. The paper then reviews the enablers needed for wide adoption, including data, computing, connectivity, skills and digital public infrastructure, and the challenges of limited public awareness, the need for high-quality data, development and maintenance cost, bias, privacy and misuse. Global ethics guidelines and regulation, as well as India's national AI strategy, are examined as frameworks for responsible use. A five-layer framework for a smart future comprising foundations, AI capabilities, sectoral applications, governance and societal outcomes is proposed, along with research gaps, objectives and a comparison of sectors. The review concludes that AI can transform human life for the better only through continued innovation combined with responsible application.

DOI: https://doi.org/10.5281/zenodo.24112523

Student Performance Prediction Using Machine Learning: A Review Of Educational Data Mining Methods For Early Intervention

Authors: Auti Kranti Somnath, Prof. S. K. Said, Prof. M. C. Kolse

Abstract: Predicting student performance is a central application of educational data mining and learning analytics, aimed at forecasting academic outcomes and identifying at-risk students early enough for teachers to intervene. This paper reviews machine learning approaches to the problem. It examines the data sources used, including prior grades, attendance, assignment submissions, learning-management-system activity and demographic factors, and the models applied, from decision trees, random forests and support vector machines to knowledge tracing and deep sequence models. Public datasets such as the Portuguese secondary-school data and the Open University Learning Analytics Dataset are discussed, together with class imbalance, feature engineering, evaluation metrics and fairness concerns. Real-world early-warning systems that combine prediction with student feedback are reviewed. A five-layer architecture comprising data integration, preprocessing, predictive modelling, explanation and intervention is proposed, with research gaps, objectives and a comparison of algorithms. The review concludes that predictive models are most useful when their outputs are explained, checked for bias and paired with supportive interventions that improve retention and personalized learning.

DOI: https://doi.org/10.5281/zenodo.24113183

Fine-Tuning Large Language Models: A Review Of Full, Parameter-Efficient And Instruction Tuning For Domain-Specific Applications

Authors: Musale Devendra Kailas, Prof. J. M. Patade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Large Language Models (LLMs) understand and generate human-like text, but general-purpose models often lack the vocabulary, knowledge and response style needed for specific domains such as healthcare, law, agriculture or customer support. Fine-tuning adapts a pre-trained model to a target task by continuing training on domain-specific data, which is far cheaper than training from scratch. This paper reviews fine-tuning methods for LLMs. It traces transfer learning from universal language model fine-tuning to text-to-text models and explains full fine-tuning, parameter- efficient fine-tuning with adapters, prefix and prompt tuning and low-rank adaptation, quantized fine-tuning, and instruction and preference tuning that align models with human expectations. Tools such as Hugging Face Transformers and PyTorch, and models such as BERT, GPT-style and LLaMA, are discussed. A five-stage proposed pipeline covering dataset collection, preprocessing, model selection, fine-tuning and evaluation is presented as a research design; no implementation results or accuracy claims are made. Research gaps, objectives, a comparison of fine-tuning methods and challenges including data quality, catastrophic forgetting, overfitting, compute cost, evaluation and safety are discussed. The review concludes that parameter-efficient methods make domain adaptation accessible to students and small organizations.

DOI: https://doi.org/10.5281/zenodo.24114144

Large Language Models: The Core Of Generative AI — A Review Of Architecture, Training, Capabilities And Limitations

Authors: Nimse Vidya Kisan, Prof. J. M. Patade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Large Language Models (LLMs) have become the core technology of Generative Artificial Intelligence, powering chat assistants, writing and coding tools, search and many other applications. This paper reviews how LLMs work and why they matter. It explains the transformer architecture with self-attention, tokenization into subword units and the decoder-only design used by generative models, then traces the development from generative pre-training through models that perform tasks from a few examples. Scaling laws and compute-optimal training explain why larger models trained on more data perform better, and emergent abilities and chain-of-thought prompting show how capabilities appear with scale. Alignment through instruction tuning and reinforcement learning from human feedback, and grounding through retrieval, are reviewed. Open model families are discussed. A five-stage LLM lifecycle architecture, comprising data, pre-training, alignment, augmentation and application, is proposed. Research gaps, objectives, a comparison of model families and limitations including hallucination, bias, cost, environmental impact, privacy and misuse are analysed. The review concludes that LLMs are a foundational general-purpose technology whose benefits depend on responsible development, grounding and human oversight.

DOI: https://doi.org/10.5281/zenodo.24114810

AI In Smart Agriculture: A Review Of AI-Based Soil Analysis And Fertilizer Management For Precision Agriculture

Authors: Sherkar Manisha Minanath, Prof. S. R. Borhade, Prof. M. C. Kolse, Prof. S. K. Sai

Abstract: Imbalanced fertilizer use is a major problem in Indian agriculture: excess nitrogen raises costs and pollutes water, while deficiencies of phosphorus, potassium and micronutrients limit yields. Effective fertilizer management requires knowing the nutrient status of each field, which conventional laboratory soil testing provides only slowly and at low spatial density. This paper reviews AI-based soil analysis and fertilizer management for precision agriculture. It examines soil sensing with in-field electrochemical NPK, pH and electrical-conductivity probes and with visible–near-infrared spectroscopy; machine learning models that classify soil fertility and predict nutrient levels from sensor and spectral data; digital soil mapping, which combines field samples with remote sensing and terrain covariates to map soil properties across landscapes; crop nitrogen status estimation; and recommendation engines that convert soil and crop information into site-specific fertilizer doses and variable-rate application. India's Soil Health Card programme is discussed as a data source and delivery channel. A five-layer architecture comprising soil sensing, data integration, AI analysis, fertilizer recommendation and farmer advisory layers is proposed, together with research gaps, objectives, a comparison of soil analysis methods and challenges such as sensor accuracy, sparse samples, model transferability, cost and farmer adoption. The review concludes that AI can make fertilizer use more efficient, profitable and environmentally sustainable.

DOI: https://doi.org/10.5281/zenodo.24115261

Secure File Encryption Systems Using AES-256-GCM And SHA-256: A Review Of Authenticated Encryption, Key Management And Integrity Verification

Authors: Bhumkar Pooja Dnyaneshwar, Prof. M. A. Auti, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Files containing academic records, financial statements, medical reports and business documents are routinely stored on laptops, removable drives and cloud folders, where they are exposed to theft, unauthorized access and silent tampering. This paper reviews the design of a secure file encryption system built on modern, standardized cryptography. The Advanced Encryption Standard with a 256-bit key in Galois/Counter Mode (AES-256-GCM) is examined as an authenticated encryption scheme that provides confidentiality and authenticity in a single operation, using a unique nonce for every encryption. The review explains why password-derived keys must be generated with a slow, salted key-derivation function such as PBKDF2 or Argon2, how the SHA-256 hash supports file integrity verification and identification, and why weaker primitives such as MD5 and SHA-1 should be avoided. A layered architecture covering the user interface, key management, cryptographic engine, integrity module and secure storage is proposed, together with an encrypted file format and a step-by-step encryption and decryption workflow. The paper also compares alternative ciphers and modes, discusses nonce misuse, key loss and post-quantum considerations, and identifies research gaps for practical tools. The review concludes that correctly combining authenticated encryption, strong key derivation and integrity checks yields a simple yet robust solution for protecting confidential files.

DOI: https://doi.org/10.5281/zenodo.24115569

AI-Based Smart Parking And Traffic Prediction Systems: A Review Of Occupancy Detection, Availability Forecasting And Spatio-Temporal Traffic Models

Authors: Pawale Gauri Sandip, Prof. M. R. Kumbhoje, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Urban growth and rising vehicle numbers have created severe traffic congestion and shortages of parking space, leading to longer travel times, higher fuel consumption, pollution and driver frustration; a significant share of urban traffic consists of drivers searching for parking. This paper reviews AI-based smart parking and traffic prediction systems. It examines parking occupancy detection with in-ground sensors and with camera images classified by deep convolutional networks on public parking datasets, prediction of future parking availability with spatio-temporal models, and traffic flow forecasting from classical regression and decision trees to deep learning and graph neural networks that model road networks. The roles of cloud computing and mobile applications in guiding drivers are discussed. A five-layer architecture combining parking and traffic sensing, data integration, an AI prediction engine, a recommendation and guidance layer and user and authority applications is proposed, with research gaps, objectives, a comparison of techniques and challenges such as sensor cost, data quality, privacy and integration with existing infrastructure. The review concludes that integrating parking availability with traffic prediction can reduce cruising, waiting time and emissions while improving the use of parking facilities.

DOI: https://doi.org/10.5281/zenodo.24115926

Agentic AI In Healthcare: A Review Of Goal-Oriented Healthcare Assistants With Symptom Guidance, Medication Adherence And Safe Escalation

Authors: Unde Rutuja Motiram, Prof. S. S. Bhor, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Healthcare systems increasingly use artificial intelligence to improve patient support, reduce manual effort and assist decision-making. Agentic AI differs from systems that only answer queries: an agent pursues goals, plans multi-step actions, uses tools such as databases and reminders and coordinates several functions to help a user manage health over time. This paper reviews agentic AI for consumer healthcare assistance. It examines the components of large-language-model agents, namely planning, tool use, memory and reflection, and the use of retrieval-augmented generation to ground answers in trusted medical sources. It reviews evidence on the clinical knowledge of large language models, their limitations including hallucination, studies of the accuracy and safety of symptom checkers, conversational diagnostic systems, and interventions that improve medication adherence. A reference architecture for a healthcare assistant is proposed with five layers: user and authentication, agent core with planning and memory, tool layer providing symptom guidance, medicine reminders, report storage and emergency information, a safety and escalation layer enforcing scope limits and red-flag detection, and an interface layer. Research gaps, objectives, a comparison of assistant capabilities and challenges such as accuracy, liability, privacy and digital literacy are discussed. The review emphasizes that such assistants must provide preliminary information only and direct users to qualified professionals.

DOI: https://doi.org/10.5281/zenodo.24116115

Ai For Plant Disease Detection: A Systematic Review Of Machine Learning, Deep Learning And Vision Transformers With Field Deployment And Dataset Bias

Authors: Yeola Tanishka Manoj, Prof. S. S. Bhor, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Crop diseases remain a serious threat to food security, causing substantial yield losses and economic hardship for farmers. Traditional detection depends on manual field inspection by experts, which is slow, subjective and often unavailable in rural regions. This paper systematically reviews AI-based approaches for crop disease prediction. It covers classical machine learning algorithms such as support vector machines, random forests and k-nearest neighbours applied to colour, texture and shape features, and deep learning architectures including convolutional neural networks, transfer-learning models and vision transformers. The complete pipeline is discussed: image acquisition through smartphones, drones and IoT sensors; preprocessing and segmentation; feature extraction; classification; and deployment on mobile and edge platforms using lightweight architectures and quantization. Commonly used datasets including PlantVillage are examined together with evidence of their background bias and the resulting gap between laboratory and field accuracy, and field datasets such as PlantDoc and the Plant Pathology challenge. Evaluation metrics of accuracy, precision, recall and F1-score are reviewed, as is the integration of environmental parameters such as temperature, humidity and soil conditions with image data for forecasting. A five-stage architecture is proposed, with research gaps, objectives, a comparison of techniques and challenges including limited real-field data, class imbalance, lighting variation, computational cost and poor generalization. Future directions identified are explainable AI, lightweight models, multimodal fusion and federated learning.

DOI: https://doi.org/10.5281/zenodo.24214661

Artificial Intelligence In Railway Systems: A Review Of Track Inspection, Rolling-Stock Health Monitoring, Delay Prediction And Intelligent Scheduling

Authors: Pawar Aditya Mahendra, Prof. M. R. Kumbhoje, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Railways move enormous numbers of passengers and freight, and their safety and punctuality depend on the condition of tracks, rolling stock and signalling and on effective scheduling. Manual inspection and fixed maintenance cycles struggle with the scale of modern networks and with the need to detect defects before they cause accidents or delays. This paper reviews Artificial Intelligence applications across railway systems. It examines deep convolutional networks and object detectors for automated track and rail-surface defect detection from inspection images, recurrent networks for vibration-based condition monitoring of axle-box bearings and other rolling-stock components, big-data analytics and machine learning for train delay prediction, and optimization and learning methods for real-time rescheduling after disruptions. Automatic train protection, including India's Kavach system, and edge deployment of compressed models on inspection vehicles are discussed. A five-layer architecture comprising multimodal sensing, edge inference, a railway data platform, AI decision support and operations control is proposed, with research gaps, objectives, a comparison of applications and challenges including harsh environments, rare failure data, safety certification and integration with legacy systems. The review concludes that AI can make railways safer, more reliable and more efficient when deployed within rigorous safety frameworks.

DOI: https://doi.org/10.5281/zenodo.24214730

Ai-Based Smart Agriculture: A Review Of Edge Disease Diagnosis, Soil Nutrient Profiling And Reinforcement Learning For Precision Irrigat

Authors: Pawar Ravi Santosh, Prof. M. R. Kumbhoje, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Agriculture faces pressure from changing climate, water scarcity, soil degradation and crop diseases, while many farmers still decide when to irrigate, how much fertilizer to apply and how to treat diseases by experience alone. This paper reviews AI-based smart agriculture with an integrated, field-deployable perspective. It examines IoT sensing of soil moisture, temperature, humidity and nutrients; deep convolutional networks, including efficient architectures such as EfficientNet, for multi-class foliar disease diagnosis; machine learning for soil macronutrient (nitrogen, phosphorus, potassium) classification and fertilizer guidance; and machine learning and reinforcement learning for precision irrigation scheduling based on evapotranspiration and weather forecasts. Deployment on resource-constrained edge gateways using model quantization is discussed so that diagnoses and irrigation decisions do not depend on continuous internet access. A five-layer architecture comprising field sensing, edge intelligence, a farm decision engine, actuation and a farmer advisory application is proposed, together with research gaps, objectives, a comparison of AI components and challenges such as cost, data scarcity, connectivity, farmer adoption and sustainability. The review concludes that integrated AI systems can raise yields while saving water, fertilizer and pesticide.

DOI: https://doi.org/10.5281/zenodo.24214845

Phishing Attacks: A Review Of Types, Attack Techniques And Prevention Strategies

Authors: Rajguru Tejas Prakash, Prof. M. R. Kumbhoje, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Phishing remains one of the most common entry points for cyberattacks, tricking people into revealing credentials, installing malware or transferring money. Phishing has diversified far beyond mass emails into targeted spear phishing, executive whaling, SMS smishing, voice vishing, clone phishing, malicious websites and social-media lures. This paper reviews phishing from three perspectives. First, it classifies phishing types by channel and targeting. Second, it examines attacker techniques, including social engineering using information from social networks, look-alike domains through typosquatting and homograph characters, cloned websites, and modern adversary-in-the-middle toolkits that intercept credentials and one-time codes to bypass multi factor authentication, along with evidence on the lifecycle and effectiveness of real campaigns. Third, it reviews prevention strategies across technical controls, such as email authentication, phishing-resistant authentication, browser warnings and detection systems, and human measures, such as training and organizational reporting cultures. A five-layer defence-in-depth framework is proposed, together with research gaps, objectives, a comparison of authentication methods and challenges. The review concludes that no single control stops phishing, but layered technical and human defences greatly reduce its success.

DOI: https://doi.org/10.5281/zenodo.24215042

Ai Agents For Autonomous Network Management: A Review Of Self-Monitoring, Self-Optimizing, Self- Healing And Self-Protecting Networks

Authors: Raskar Kasturi Ramesh, Prof. M. R. Kumbhoje, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Modern enterprise, cloud and mobile networks are large, complex and dynamic, and manual monitoring and configuration are slow, error-prone and unable to react quickly to failures, congestion and attacks. AI agents offer an intelligent approach in which software entities use machine learning, reasoning and automation to monitor, analyse and manage network operations with minimal human intervention. This paper reviews AI agents for autonomous network management. It traces the concept from the autonomic computing vision and its monitor–analyse–plan–execute loop, through software-defined networking and intent-based networking, to industry frameworks for zero-touch and autonomous networks. Machine learning techniques for traffic prediction, anomaly and fault detection, reinforcement learning for routing and resource management, graph neural networks for performance modelling, and language-model agents for configuration and troubleshooting are examined. Future directions such as 5G and 6G network automation, open radio access network controllers and network digital twins are discussed. A layered architecture with data collection, analytics, agent decision, orchestration and intent interfaces is proposed, together with research gaps, objectives, a comparison of techniques and a discussion of challenges including trust, safety of automated actions and data quality. The review concludes that AI agents can make networks more adaptive, reliable and secure while reducing operational cost.

DOI: https://doi.org/10.5281/zenodo.24215184

Artificial Intelligence In Data Analysis: A Review Of Automated Data Preparation, Predictive Analytics, Natural-Language Querying And Explainability

Authors: Shelke Rutvik Santosh, Prof. S. R. Borhade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: The volume, speed and variety of data produced by digital systems, sensors and online platforms have outgrown traditional analytical methods. Artificial intelligence enables organizations to extract insights from large, complex and unstructured data with greater speed and accuracy. This paper reviews how AI techniques, particularly machine learning, deep learning and natural language processing, are applied across the data analysis pipeline: automated data cleaning and preparation, automated machine learning for model and feature selection, pattern recognition and predictive modelling with ensemble methods, anomaly detection in transactions and sensor streams, and natural-language interfaces that translate questions into database queries so that non technical users can explore data conversationally. Applications in healthcare, finance and retail are discussed. The review also examines explainable AI methods that attribute predictions to input features, and the problems of data quality, bias and fairness. A five-layer AI-assisted analytics architecture comprising data ingestion, preparation, modelling, explanation and conversational insight layers is proposed, together with research gaps, objectives, a comparison of traditional and AI-driven analysis and challenges such as privacy, compute cost and skill shortages. The review concludes that AI augments rather than replaces human analysts, freeing them for interpretation and strategic decisions.

DOI: https://doi.org/10.5281/zenodo.24216130

Ai-Based Elderly Care And Fall Detection: A Review Of Wearable, Ambient And Vision Sensing With Machine Learning And Iot Alerting

Authors: Shinde Sakshi Jaysing, Prof. S. R. Borhade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Falls are a leading cause of injury among older adults, and a person who falls while living alone may remain on the floor for a long time before help arrives. AI-based elderly care and fall detection systems monitor movement and alert caregivers when a fall is likely. This paper reviews such systems. It examines wearable approaches using accelerometers and gyroscopes with threshold and machine learning classifiers; ambient approaches using floor, radar and Wi-Fi channel sensing that require nothing to be worn; and vision approaches using cameras and human pose estimation. Public datasets of simulated falls and activities of daily living, deep learning models such as CNNs and LSTMs on motion sequences, and IoT alerting through mobile applications are discussed, along with activity monitoring and basic health tracking within ambient-assisted living. A five-layer architecture comprising sensing, edge inference, connectivity, care platform and caregiver application layers is proposed, together with research gaps, objectives, a comparison of sensing approaches and challenges such as false alarms, missed falls, battery life, connectivity, privacy and the gap between simulated and real-world falls. The review emphasizes that such systems support but do not replace medical professionals and emergency services, and concludes that combining sensing modalities with on-device AI can provide timely, reliable and acceptable support for independent living.

DOI: https://doi.org/10.5281/zenodo.24216390

Artificial Intelligence In Industries: A Review Of Predictive Maintenance, Automated Quality Inspection, Demand Forecasting And Industry 4.0 Adoption

Authors: Sandbhor Atharva Vikas, Prof. S. R. Borhade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Artificial intelligence is changing how industries design, produce, maintain and deliver products. This paper reviews AI in industrial operations within the Industry 4.0 paradigm, where cyber-physical systems, the industrial Internet of Things and data analytics connect machines, processes and people. It examines four major application areas: predictive maintenance, which uses vibration, temperature and current signals to estimate machine health and remaining useful life; automated visual quality inspection, which uses deep learning and unsupervised anomaly detection to find defects; demand and inventory forecasting with probabilistic deep learning models; and digital twins that mirror physical assets for monitoring and optimization. Supporting topics such as robotics, workplace-safety monitoring and the cross-domain use of AI for information verification and security are discussed briefly. A five-layer industrial AI architecture comprising the shop-floor, industrial IoT, data platform, AI services and enterprise decision layers is proposed, together with research gaps, objectives, a comparison of application areas and challenges such as data quality, legacy equipment, cybersecurity, skill shortages, cost and over-reliance on automation. The review concludes that AI delivers the greatest value when it is integrated with domain knowledge and human oversight.

DOI: https://doi.org/10.5281/zenodo.24214956

Phishing Attack Detection And Cybersecurity Defense Mechanisms: A Review Of Url, Content And Visual Machine Learning Approaches

Authors: Suryavanshi Aditi Krushnakant, Prof. S. R. Borhade, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: —Phishing attacks use fake emails, messages and websites to trick people into revealing passwords, banking details and other personal data, causing financial losses and privacy violations. Attackers continuously create new sites and adapt their techniques, so blocklists and static rules often fail against advanced and short lived phishing campaigns. This paper reviews phishing detection with emphasis on website and URL analysis, and places detection within a broader defence-in-depth strategy. It surveys feature-based approaches that analyse URL structure, domain information, page content and HTML elements, including the influential phishing website feature set of Mohammad et al., and machine learning classifiers such as Random Forest, Decision Tree, Support Vector Machine and Logistic Regression. Deep learning on raw URL strings and visual similarity methods that compare page screenshots and logos against legitimate brands are examined. Defence mechanisms beyond detection, including browser warnings, email authentication, phishing-resistant multi-factor authentication and user awareness training, are discussed. A layered detection and defence architecture is proposed, along with research gaps, objectives and a comparison of algorithms. The review concludes that combining automated multi view detection with resistant authentication and informed users provides the most effective protection against phishing.

DOI: https://doi.org/10.5281/zenodo.24216546

Ai In Education: A Review Of Multimodal Student Engagement Detection, Learning Analytics And Adaptive Intervention Systems

Authors: Tavhare Pranjal Pravin, Prof. S. S. Bhor, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Student engagement is closely linked to achievement, retention and the quality of learning, yet teachers in large physical classrooms and online courses cannot easily monitor every learner in real time. This paper reviews AI-driven student engagement systems for classroom and online settings. It begins with the educational view of engagement as behavioural, emotional and cognitive, and then examines automatic engagement recognition from facial expressions and head pose, datasets of learners recorded in the wild, the dynamics of affective states such as confusion and boredom during learning, and learning analytics from activity logs, attendance and quiz performance, including studies of disengagement patterns in massive open online courses and knowledge tracing. Machine learning and deep learning methods such as CNNs, random forests and SVMs for classifying engagement as high, moderate or low are compared using accuracy, precision and recall. An adaptive recommendation module that suggests content and interventions and a teacher dashboard are discussed. A five-layer architecture comprising multimodal data, feature extraction, engagement classification, adaptive intervention and teacher analytics layers is proposed, together with research gaps, objectives, a comparison of data modalities and challenges such as privacy, consent, bias and scalability. The review concludes that ethical, teacher-centred AI can support more inclusive and student-centred education.

DOI: https://doi.org/10.5281/zenodo.24216799

Ai In Healthcare: A Review Of Smart Health Monitoring With Iot Wearables, Early Warning Scores And Predictive Deterioration Analytics

Authors: Thorat Pradnya Ajit, Prof. S. S. Bhor, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: —Continuous monitoring of patient health is essential for early diagnosis and timely treatment, but conventional care depends on periodic manual checks that can miss critical changes between rounds. This paper reviews AI-based smart health monitoring for hospitals and home care. It examines IoT-enabled wearable sensors and medical devices that measure heart rate, temperature, blood oxygen saturation, blood pressure and activity; clinical early warning scores that convert vital signs into a risk score; and machine learning models that predict clinical deterioration more accurately than conventional regression, along with deep learning on electrocardiograms for arrhythmia detection, continuous prediction of acute kidney injury from electronic health records and large-scale smartwatch studies of atrial fibrillation. Public critical-care databases that enable model development are discussed. CNNs, random forests and SVMs for anomaly detection and classification of patient status as normal, warning or critical are compared using accuracy, precision and recall. A five-layer architecture comprising sensing, edge processing, secure communication, clinical analytics and care-team dashboard layers is proposed, together with research gaps, objectives, a comparison of monitoring approaches and challenges such as alarm fatigue, data quality, privacy, interoperability and clinical validation. The review concludes that AI assisted monitoring can enable earlier intervention when designed to support clinicians.

DOI: https://doi.org/10.5281/zenodo.24217444

Artificial Intelligence For Software Testing: A Review Of Automated Test Generation, Defect Prediction, Test Prioritization And LLM-Assisted Testing

Authors: Varhadi Shreya Sachin, Prof. S. S. Bhor, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Software testing consumes a large share of development effort, and manual test design cannot keep pace with frequent releases. Artificial intelligence and machine learning can automate test case generation, predict where defects are likely, prioritize which tests to run and help analyse results. This paper reviews AI for software testing. It examines search-based and feedback-directed automated test generation, including evolutionary generation of unit test suites and random testing guided by execution feedback, and whitebox fuzzing that uses symbolic execution to reach deep program paths. It reviews defect prediction from code metrics and change history, systematic evidence on fault-prediction performance, and reinforcement learning for test case prioritization and selection in continuous integration, where the goal is to find failures early within a limited time budget. It also reviews the use of large language models trained on code to generate unit tests, with empirical findings on coverage, correctness and the need for human review. A five-layer architecture comprising code and history data, risk modelling, test generation, execution and triage, and feedback layers is proposed, along with research gaps, objectives, a comparison of techniques and challenges such as the oracle problem, flaky tests, dataset quality and trust. The review concludes that AI makes testing faster and better targeted but does not remove the need for human judgement.

DOI: https://doi.org/10.5281/zenodo.24217564

Artificial Intelligence In Disaster Prediction: A Review Of Flood, Earthquake And Wildfire Forecasting With Machine Learning And Early Warning Dashboards

Authors: Vishwkarma Mohit Santapal, Prof. S. S. Bhor, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Natural disasters such as floods, earthquakes, cyclones, landslides and forest fires cause severe loss of life, property and environment. Conventional disaster prediction depends on historical records, weather bulletins and expert analysis, and delayed or incomplete information makes response difficult. This paper reviews artificial intelligence and machine learning for disaster prediction and early warning. It examines flood prediction with machine learning and deep learning on rainfall, river level and catchment data, including rainfall–runoff modelling with recurrent networks and global flood forecasting in ungauged basins; precipitation nowcasting with convolutional recurrent and generative models on radar data; earthquake detection and phase picking with deep networks and the scientific limits of earthquake early warning; wildfire risk prediction; and satellite and aerial damage assessment with social-media analysis for post-disaster situational awareness. It discusses data sources, from meteorological services and river gauges to satellite imagery and IoT sensors, and the design of risk dashboards and alerting that follow people-centred early warning principles and India's disaster management framework. A five-layer architecture comprising data acquisition, integration, risk modelling, warning and dissemination, and response support is proposed, with research gaps, objectives, a comparison of hazards and challenges such as data scarcity, rare events, false alarms and last-mile communication. The review stresses that AI predictions supplement, never replace, official warnings.

DOI: https://doi.org/10.5281/zenodo.24217677

Artificial Intelligence In Vehicle Automation: A Review Of Advanced Driver Assistance Systems, Perception, Sensor Fusion And Safety

Authors: Wajage Aryan Sujit, Prof. S. S. Bhor, Prof. M. C. Kolse, Prof. S. K. Said

Abstract: Advanced Driver Assistance Systems (ADAS) use cameras, radar, LiDAR, ultrasonic sensors and artificial intelligence to monitor a vehicle's surroundings and help drivers avoid crashes. Features such as automatic emergency braking, lane departure warning and lane keeping, adaptive cruise control, blind-spot detection, parking assistance and collision avoidance form the building blocks of vehicle automation, which ranges from basic assistance to highly automated driving. This paper reviews the role of AI in ADAS and vehicle automation. It explains the levels of driving automation, the sensor suite and its trade-offs, and the AI techniques used for lane detection, object detection on camera images and LiDAR point clouds, bird's-eye-view perception and multi sensor fusion. Evidence on the safety benefits of ADAS features is discussed, together with the difficulty of demonstrating the reliability of highly automated systems. Safety and security engineering, including functional safety, safety of the intended functionality, cybersecurity standards and adversarial attacks on perception, are examined. A layered ADAS architecture covering sensing, perception, fusion and prediction, planning and decision, and control and human–machine interface is proposed, with research gaps, objectives and a comparison of sensors. The review concludes that AI-driven ADAS can greatly reduce road accidents, provided that cost, adverse weather, processing demands and cyber-physical security are addressed.

DOI: https://doi.org/10.5281/zenodo.24217776