International Journal of Science, Engineering and TechnologyAuthors: Disha Mudgal, Amey Bhogle, Dr. Jasbir Kaur, Ifrah Kampoo
Abstract: Smallholder farmers often make crop decisions with incomplete and delayed information about soil, season, and weather. AgriCare addresses this gap through a Flutter-based mobile application that combines cloud reasoning with on-device machine learning. The proposed system first attempts to generate structured crop recommendations using a Groq-hosted Llama 3.3 70B model, enriched with live weather and reverse-geocoded location context. If the cloud path fails because of poor connectivity, an API error, or a timeout, the application automatically falls back to a TensorFlow Lite classifier stored on the device. This hybrid design keeps the application useful in rural conditions where mobile data may be inconsistent, while still offering richer explanations whenever cloud inference is available. Prototype evaluation shows that the local TFLite path returns predictions in under 100 ms, while the cloud path produces more detailed ranked crop cards within a few seconds on stable connectivity. The paper presents the architecture, data flow, mathematical preprocessing, evaluation results, limitations, and future improvements of AgriCare as a practical agricultural advisory system.
DOI: http://doi.org/10.5281/zenodo.20826060
