A real-time multimodal, multi-agent AI advisory system for plant disease management (2025–2026).
What I did
Combined CNN-based image classification, a multi-agent architecture, domain fine-tuned LLM-driven advisory generation and real-time weather data to give crop-specific recommendations.
Built specialised Image, Context and Safety agents using modular Retrieval-Augmented Generation (RAG), web scraping and ReAct-based reasoning to produce agronomically accurate, safe and regulation-compliant outputs, validated against authoritative agricultural and health sources.
Used TensorFlow and PyTorch CNNs (MobileNet architecture) for disease detection; PEFT/LoRA fine-tuning of LLMs; Direct Preference Optimization (DPO) with Unsloth to align outputs with real farmer needs; Open-Meteo API for weather; React frontend; Flask backend.
Evaluated with domain experts, achieving more than 85% on every evaluation metric used.
Published at IEEE ICARC 2026 and cited in MDPI Technologies.