AI-Powered Local Agri-Market Intelligence and Farmer-Buyer Connect System

ABSTRACT Agriculture is a vital part of the Indian economy because it provides jobs for more than half of the country's workers and contributing significantly to food security. Despite advances in agricultural production, Indian farmers — particularly those at the small and medium scale — continue to face severe challenges in post-harvest market access, price discovery, and buyer connectivity. The absence of real-time, localized, and actionable market intelligence forces farmers to depend on intermediaries, resulting in below-market realizations and reduced profitability. This thesis presents the design, architecture, and implementation of an AIPowered Local Agri-Market Intelligence and Farmer-Buyer Connect System — a multi-agent, bilingual, explainable artificial intelligence platform that enables farmers to access real-time agricultural market intelligence through natural language queries in Hindi and English. The system addresses the critical gap between available government mandi data and the farmer's ability to use this information to make smart choices. The proposed system integrates a ReAct (Reasoning and Acting) AI Agent powered by the Gemini 2.5 Flash Large Language Model with a Multi-Agent Architecture comprising ten specialized agents: Intent Detection, Location Resolution, Mandi Intelligence, Tavily Internet Search, Reasoning, Price Prediction, Fact-Check, Answer Generation, and Sell Decision agents. The system employs RetrievalAugmented Generation (RAG) using a FAISS vector database with sentencetransformer embeddings to retrieve relevant historical price context, augmenting the LLM's reasoning with verified factual data. Live mandi price data is fetched from the Government of India's Agmarknet API and eNAM platform, which serve as the trusted source of truth. The Tavily Search tool enables the agent to retrieve current internet-based market trends, agricultural news, and demand signals. A dedicated Fact-Check Agent cross-validates AI-generated insights against government data and search results, assigning confidence scores to prevent hallucination. A Prediction Agent employs moving-average trend analysis to forecast price trends and generate sell-or-wait recommendations with quantified confidence. The system's bilingual Answer Generation Agent produces farmer-friendly explanations in both Hindi and English, ensuring accessibility across linguistic demographics. The frontend, built with React.js and Tailwind CSS, features a voiceenabled AI chatbot interface, animated mandi dashboards, trend charts, and an Explainable AI reasoning panel. The backend system is organized and managed using Node.js with Express.js and MongoDB, while the AI microservice layer is implemented using Python and FastAPI. Experimental results show that the system correctly understands what the user wants and finds the right local market prices, performs multi-market comparative reasoning, and quickly creates helpful replies in bilingual responses based on the user's specific situation. The architecture is modular, scalable, and production-ready, suitable for deployment on cloud infrastructure via Vercel, Render, and MongoDB Atlas. This research contributes to the fields of Agentic AI, Agricultural Informatics, and Explainable AI by demonstrating that a multi-agent ReAct framework grounded in government-verified data can bridge the digital divide in agricultural market access and empower rural farmers with intelligent, trustworthy, and accessible decision support. Keywords: Agentic AI, ReAct Agent, Multi-Agent System, RAG, FAISS, Mandi Intelligence, Agricultural Market, Gemini LLM, NLP, Explainable AI, Bilingual AI, Farmer-Buyer Connect. Supervised by: Dr. Rajiv Misra, IIT Patna & Abhinandan Kumar, NDRI Karnal