KrishiBondhu

Problem: Bangladeshi smallholder farmers who see disease or pests on a crop usually ask the local pesticide seller, who recommends whatever is in stock. The Department of Agricultural Extension (DAE) publishes good guidance and a register of approved pesticides for each crop. That information sits in PDFs and websites that few farmers read. Each local agriculture officer also serves thousands of farmers, so personal advice is hard to get. Solution: KrishiBondhu is a Bangla-first AI crop-health advisor. A farmer sends a photo, a voice message, or text, and the agent: - Diagnoses from the photo. Meta Muse identifies the problem, and SAM 3.1 marks the affected area and measures how much of the leaf is diseased. - Answers from official sources. Advice comes from DAE's Krishoker Janala (998 problems across 122 crops), and every pesticide is checked against the DAE register for that crop. Unregistered ones are flagged and never recommended. - Builds a day-by-day treatment plan. It checks the weather forecast and asks for follow-up photos. A measured worsening escalates to a booking with the district agriculture officer. - Understands Bangla speech. It uses Meta ASR for voice input and remembers each farmer's land, crops, and history. - Keeps farmer data private. Postgres row-level security isolates each farmer's data. - I plan to bring it to WhatsApp and a phone call version so it reaches even more people. Tech: Meta Muse, SAM 3.1 and Meta ASR; a LangGraph agent; a Next.js 16 front end with OpenUI for interactive cards and timelines; Postgres; LangSmith tracing. Run the app: https://krishibondhu-ai-meta-hackathon.vercel.app/chat/ (Use any Bangladeshi number to enter; +8801700000000 would do.) Impact: It turns official agricultural knowledge into trustworthy, personalised, follow-up advice in the farmer's own language. This should reduce misuse of pesticides and crop loss.