KisanMind

A smallholder farmer deciding what to plant needs soil, weather, mandi prices and scheme eligibility at once, and that information lives in four places and two languages. HackWarts gave us 24 hours. I took the AI service; Ved, Nimit and Parth took the Next.js app and the Node backend.

The crop node is deliberately not an LLM call. It loads a catalogue of twenty-plus crops, eliminates by rules (soil, season, water, budget) and records why each crop was rejected, scores the survivors, and only then asks Gemini to re-rank the top three with reasoning. That makes the recommendation explainable and testable: 152 test functions cover the agents, the graph, the weather service, the chat router and the voice path, and they were written during the hackathon, not after.

Evaluation. The pipeline has tests but no measured quality: no agreement rate against an agronomy reference, no word-error rate on Hindi speech. Until a 30-profile evaluation set exists and its results are published here, this page does not carry a measured mark. That is the next piece of work on this project.