🌾 Kisan Mitra — AI Farming Assistant

Bridging the information gap between India's farmers and modern AI — in their own language.

💡 Inspiration

Agriculture is the backbone of India, yet millions of farmers struggle to access timely, accurate, and localized advice. Between pest attacks, unpredictable weather, and complex government schemes, the learning curve is steep.

Kisan Mitra was born from a simple belief: every farmer deserves expert knowledge in their native tongue — no middlemen, no barriers. We used modern AI to make agricultural guidance and government subsidies accessible through a simple conversational interface.

🤖 What It Does

Kisan Mitra is a multilingual, Agentic AI farming assistant. Ask it anything — crop management, pest control, irrigation, or government subsidies — and it responds in the language you asked in.

Core capabilities:

🌱 Crop & Soil Advice — Powered by a local RAG knowledge base with tailored, actionable guidance
📜 Govt Scheme Finder — Real-time web search for the latest subsidies, crop insurance & policies
🗣️ True Multilingualism — Hindi, English, Marathi, Bengali, and more Indian languages
🔁 Agentic Reasoning — An intelligent loop that decides when to search locally vs. the web

🏗️ How We Built It

Layer	Technology
Data Ingestion & Storage	Apache Spark (PySpark) + Databricks + Delta Lake
Embeddings & Vector Store	HuggingFace `all-MiniLM-L6-v2` + FAISS / ChromaDB
Agentic Framework	LangGraph (ReAct pattern)
LLM Engine	DeepSeek-R1 via HuggingFace Endpoints
Deployment	Databricks App (scalable REST endpoint)

The agent intelligently routes every query to either the local RAG tool (search_rag) for crop knowledge, or the web search tool (search_schemes) for live government data — then synthesizes a clear, numbered response.

🚧 Challenges We Ran Into

Language Consistency — Keeping the LLM from slipping into English mid-response required strict system prompting and output parsing.

Routing Accuracy — Getting the ReAct agent to reliably choose RAG vs. web search took extensive prompt engineering and loop-limit tuning.

Cloud Infrastructure — Persisting FAISS/ChromaDB vector stores on Databricks' distributed file system needed careful DBFS path management.

🏆 Accomplishments We're Proud Of

✅ Fully deployed a scalable LangGraph agentic framework within the Databricks ecosystem
✅ Achieved accurate multilingual intent recognition with lightweight open-source models
✅ Dynamically bridges a static knowledge base + real-time internet for farming queries

📚 What We Learned

LangGraph's ReAct pattern is powerful but demands careful orchestration vs. simple chains
Delta Lake + Spark integrates seamlessly with vector stores as a unified AI data backend
Prompting open-source models for multilingual zero-shot reasoning has deep nuance

🚀 What's Next for Kisan Mitra

🎙️ Voice integration — Speech-to-Text & Text-to-Speech so farmers never need to type
💬 WhatsApp bot — Frictionless access for rural users on familiar infrastructure
🌦️ Localized weather APIs + real-time crop market prices added to the Delta Lake dataset