KrishakSarthi — AI-Powered Agriculture Assistant for Every Farmer
Tagline
From smartphone to WhatsApp — one AI companion covering crop planning, disease detection, and market access for India's 140 million farming households.

The Problem
Most Indian farmers don't have access to timely agronomic advice. Agricultural extension officers are stretched thin, internet literacy is low, and existing agri-apps require a smartphone and stable connectivity. Farmers lose crops to diseases they could have prevented and miss market opportunities because they can't find buyers.

What We Built
KrishakSarthi is a full-stack AI agriculture platform with three deeply integrated layers:

Mobile App (React Native + Expo)
A farmer-friendly Android/iOS app with large touch targets, high-contrast green UI, and full trilingual support (Hindi, Bengali, English).
AI Crop Disease Detection — photograph a diseased leaf; Gemini Vision identifies the pathogen, severity, and treatment within seconds
Crop Prediction Engine — combines live weather data with soil and location inputs to recommend the best crop for the season
Crop Planner + Calendar — step-by-step guided farming plan with a visual activity calendar
Document Analyzer — farmers photograph government forms, seed packets, or invoices and get plain-language summaries
AI Chatbot — RAG-powered conversational assistant backed by a curated agriculture knowledge base
Farmer-Buyer Marketplace — supply chain screen connecting farmers directly with verified buyers; live GPS location sharing for last-mile logistics
Daily Routine Tracker — task scheduler tailored to crop growth stages

RAG Knowledge Backend (Node.js + Vercel)
A Retrieval-Augmented Generation API that turns agriculture PDFs and documents into an intelligent Q&A engine.
PDF ingestion pipeline chunks and embeds agricultural documents into Pinecone vector store
Queries are semantically searched, top chunks retrieved, and answers synthesized by Gemini with source citations
Exposes a clean REST API (POST /api/chat) consumed by both the mobile app and WhatsApp service
Multi-language translation layer for Hindi ↔ English responses
3. WhatsApp Bot (Node.js + Meta Cloud API)
For farmers who only have a basic WhatsApp account.

Hooks into Meta's WhatsApp Cloud API via webhook
Farmers ask agriculture questions in Hindi or English and get back RAG-powered answers — no app download needed
Natural-language Crop Planning: farmer types "5 acre mein basmati dhan, Sodepur, 15 June se" — the bot extracts all fields, confirms, generates a full 90-day farming plan PDF, and delivers it back over WhatsApp
Voice call support for farmers who prefer speech

Challenges We Solved

Zero-connectivity farmers: WhatsApp channel needs only basic 2G; no app install required
Multilingual NLP extraction: Free-form Hindi messages parsed by Gemini to extract structured crop fields (crop name, area, location, date) reliably
PDF over WhatsApp: Generated PDFs uploaded to Meta Media API and delivered as WhatsApp documents, not just text
Encoding pipeline: Built a robust ingestion pipeline to chunk, embed, and upsert heterogeneous agriculture PDFs with metadata preserved for source citations
Impact Potential
Targets 140M+ smallholder farmers in India who rely on WhatsApp daily
Reduces crop loss from undetected disease with sub-10-second AI diagnosis
Bridges the farmer-buyer gap directly in the app, cutting out middlemen
Works in the language the farmer is most comfortable in

Challenges we ran into
RAG queries took 30+ seconds, exceeding Twilio's hard limit
Solution: Implemented async job queue with polling mechanism—respond immediately to Twilio, process in background, caller polls for result
Audio Processing Pipeline

WhatsApp Integration Challenges
Multi-Turn Conversation State

Keeping conversation context across multiple message exchanges for crop planning
Handling WhatsApp message retries (duplicate IDs)
Solution: Redis-based session storage with auto-expiration
Language Support (Hindi/English Mix)

WhatsApp media upload constraints (file size limits)
Proper multipart form-data formatting for PDF uploads
Solution: PDF generation with size optimization and chunked delivery
Rate Limiting & Quota Management

Handling API rate limits from multiple providers simultaneously
Managing quota exhaustion gracefully
Solution: Multiple API key rotation and fallback models