Gram Saathi - Overview
Gram Saathi - Farmer Profile
Gram Saathi - System Health
Gram Saathi - Analytics
Gram Saathi - Call History

Gram Saathi

Inspiration

India has millions of farmers who have no smartphone or internet access. They are completely excluded from the AI revolution transforming agriculture. Every existing agri-tech solution requires an app or data connectivity, leaving the most vulnerable farmers behind. We asked ourselves: what if a farmer could access live market prices, weather forecasts, and government scheme information just by giving a missed call from a ₹500 feature phone? That question became Gram Saathi, a voice-first AI assistant that meets farmers where they already are: on a basic phone call, in their own language, at zero cost.

What it does

Gram Saathi is a voice-only AI agricultural assistant that works through a missed call trigger. A farmer dials a number —> the call auto-rejects (costing them nothing) —> Gram Saathi calls them back with a full AI-powered voice conversation in their regional language. It provides:

Live mandi prices from 1,266+ agricultural markets across India via official government data
5-day hyperlocal weather forecasts with severe weather alerts (heavy rain, heatwave, frost)
Government scheme eligibility matching against 45+ central and state schemes
Crop-specific advisory — season-aware, region-specific guidance on sowing, irrigation, pest management, and fertilizer
Progressive profiling — builds the farmer's profile naturally through conversation (no forms, no apps)
Analytics dashboard for monitoring calls, transcripts, user profiles, and usage patterns in real-time

All of this works across 11 Indian languages: Hindi, Tamil, Telugu, Kannada, Marathi, Bengali, Gujarati, Punjabi, Malayalam, Odia, and English.

How we built it

The architecture is a real-time voice pipeline stitching together speech, AI reasoning, and telephony:

Farmer speaks → LiveKit (VAD) → Sarvam STT → AWS Bedrock LLM (with Nova2) → Sarvam TTS → LiveKit → Farmer hears

Backend: Python 3.12 + FastAPI (async), PostgreSQL + Redis, running on a single AWS EC2 instance in Mumbai (ap-south-1).

Voice pipeline: LiveKit Agents SDK handles real-time WebRTC media. We built a custom Bedrock LLM plugin to support tool use with Nova 2 via the non-streaming Converse API. Sarvam AI provides native Indian-language STT (saaras:v3) and TTS (bulbul:v3) at 8kHz PCM for telephony quality.

Telephony: Twilio provides the phone number and Elastic SIP Trunk. On a missed call, a webhook triggers an outbound callback through LiveKit's SIP integration, the farmer never pays a paisa.

Data sources: data.gov.in for live mandi commodity prices (30-min cache), Open-Meteo for hyperlocal weather (2-hr cache), a curated schemes.json for government scheme matching, and structured crop advisory data with season detection.

Frontend: Next.js 16 + Tailwind CSS + shadcn/ui dashboard with TanStack React Query, showing live call monitoring, call history with transcripts, user profiles, and analytics charts.

Deployment: Docker Compose orchestrating 7 services (PostgreSQL, Redis, LiveKit server, SIP, backend, agent worker, frontend) on a single t3.medium instance.

Challenges we ran into

Message ordering rules — Bedrock enforces strict message alternation (user/assistant) and requires specific block ordering. Debugging silent failures from malformed message histories was painful.
11-language number formatting — Digits and currency symbols can't be spoken by TTS. We built language-specific number-to-word conversion so "₹2,450" becomes "do hazaar chaar sau pachaas rupaye" in Hindi or "இரண்டாயிரத்து நானூற்று ஐம்பது ரூபாய்" in Tamil.
Marker stripping — The LLM emits <thinking> blocks, <<<LANG:xx-YY>>> markers, and <<<END_CALL>>> tags that must never reach the TTS engine. Getting this filtering reliable across all edge cases took multiple iterations.
Mandi price API reliability — The data.gov.in API is inconsistent. We implemented a dual-endpoint strategy (primary date-specific + fallback rolling 3-day window) with 30-minute caching to ensure farmers always get data.
Feature phone audio quality — 8kHz narrowband audio over Indian telecom networks means STT accuracy drops. Tuning VAD (voice activity detection) thresholds and prompt engineering for brevity was critical.

Accomplishments that we're proud of

Zero cost to the farmer — The entire missed call → callback model means a farmer with no balance, no internet, and no smartphone can access AI-powered agricultural intelligence for free.
11 languages from day one — Not as an afterthought, but as a core design principle. The system prompt uses feminine verb conjugations, culturally appropriate greetings, and region-specific agricultural knowledge in every supported language.
Sub-3-sentence responses — We engineered the LLM to give phone-friendly answers. No walls of text — just the information the farmer needs, spoken naturally.
Progressive profiling — Instead of a registration form (impossible on a feature phone), the system builds the farmer's profile organically through conversation. By the second call, it already knows their state, district, crops, and land size.
Production-ready on a single EC2 — The entire stack (7 Docker services, real-time voice pipeline, database, caching) runs on a single t3.medium instance (~$30/month), making it viable for NGOs and government deployment.
Full observability dashboard — Real-time call monitoring, live transcripts, language distribution analytics, and tool usage tracking — all in a polished Next.js UI.

What we learned

Voice-first design is fundamentally different from screen-first. You can't show a table of prices — you have to narrate the most relevant one. Every UX decision had to be rethought for audio.
Indian language NLP is still an emerging frontier. STT accuracy varies significantly across languages and dialects. Tamil and Odia were notably harder than Hindi.
Government APIs need defensive engineering. data.gov.in has inconsistent response formats, occasional downtime, and no SLA. Building resilient fallbacks is not optional.
LLM tool use over voice requires tight latency budgets. Every tool call adds a round-trip. We learned to make the LLM decisive — call the right tool on the first try, don't ask clarifying questions when you have enough context.
Farmers don't behave like chatbot users. They speak in long, context-rich sentences mixing multiple requests. The system prompt had to be engineered to handle "what's the price of wheat and will it rain tomorrow and am I eligible for PM-KISAN" as a single turn.

What's next for Gram Saathi

WhatsApp integration — Extend the same voice AI to WhatsApp voice notes for farmers who have basic smartphones but prefer voice over text.
Proactive outbound alerts — Call farmers before a severe weather event, when mandi prices spike, or when a government scheme deadline approaches — without them having to call in.
Regional dialect fine-tuning — Partner with Sarvam AI to improve STT accuracy for underserved dialects (Bhojpuri, Chhattisgarhi, Rajasthani).
Farmer-to-farmer knowledge network — Let farmers record and share crop tips in their own voice, building a community knowledge base.
Integration with FPOs and cooperatives — Enable Farmer Producer Organizations to broadcast price negotiation updates and collective selling opportunities.
Offline-capable edge deployment — Explore running lightweight models on local hardware at village telecom towers to reduce latency and cloud costs.
Government and NGO partnerships — Pilot with state agriculture departments and organizations like Digital Green to reach millions of farmers across India.