AI-powered voice & SMS agricultural intelligence for India's small and marginal farmers in their own language, on the phone they already own.
Built for Build with AI: Code for Communities a Google Cloud hackathon where every problem statement was written by an MP's office, and the best solutions get piloted in a real constituency.
86.1% of India's ~146 million landholdings are small and marginal (Agriculture Census, 2015–16). Most of these farmers choose what to plant based on habit and hearsay — not on what the soil, the groundwater, or the sky are actually saying. When the guess is wrong, the loss is total: no insurance cushion, no smartphone to check an app, and one extension officer covering dozens of villages.
AgriVaani doesn't ask farmers to change how they live. It meets them on a phone call.
AgriVaani is three connected AI systems, reachable entirely by voice IVR and SMS — no app, no smartphone, no literacy requirement:
| Pillar | What it does | |
|---|---|---|
| 🛰️ | Smart Crop Recommendation | Fuses satellite NDVI/NDMI, soil pH, groundwater depth, and rainfall history into a ranked, plain-language crop recommendation per plot, per season. |
| ⚡ | Real-Time Advisory & Dry-Spell Alerts | Fuses weather forecasts with real field sensor data (Arduino Mega 2560 + GSM + pH sensor) to push proactive irrigation/fertigation alerts by voice and SMS — before the damage happens, not after. Can optionally auto-actuate an irrigation pump via a 4-channel relay, gated by hard safety limits. |
| 🔬 | Crop Health Logging | A farmer sends a photo or voice note of a sick plant. Gemini's multimodal model gives an instant diagnosis and confidence score. Low-confidence or high-severity cases are automatically escalated to a human expert at the nearest Rythu Seva Kendra. |
FARMER TOUCHPOINTS
📞 Toll-free voice call (IVR) 💬 SMS (inbound/outbound)
│ │
┌─────────▼──────────┐ ┌──────────▼─────────┐
│ Telephony Gateway │◄───────┤ SMS Gateway │
│ (Twilio / Exotel) │ │ (Twilio / Exotel) │
└─────────┬──────────┘ └──────────┬─────────┘
└───────────────┬──────────────┘
▼
┌───────────────────────────┐
│ Orchestration API │ Node.js / TypeScript
│ session state · routing │
│ IVR state machine │
└──────────────┬────────────┘
┌──────────────┬───────────┼───────────────┬──────────────┐
▼ ▼ ▼ ▼ ▼
┌──────────────┐ ┌───────────┐ ┌───────────┐ ┌────────────┐ ┌─────────────┐
│ Recommend │ │ Advisory │ │ Crop │ │ Language │ │ RSK Officer │
│ Service │ │ Alert │ │ Health │ │ Provider │ │ Dashboard │
│ FastAPI + │ │ Engine │ │ Diagnosis │ │ Bhashini / │ │ Next.js 14 │
│ XGBoost │ │ Node cron │ │ Gemini │ │ Google STT │ │ │
│ │ │ + rules │ │ Vision │ │ /TTS │ │ │
└──────┬───────┘ └─────┬─────┘ └─────┬─────┘ └────────────┘ └──────┬──────┘
│ │ │ │
▼ ▼ ▼ │
┌──────────────────────────────────────────────────┐ │
│ Data Layer │ │
│ Earth Engine · Soil Health Card · CGWB │ │
│ IMD / weather · Field hardware (GSM/GPRS + SMS) │ │
└────────────────────────┬─────────────────────────┘ │
▼ │
┌────────────────────────────┐ │
│ PostgreSQL · Firestore · │◄───────────────────────┘
│ BigQuery │
└────────────────────────────┘
🌱 Field 📡 GSM / GPRS
┌─────────────────────┐ ┌───────────────────────┐
│ Arduino Mega 2560 │ ── pH readings ──► │ Orchestration API │
│ + SIM800L GSM │ │ /v1/sensors/:id/... │
│ + pH sensor │ ◄── commands ──── │ │
│ + 4-ch relay │ └───────────────────────┘
│ ├─ irrigation pump │
│ ├─ fertigation valve A ▼
│ ├─ fertigation valve B Alert Engine (cron)
│ └─ spare fuses forecast + soil signal
│ safety: 30-min max runtime, → voice + SMS alert
│ local manual override always wins → optional auto-actuation
└────────────────────┘ (opt-in, critical-only)
| AI / ML | Vertex AI · Gemini 2.x (multimodal) · XGBoost recommender + rules-based fallback |
| Geospatial | Google Earth Engine (Sentinel-2 NDVI/NDMI) · ISRO Bhuvan · CGWB / India-WRIS groundwater |
| Weather | IMD API · NASA POWER / OpenWeatherMap fallback |
| Voice & SMS | Twilio Programmable Voice + SMS (Exotel-ready via provider interface) |
| Language | Bhashini (primary) · Google Cloud Speech-to-Text & Text-to-Speech (fallback, live) |
| Backend | Node.js + TypeScript (orchestration) · Python + FastAPI (recommendation engine) |
| Database | PostgreSQL (Cloud SQL / Neon / Supabase) · Firestore · BigQuery |
| Dashboard | Next.js 14 + TypeScript + Tailwind CSS |
| Field Hardware | Arduino Mega 2560 · SIM800L GSM module · analog pH sensor · 4-channel relay module |
agrivaani/
├── apps/
│ ├── orchestration-api/ # Node.js/TypeScript — telephony, sessions, IVR, alerts
│ ├── recommendation-service/ # Python FastAPI — crop scoring + Gemini rationale
│ └── officer-dashboard/ # Next.js 14 — RSK escalation queue
├── firmware/
│ └── sensor_node/sensor_node.ino # Arduino Mega 2560 + GSM + pH + relay
├── scripts/
│ ├── seed_demo_data.ts
│ ├── telemetry_replay.py
│ └── fake_call.ts
├── db/
│ └── migrations/001_initial_schema.sql
├── docs/
│ ├── DEMO_SCRIPT.md
│ └── KNOWN_GAPS.md
└── .env.example
This project runs natively on your machine. No virtualization needed.
Create a free instance at neon.tech or supabase.com, then:
psql <your_connection_string> -f db/migrations/001_initial_schema.sqlcp .env.example .env
# fill in DATABASE_URL at minimum — everything else falls back to
# deterministic mocks if a key is missing (see KNOWN_GAPS.md)cd scripts && ts-node seed_demo_data.ts# Terminal 1 — orchestration API
cd apps/orchestration-api && npm install && npm run dev
# Terminal 2 — recommendation engine
cd apps/recommendation-service && python -m venv venv
source venv/bin/activate # Windows: venv\Scripts\activate
pip install -r requirements.txt
uvicorn main:app --reload --port 8000
# Terminal 3 — officer dashboard
cd apps/officer-dashboard && npm install && npm run devngrok http 3000
# point your Twilio number's webhook at <ngrok-url>/v1/voice/inboundFull walkthrough: docs/DEMO_SCRIPT.md
AgriVaani is built so that every mocked component sits behind the same interface a real one would use — swap a credential, not the architecture. Full detail in docs/KNOWN_GAPS.md.
| Component | Status |
|---|---|
| Crop scoring (rules-based) | ✅ Live, unit-tested |
| Gemini rationale generation | ✅ Live (falls back to static strings without a key) |
| Google Cloud Speech-to-Text / Text-to-Speech | ✅ Live |
| Arduino Mega 2560 firmware (pH + relay + safety cutoffs) | ✅ Written, hardware-ready |
| Alert engine (deterministic dry-spell + pH thresholds) | ✅ Live |
| Gemini Vision crop diagnosis | ✅ Live (falls back to a mock diagnosis without a key) |
| RSK officer dashboard | ✅ Live |
| Twilio voice/SMS | ✅ Live with a trial account + ngrok |
| Earth Engine satellite sync | 🟡 Mocked (realistic randomized NDVI/NDMI) |
| XGBoost trained model | 🟡 Rules-based fallback in its place |
| Bhashini Indic language APIs | 🟡 Mocked (enterprise account required) |
| Officer authentication | 🟡 Stubbed (pilot-blocker, not a hackathon-blocker) |
- Auto-actuation is opt-in per plot, critical-alerts-only — advisory (voice/SMS) is always the default response. The system never silently starts controlling a farmer's irrigation pump.
- Hardware-level guardrails: a 30-minute max relay runtime regardless of the source of the command, and a local manual override switch that always wins over a remote signal.
- Alert fatigue control: no duplicate alert for the same unresolved condition within 24 hours.
- Explicit escalation rule, not implicit LLM judgment:
needs_expert = confidence < 0.6 OR severity == "high"lives in code, not inside a prompt.
- Crop recommendation engine with rules-based fallback
- Real-time alert engine with hardware sensor fusion
- Crop health diagnosis with human escalation
- Voice/SMS orchestration with live Twilio + Google TTS
- Real Earth Engine satellite integration
- Trained XGBoost model on labeled yield data
- Bhashini production integration for full Indic language coverage
- Pilot deployment with one Rythu Seva Kendra cluster
- District-scale rollout in partnership with the state agriculture department
SSY – Solve. Scale. Yield. — Indore
Built for Build with AI: Code for Communities (Google Cloud Hackathon)
Krishi Buddhimatta, Har Kisan Tak Agricultural intelligence, to every farmer.