AgriSakha

Knowledge. Credit. Markets. In Your Voice.

Comment

AgriSakha 🌾

Voice-First AI for India's Smallholder Farmers

💡 Inspiration

India's 120 million smallholder farmers are the backbone of the nation's food security — yet they remain the most underserved by technology. They face a brutal trifecta of problems every single day:

  • No credit access — Without formal credit histories, banks reject 70%+ of agricultural loan applications, forcing farmers into debt traps with moneylenders charging 40–60% annual interest.
  • No agronomic guidance — A single agricultural extension worker serves tens of thousands of farmers. Timely, personalized advice on pests, fertilizers, or market prices is virtually out of reach.
  • No digital literacy — Most farmers still rely on basic feature phones. Existing AgriTech apps exclude the very people who need them most.

We were inspired by a simple but powerful question:

What if a farmer could just speak in Hindi and get the same quality of advice available to a city-based agribusiness?

AgriSakha was born from that belief — that the only barrier between a farmer and life-changing information should be nothing more than their own voice.

🤖 What it does

AgriSakha (meaning "Farmer's Friend" in Hindi) is a voice-first AI platform designed to run on even the most basic feature phones via USSD — no smartphone, no internet browser, no literacy required.

🗣️ Voice AI — Speak, and AgriSakha Listens

Farmers speak naturally in Hindi. AgriSakha uses offline Hindi speech recognition (Vosk ASR) to transcribe the query and responds instantly in spoken Hindi via Coqui TTS:

Farmer Query (Hindi) AgriSakha Response
"मेरी फसल में कीड़े लग गए हैं" नीम तेल 5 मिली/लीटर — 3 दिन अंतराल पर छिड़काव करें
"गेहूं के लिए कितना यूरिया डालें?" 50 किलो यूरिया प्रति एकड़
"मेरा क्रेडिट स्कोर क्या है?" क्रेडिट स्कोर: 720 — ₹1,00,000 तक के कर्ज के पात्र हैं
"आज टमाटर की कीमत?" ₹40/किलो — मंडी भोपाल

💳 Satellite-Powered Credit Scoring

AgriSakha generates a credit score without a bank account or CIBIL history — using satellite imagery instead. By integrating with ISRO's Bhuvan API, the engine evaluates:

  • Crop Health (60%) — NDVI index derived from satellite data
  • Soil Moisture (20%) — Remote sensing hydration levels
  • Acreage (20%) — Verified land area from geospatial records

Every credit score event is cryptographically stored on a custom SHA-256 blockchain, creating a tamper-proof, auditable credit trail that financial institutions can trust.

🌱 AI Soil Analysis

Farmers input soil sensor readings (N, P, K, pH, Organic Carbon) and receive an instant Hindi-language fertilizer report — including exact Urea dosage in kg/acre and nutrient status labels:

मिट्टी विश्लेषण:
नाइट्रोजन: 35 ppm (सामान्य)
फॉस्फोरस: 20 ppm
सुझाव: 15.0 kg यूरिया/एकड़

📱 USSD Menu (Feature Phone Compatible)

No smartphone needed. Farmers simply dial *123#:

AgriSakha मेनू:
1. कर्ज (Loan Eligibility)
2. कीट (Pest Control)
3. मिट्टी (Soil Test)

🛠️ How we built it

We built AgriSakha as a modular Python system with three independent, composable engines following an edge-first philosophy — models run locally, so the system functions even with zero connectivity.

Component Technology Why We Chose It
Hindi ASR Vosk vosk-model-small-hi-0.22 Runs fully offline — critical for low-connectivity rural areas
Hindi TTS Coqui TTS tts_models/hi/indic-tts Natural Indic prosody, no cloud dependency
Soil Analysis NumPy weighted scoring model Lightweight, fast inference at the edge
Credit Engine Custom satellite-weight algorithm Democratizes credit without traditional financial data
Blockchain Custom SHA-256 chain (Python) Immutable, auditable record of every credit event
Web Demo UI Streamlit Rapid prototyping and demonstration of the soil pipeline
Data NFHS-5 (data.gov.in), Rajya Sabha AU records Grounded in real agricultural and demographic context

Voice Processing Pipeline

Farmer speaks → Vosk ASR transcribes → Intent matched
     → Hindi response generated → Coqui TTS speaks back

Credit Scoring Formula

score = (crop_health × 0.6) + (soil_moisture × 0.2) + (acreage × 0.2)
# Each score event → immutably stored as a SHA-256 blockchain block

🧗 Challenges we ran into

  1. Hindi ASR quality in noisy environments — Background noise from fields, wind, and livestock severely degrades speech recognition. We tuned Vosk model parameters and explored noise-reduction pre-processing with librosa and pydub.

  2. Running TTS on constrained hardware — The Coqui Indic TTS model is resource-intensive. Getting it to run reliably on lower-end hardware at acceptable latency required exploring model quantization and audio format trade-offs.

  3. Satellite API availability — ISRO's Bhuvan API is not publicly available for real-time programmatic access. We built a realistic mock mirroring the expected data schema (NDVI, soil moisture, acreage), requiring deep research into how satellite-based agricultural indices are computed.

  4. Designing for zero digital literacy — Every UI decision had to assume no familiarity with apps, menus, or text. The USSD *123# flow had to deliver maximum information in minimum steps, with every response tested for clarity when read aloud at natural Hindi speaking speed.

  5. Blockchain speed vs. integrity — Each credit event requires a hash computed over the full prior chain. We carefully designed the block structure so validation stays fast as the chain grows, without sacrificing tamper-proof guarantees.

🏆 Accomplishments that we're proud of

  • ✅ Built a fully offline, Hindi-first voice AI — a rare feat in Indian AgriTech, where most tools require internet connectivity and English literacy.
  • ✅ Created a credit score that works without a bank account — using satellite imagery as the primary signal, potentially unlocking formal credit for millions of previously "unscoreable" farmers.
  • ✅ Implemented a working blockchain for agricultural credit records — demonstrating how immutable, decentralized record-keeping builds institutional trust in farmer data.
  • ✅ Grounded in real data — NFHS-5 demographic data and Rajya Sabha agricultural policy records informed our design, ensuring AgriSakha addresses real, documented needs.
  • ✅ End-to-end pipeline in one hackathon — Voice I/O, soil analysis, credit scoring, blockchain storage, and a web demo UI — all integrated and test-covered.

📚 What we learned

  • Language is the biggest barrier in rural tech adoption — Not the smartphone gap, not the internet gap. If an interface doesn't speak a farmer's language, it will not be used. Building Hindi-first forced us to rethink every design assumption.
  • Offline-first is non-negotiable for Bharat — Cloud-dependent architectures fail at the last mile. Vosk's lightweight offline ASR was a revelation — genuinely high quality at a fraction of cloud compute cost.
  • Satellite data is the unbanked farmer's credit history — A farmer who has never touched a bank still has years of verifiable crop health data visible from space. This reframe fundamentally changes how we think about financial inclusion.
  • Blockchain solves a real trust problem here — In contexts where local record-keeping is unreliable or prone to manipulation, a SHA-256 immutable chain provides a foundation both farmers and lenders can trust.
  • Simplicity is a technical achievement — Building a system sophisticated enough to help farmers while simple enough for a first-time USSD user is harder than building a feature-rich app. Constraints drive better design.

🚀 What's next for AgriSakha

Immediate Roadmap (Next 3 Months)

  • 🛰️ Integrate real ISRO Bhuvan API — Move from mocked satellite data to live NDVI and soil moisture feeds for genuine, real-time credit scores.
  • 🗣️ Expand to 10+ Indian languages — Add Marathi, Tamil, Telugu, Bengali, and Kannada using AI4Bharat's IndicTTS and IndicASR models.
  • 📡 IoT soil sensor integration — Replace manual data entry with real-time readings from low-cost NPK sensors over LoRa networks.

Medium-term Vision (6–12 Months)

  • 🏦 Bank and MFI partnerships — Present blockchain-anchored credit scores to rural banks (RRBs) and microfinance institutions as verifiable alternative credit signals.
  • 📊 Live mandi price integration — Real-time commodity prices from the AGMARKNET API, delivered by voice, so farmers sell at the right time and place.
  • 🐛 Pest outbreak early warning — Aggregate query patterns across farmers to detect regional pest outbreaks before they spread, alerting extension workers proactively.

Long-term Goal

Give every Indian farmer access to the same quality of agricultural intelligence available to a large agribusiness — for free, in their own language, from any phone.

AgriSakha is not just an app. It is infrastructure for agricultural equity.

🛠️ Built With

Python · Vosk · Coqui TTS · NumPy · Streamlit · SHA-256 Blockchain · sounddevice · pydub · librosa · ISRO Bhuvan API · AI4Bharat IndicNLP

🙏 Acknowledgements

  • AI4Bharat — for Hindi NLP tools and IndicTTS support
  • Alphacephei (Vosk) — for the lightweight offline Hindi ASR model
  • ISRO Bhuvan — for satellite data API inspiration
  • The/Nudge Institute — for the Pragati AI for Impact Hackathon platform
  • data.gov.in — NFHS-5 Factsheets for agricultural demographic context
Share this project:

Updates