Farm Sutra

"AI-powered farming assistant for rural India — multilingual, 2G-ready, and built on Amazon Nova."

Comment

🌾 Farm Sutra — AI-Powered Farming Assistant for Rural India

💡 Inspiration

India has over 140 million farmers, yet most of them make critical decisions — when to sell, what disease is attacking their crop, whether to take a loan — based on word of mouth or outdated information.

The breaking point for us was a simple observation: a farmer in rural Haryana loses 30–40% of his crop yield every season not because he lacks hard work, but because he lacks access — access to timely market prices, access to expert crop diagnosis, access to fair credit.

Existing agri-apps are built for smartphone-savvy users in English. They ignore the reality that most Indian farmers:

  • Speak Hindi, Punjabi, Tamil, or Haryanvi — not English
  • Are on 2G networks with limited data
  • Have no formal credit history, making banks turn them away

We asked ourselves: What if AI could sit in every farmer's pocket — speaking his language, understanding his crop, helping him sell at the right price?

That question became Farm Sutra.

🛠️ How We Built It

Farm Sutra is a full-stack mobile + cloud application built on the AWS ecosystem, powered at its core by Amazon Nova Pro via Amazon Bedrock.

Architecture Overview

React Native (Expo) App
        │
        ├── AWS Amplify (API Gateway + Auth layer)
        │         │
        │         └── AWS Lambda (Node.js)
        │                   │
        │                   └── Amazon Bedrock
        │                           │
        │                           └── Amazon Nova Pro v1
        │
        ├── Open-Meteo API (Live GPS Weather)
        └── AsyncStorage (Offline chat persistence)

Feature Breakdown

🤖 Farm Sutra AI (ChatBot)

The AI chatbot uses Amazon Nova Pro via Bedrock for all responses. Every query is augmented with a language instruction so Nova replies strictly in the farmer's chosen language — Hindi, Punjabi, Tamil, Haryanvi, or English.

Crop Scanner — the most technically complex feature — sends both a base64-encoded image and a fully translated prompt to Nova's multimodal endpoint. The prompt itself is written in the farmer's language, structured to extract disease name, treatment plan, urgency, and confidence level.

$$\text{Crop Health Score} = f(\text{Image}{\text{base64}}, \text{Prompt}{\text{lang}}, \text{Nova Pro Multimodal})$$

🛒 Mandi (Market)

A real-time marketplace where farmers list their crops and buyers post demand. Built with AsyncStorage for persistence and includes live GPS-based weather context at the top of the screen.

🌦️ AgriWeather

Uses Open-Meteo's free GPS API to pull 7-day live forecasts. The AI layer converts raw WMO weather codes into actionable farming advice — "High humidity for 2 days — fungal infection risk for your wheat".

🏦 Credit Score

A proprietary farmer credit scoring system that rewards positive farming behaviors — scanning crops, checking weather, selling in the Mandi. This creates a verifiable activity log for farmers who have zero formal credit history.

$$\text{Credit Score} = \text{Base Score} + \sum_{i=1}^{n} \text{ActivityPoints}_i$$

Where each activity (crop scan, weather check, market listing) contributes weighted points to build a farmer's financial credibility over time.

🌍 Multilingual Engine

Every single string in the app — UI labels, AI prompts, error messages, weather alerts, pest warnings — is translated into 5 languages:

Language Region
English Default
Hindi (हिन्दी) North India
Punjabi (ਪੰਜਾਬੀ) Punjab
Haryanvi (हरियाणवी) Haryana
Tamil (தமிழ்) South India

📚 What We Learned

  • Prompt engineering is everything for multilingual AI. Simply asking Nova to "reply in Hindi" wasn't enough — we had to write the entire system prompt in the target language to get culturally accurate, dialect-appropriate responses. The Haryanvi dialect prompt took multiple iterations.

  • Multimodal inputs require careful image compression. Sending full-resolution images to Bedrock on a 2G network was impractical. We settled on quality: 0.5 in Expo ImagePicker — enough for accurate disease diagnosis while staying 2G-friendly.

  • Offline UX is a feature, not an afterthought. For rural farmers, connectivity is unreliable. Building graceful offline handling — persistent chat history, informative banners, queue messaging — required as much thought as the AI integration itself.

  • Credit scoring for the unbanked is a design challenge. Traditional credit models require financial history. We learned to think differently — behavior as credit — rewarding consistent, verifiable farming activities instead.

🚧 Challenges We Faced

1. Haryanvi Dialect

Haryanvi is not a standard language with NLP support. Getting Amazon Nova to respond in authentic Haryanvi farmer dialect — with words like "सै", "कोन्या", "आपणा" — required hand-crafting every prompt and testing with native speakers. No existing dataset or library helped here.

2. 2G Network Optimization

Most agri-apps assume 4G. We had to rethink every API call:

  • Image compression before Bedrock upload
  • AsyncStorage chat persistence to avoid re-fetching
  • Open-Meteo chosen specifically for its lightweight JSON response

3. Multimodal Prompt Translation

Writing the crop scan diagnostic prompt in 5 languages — with the exact same structured output format (emojis as anchors: 🌿, 🔍, 🦠, 💊, ⏰) — so that parsing remains consistent regardless of language was a significant challenge.

4. Credit Score Without Financial Data

Building a meaningful credit scoring system for farmers who have zero bank history meant inventing new signals. We used in-app behavioral data as a proxy for creditworthiness — a concept we're proud of and believe has real-world applicability.

🔮 What's Next for Farm Sutra

  • Amazon Nova 2 Sonic integration — Replace mock voice input with real multilingual speech-to-speech using Nova 2 Sonic, enabling truly voice-first interaction for low-literacy farmers
  • Nova 2 Omni — Video-based crop scanning, where farmers can record a 10-second video of their field for richer disease detection
  • Bank API integration — Connect the Credit Score tab directly to Kisan Credit Card loan applications
  • Government scheme push— Partner with state agriculture departments to deliver PM-KISAN and soil health scheme notifications directly to farmers

"Farm Sutra is not just an app. It's a voice for the farmer who never had one."

Built With

Share this project:

Updates