AI-Based Market Intelligence Systems for Farmer Collectives: A Case Study from India
📚 Published in ACM Digital Library
💰 Google AI for Social Good – Funding
This project is a sub-part of a larger initiative under the Google AI for Social Good program in collaboration with:
- ACT4D – IIT Delhi
- Gram Vaani
- Centre for Collective Development (CCD)
It focuses on improving market outcomes for smallholder farmers by building AI-assisted forecasting and recommendation tools for crop price decision-making, specifically targeting non-perishable crops like soybean.
-
Digitalization Tool:
Developed Android-based ODK forms for CCD field officers to collect mandi price data, eliminating manual recordkeeping. -
Market Intelligence Development:
Designed the foundation for a digital market surveillance system using the CoRE Stack (Commoning for Resilience and Equality). -
Recognition:
Awarded Honorable Mention for Best Presentation at the Google AI4SG Mid-Program Workshop for impactful field deployment and iterative product design.
🎓 Workshop Projects
The core frontend system was developed as a mobile app called Rythu Vaani, built for deployment across farmer cooperatives supported by CCD.
-
Historical Pricing Trends:
Allows comparison of Adilabad mandi prices with surrogate mandis to support more informed sale decisions. -
Forecast & Recommendations:
Presents top 3 recommended sale dates based on AI-generated price forecasts. -
Review Past Advice vs Reality:
Lets field officers and cooperatives verify the accuracy of past recommendations against actual market outcomes. -
Lightweight Design:
Works offline and syncs with Firebase when connectivity is available. Designed for low-resource rural environments. -
Field Tested:
Actively deployed across 16 cooperatives. Iterated based on weekly calls with CCD staff to improve usability.
-
LSTM (Long Short-Term Memory):
Recurrent model used to learn long-range trends in mandi prices. -
Temporal Convolutional Network (TCN):
Used for short-term multivariate time-series forecasting. Performed better than LSTM in both accuracy and training stability.Inputs:
- Local + surrogate mandi prices
- Arrival volumes
- Seasonality encoded as cyclic features (day-of-year)
- Built a 5-model TCN ensemble
- Incorporated Prospect Theory to rank sale dates, simulating risk aversion under price uncertainty
- Metrics used for evaluation:
- Probability of Accurate Prediction (PAP)
- Net Gain (NG)
- Oracle Gain (upper bound)
These models power the recommendations shown in the Android app.
| Resource | Link |
|---|---|
| 🏗️ System Architecture (Figma) | View |
| 🔁 Application FlowChart (Figma) | View |
| 🎥 Google AI4SG Workshop Presentation | YouTube |
| 🖼️ App UI Screenshots | Google Drive |
| 📽️ Project Slideshow | Slides |
| 📱 APK Download | Download APK |
| 🔥 Firebase Console | Firebase |
- Install Android Studio
- Clone this repository
- Open the project inside the
android app/directory - Sync Gradle, connect your emulator/device, and run the app
-
Earlier debug versions of the APK were large. For deployment, it's advised to create a signed release APK with a private keystore.
🔐 App Signing Guide -
The original version used CSV files as lightweight structured data pushed to Firebase. We are currently migrating to a more stable Room database.
Built for real-world deployment with smallholder farmers, this app brings AI-assisted decision-making directly into the hands of rural cooperatives.