About AgriSahayak: AI Farming Assistant for Rural India

What Inspired Me

Growing up in Rajasthan, I've seen firsthand how farmers in my region struggle with basic agricultural decisions. My uncle, who has a small 1.5 hectare farm, often loses crops to diseases he can't identify quickly enough. He relies on traveling to the nearest agricultural center, which is 25 km away, just to get basic advice.

This got me thinking about India's massive agricultural challenge. We have over 600 million farmers, but most of them are small landholders who don't have access to the kind of expert advice that could dramatically improve their yields. Traditional extension services reach less than 10% of farmers effectively. With IBM's Granite models and agent framework becoming available, I realized there was an opportunity to build something that could actually make a difference.

What I Learned

Technical Learnings

Building this project taught me a lot about working with foundation models in real-world scenarios. I learned how to chain different AI agents together using IBM's Agent Development Kit, which was completely new to me. The biggest learning curve was understanding how to make Granite models work effectively with agricultural data and Indian languages.

I also discovered that multimodal AI integration is much more complex than I initially thought. Getting the system to seamlessly handle voice inputs in Hindi, crop photos, and text queries required careful prompt engineering and a lot of trial and error.

Domain Insights

Through interviews with farmers in my area, I learned that they don't just want generic advice. They need specific, actionable recommendations that consider their exact location, crop variety, and current conditions. I also realized that trust is everything in this space. Farmers have been burned by bad advice before, so any AI system needs to explain its reasoning clearly.

Voice interfaces turned out to be crucial. Many older farmers are more comfortable speaking than typing, and they often prefer communicating in their local dialect rather than formal Hindi.

How I Built It

Architecture

I built AgriSahayak using IBM's Agent Development Kit as the core orchestration layer. The system consists of several specialized agents that work together:

User Input → Input Processing Agent → Knowledge Synthesis Agent → Action Planning Agent → Response Generation Agent

Technical Implementation

Core Agent Framework: Used IBM's ADK to create a pipeline of specialized agents. Each agent has a specific role in processing user queries and generating recommendations.

Language Model: Fine-tuned IBM Granite-13B-Chat on agricultural datasets I collected from government sources and translated into Hindi. This was probably the most time-consuming part of the project.

Computer Vision Component: Built a crop disease detection model using transfer learning. I started with a pre-trained vision model and fine-tuned it on Indian crop disease datasets I found from agricultural research institutions.

Data Integration: Connected real-time APIs for weather data from IMD, market prices from eNAM platform, and government scheme databases.

Key Features I Implemented

Multimodal Input Handling: The system can process voice messages in Hindi/English, crop photos for disease detection, and text queries. I used speech-to-text APIs for voice processing and built custom image analysis pipelines.

Contextual Recommendations: Instead of generic advice, the system considers the user's specific location, current weather, soil type, and crop growth stage to provide personalized recommendations.

Action-Oriented Responses: Rather than just providing information, the system gives specific next steps. For example, instead of saying "your crop has fungal infection," it says "spray Mancozeb fungicide tomorrow morning before 9 AM, avoid watering for 48 hours."

Challenges I Faced

Data Quality Issues

The biggest challenge was finding quality agricultural data in Indian languages. Most research datasets are in English, and government data is often inconsistent or outdated. I had to spend a lot of time cleaning and translating datasets to make them usable.

Language Complexity

Hindi has so many dialects and variations. Farmers in different regions use different terms for the same concepts. I initially trained the model on standard Hindi, but realized I needed to incorporate regional variations and even some English words that farmers commonly use.

Making It Practical

My first prototype was too academic. It gave long explanations that farmers didn't want. I had to completely redesign the response system to be more direct and actionable. This meant understanding not just what information to provide, but how to present it in a way that busy farmers would actually use.

Trust and Explainability

During testing with local farmers, I realized they wanted to understand why the AI was making specific recommendations. I had to build an explanation system that could break down complex reasoning into simple, understandable steps.

Technical Constraints

Rural internet connectivity is terrible. My initial version required constant internet connection, which made it practically useless. I had to redesign the architecture to work offline for basic queries and sync data when connection is available.

Testing and Results

I tested the prototype with about 15 farmers in my area over two weeks. The results were encouraging:

• Most farmers could successfully use the voice interface after a brief demonstration
• Disease identification accuracy was around 78%, which is better than what farmers typically achieve on their own
• The government scheme lookup feature was particularly popular since most farmers weren't aware of all available programs

The main feedback was that farmers wanted more specific local information. Generic advice doesn't work in agriculture since conditions vary so much even within a few kilometers.

What's Next

This hackathon project is just the beginning. I want to expand it to include:

• Integration with satellite imagery for crop monitoring
• Peer-to-peer farmer networks for sharing local knowledge
• Direct connections to input suppliers for ordering seeds and fertilizers
• Financial literacy modules to help farmers make better economic decisions

The goal is to create something that actually gets used in the field, not just wins hackathons. Agriculture in India needs practical solutions, and I think AI agents can provide that if we build them thoughtfully.

Built With

• celery
• docker
• enam-api
• fastapi
• html/css
• ibm-adk
• ibm-granite
• ibm-watson
• imd-api
• javascript
• mongodb
• ntlk
• opencv
• postman
• python
• redis
• spacy
• sqlite
• tensorflow