My Journey: Building AgriSathi

What Inspired Me

Living in a country where agriculture is the backbone of the economy, I noticed a painful irony: while the internet is overflowing with data, the people who need it most—our farmers—are often the most disconnected. I realized that for a farmer in rural India, the "Digital Divide" isn't just about a lack of internet; it’s a barrier of literacy, language, and hardware.

I asked myself: What if a farmer didn't need a smartphone or data pack? What if they could just pick up any basic phone, dial a number, and talk to the internet in their own mother tongue? That "what if" became AgriSathi. My mission was to turn a simple phone call into a gateway for expert agricultural wisdom.

How I Built It

I designed AgriSathi as a first-of-its-kind, end-to-end Voice-to-Voice AI platform. I focused on making the interaction as natural as talking to a neighbor.

The Technical Workflow

Telephony & Orchestration: I used LiveKit to handle the SIP trunking, allowing farmers to connect via a standard cellular call and their AI agent SDK for entire Agentic Orchestration.
The Speech Pipeline:
- STT (Speech-to-Text): I integrated Deepgram Nova-3 for its exceptional multilingual support, converting local dialects into text.
- The Brain (LLM): I utilized Gemini 3 for ultra-low latency processing, coupled with the excellent reasoning to appropriately call the tools and talk to the user in a respectful and easy manner, making sure the answers are properly formatted and are in a multilingual way for the TTS module to convert it into voice.
- TTS (Text-to-Speech): To give our agent a human-like voice, I used Cartesia Sonic 3, which provides high-quality, multilingual vocal outputs.
Information Retrieval: The system isn't just a chatbot; it's a researcher. It uses Google Web Search APIs to fetch real-time Mandi prices, weather forecasts, and the latest policy schemes.

Cost Efficiency

I optimized the architecture to keep the total estimated cost at approximately $0.0698/min, ensuring that the solution remains scalable and economically viable for public impact.

Challenges I Faced

The road wasn't easy. Our biggest hurdle was Latency. In a voice call, a delay of even two seconds feels like an eternity. I had to move toward an asynchronous programming model using Python and specialized LLMs like Gemini to ensure the "Voice-to-Voice" loop felt instantaneous.

Another challenge was Context Retention. Farmers often call back with follow-up questions. I solved this by implementing a Database system that generates a conversation summary after every call, allowing the AI to "remember" the farmer’s previous issues (like a specific pest infestation) when they call again.

What I Learned

This project taught me that accessibility is the highest form of innovation. Building a complex LLM is one thing, but making it accessible to someone with a ₹1,200 feature phone is where the real impact lies. I learned how to orchestrate multiple cutting-edge AI tools (LiveKit, Gemini) into a single, seamless pipeline that bridges the gap between high-tech AI and the grassroots level of Bharat.