Kisan Saathi - Hindi Voice Agent for Crop Triage
Link to open source: https://github.com/yashtomer/kisan-saathi
Link to Live Project: https://kisan-saathi-theta.vercel.app
Kisan Saathi is a Hindi-first voice agent that helps Indian farmers triage crop problems just by talking - no typing, no app install, no English required. A farmer describes what they are seeing in their field, the agent asks follow-up questions in Hindi, and comes back with a likely cause and practical next steps.
Most farm advisory tools quietly assume a literate, English-comfortable user with a steady data connection. Kisan Saathi flips that assumption: voice in, voice out, in the farmer's own language.
The part that matters most is what happens when the agent is not sure. Instead of guessing, low-confidence cases are escalated to a human agronomist who joins the same live channel and takes over the conversation. The AI absorbs the routine volume, and experts spend their time on the cases that actually need them.
It is built on the Agora Conversational AI Engine, using RTC for real-time audio and RTM as the data channel for live transcripts and agent-state events. Speaker-identified live transcription, an agent visualizer that surfaces conversation state, and per-stage latency metrics (AGENT_METRICS) make the whole voice pipeline observable end to end.
Stack: Next.js (App Router), React, TypeScript, Tailwind CSS, Agora RTC + RTM, and Agora-managed STT/LLM/TTS with BYOK support for Deepgram, OpenAI and ElevenLabs. Open source under MIT.
