Most farmers in India already carry the one device that matters for this story: a phone with WhatsApp on it. Many are more comfortable speaking than typing, and more comfortable in Hindi or their own language than in English. Agricultural advice, on the other hand, has tended to arrive as printed leaflets, long web pages or calls to a helpline during office hours.
Krishi Sathi brings the advice to where farmers already are. It is a WhatsApp assistant, built on BharatGen's AI models, that a farmer can simply talk to. This case study covers what it does, how a conversation with it works, and what building it taught us.
What Krishi Sathi does
Krishi Sathi answers the everyday questions that decide how a season goes. It runs on a small language model trained for agriculture, on a body of farming knowledge that covers planting windows, soil conditions, water needs and pest management.
- Pest management. A farmer describes what they are seeing on the crop and gets advice on what it could be and what to do about it.
- Crop yield prediction. Guidance on what to expect from a crop, so that decisions about inputs and selling can be planned.
- Everyday farm questions. Storage, irrigation and the practical problems that come up between sowing and sale.
The first version focused on two crops grown widely around Nashik in Maharashtra: onions and grapes. Krishi Sathi can work in all 22 scheduled Indian languages. The first demo was built in Hindi and English.
Built around the voice note
Every design choice starts from how farmers already use their phones.
- WhatsApp, not a new app. There is nothing to download and no account to create. Krishi Sathi is a contact in the farmer's chat list.
- Voice in, voice out. The farmer records a voice note, just as they would for a relative. Krishi Sathi turns it into text, shows the farmer what it heard, and replies in writing and as a voice note. That makes it usable for farmers who read with difficulty, and lets everyone check that they were understood.
- Clarifying questions. When a question is vague, Krishi Sathi asks before it answers, rather than guessing.
- A simple menu. Typing M at any point brings back the main menu.
A conversation with Krishi Sathi
Here is a real exchange, translated from Hindi. The farmer's messages were voice notes.
| Who | What was said |
|---|---|
| Farmer | They're kept in the storeroom, but now they're getting old. What can I do to store them well, so they don't spoil? |
| Krishi Sathi | When you say "they", what do you mean? Are these seeds, fruit and vegetables, or something else? |
| Farmer | Onions. |
| Krishi Sathi | Understood. Are these onions kept in a cool, dry place, or somewhere else? |
| Farmer | In the storeroom. |
| Krishi Sathi | Understood. To keep onions from spoiling in storage, keep them somewhere well ventilated, out of direct sunlight, at 10°C to 15°C. Keep them away from moisture and humidity. Would you like more detailed advice on storage? |
The farmer's first message, as Krishi Sathi transcribed it: स्टोर रूम में रखे हुए है लेकिन वो अभी पुराने हो रहे है क्या कर सकते हो उनको अच्छे से रखने ताकि वो खराब ना हो।
Two things in this exchange are worth noticing. The opening question never names the crop, and Krishi Sathi asks rather than assumes. And the answer is short and practical: a temperature range, what to avoid, and an offer to go further.
See it in action
This demo video shows Krishi Sathi at work.
What farmers taught us
The team trained Krishi Sathi on growing crops: when to plant, how to manage soil and water, how to deal with pests. Then a farmer near Nashik, someone who had farmed all his life, sat down with it and asked for help storing his onions.
Nobody had planned for that question, and it is a good one. For an onion farmer, a harvest that rots in storage is lost just as surely as one that fails in the field. It was an early lesson that has shaped how BharatGen works since: you cannot fully anticipate what people need until you are in the room with them. Models that do well in evaluation still meet phrasings, problems and situations in the field that nobody thought to test.
That is why we work closely with people who know a domain from the inside, and why we test with real users early, before thinking about scale.
Why it matters
Farming is the livelihood of a very large share of India's people, and it is full of decisions that good, timely information can improve. Much of the knowledge farmers need already exists. The hard part is getting it to them in their language, in a form they can use, at the moment they need it.
Krishi Sathi showed that an assistant built on AI models made for India can do that over the most familiar app on a farmer's phone, using their voice. More on the AI models behind work like this is on our text AI models page.
Since this post, our agriculture work has continued with AgriParam, a language model for Indian agriculture, and BhashaBench-Krishi, a benchmark that tests AI models on Indian agricultural knowledge.
