Two Desais on Krishi.System: where AI advice for farmers actually comes from

Pratik sat down with Sachi Desai (VP of AI GTM and Partnerships at The Climate Corporation) and Venky Ramachandran for a 30-minute conversation on Krishi.System. The episode is titled Artificial Intelligence Deepdive with Two Desais, and Venky’s framing is the part of the conversation I want to surface here.

The question Venky opens with is not the one most AI-in-agriculture conversations land on. The usual question is “can AI deliver better advice to farmers?” Venky’s reframe is sharper: what is the AI advice drawn from?

The argument runs like this: medicine, law, and finance digitised their knowledge a generation ago, while the knowledge that actually runs farms stayed oral, local, deeply contextual, passed on generation to generation by working the same land together. Almost none of it is on the internet, which means almost none of it is in any training corpus.

That chain is breaking now. In large-holding contexts like the US and smallholding ones like India alike, the next generation does not want to farm. The transfer of tacit field knowledge is slowing. So when a farmer asks a model something, the question of what the model is retrieving from becomes the question.

Sachi and Pratik bring two complementary answers to that question.

Sachi comes from a world where the gap is less about farmers not knowing things and more about wanting confirmation before high-stakes decisions. A soy farmer in Illinois calls her advisor not because she is uninformed but because farming is capital-intensive and irreversible. Talking through a decision is how she builds confidence to act.

Pratik comes from a world where neither the extension officer nor the model is present in any meaningful way. When the team trained the first version of Dhenu in 2023, the training set was built from hundreds of thousands of real farmer exchanges, and the corpus has grown well past a million since, because the way a smallholder phrases an agricultural question is almost nothing like the way it appears in any text online. Venky picks out a specific example from the episode: “jilli” in a particular dialect of Marathi refers to a caterpillar pest at a specific lifecycle stage on a cotton crop, and the word means something different in a soybean context three districts away. A generic language model can only give a satisfactory answer to someone who does not know better. A grower who knows cotton will immediately see where the answer falls apart.

The business model layer is where the two perspectives converge. Both Sachi and Pratik operate on a B2B logic. The AI advisory layer is deployed to agribusinesses, input companies, retailers, and cooperatives who then surface it to farmers. The farmer interacts with an AI that has been configured and constrained by a business whose commercial interest is not identical to the farmer’s. We live this in every deployment conversation I sit in: the retailer wants the assistant to move product, the farmer wants the answer that saves the crop, and the configuration work is where those two interests get reconciled. That is the constraint the design has to live inside, and the part of the conversation that Venky paywalls is where the two of them got into the harder parts of resolving it.

Listen to the whole conversation at krishidotsystem.com/p/artificial-intelligence-deepdive. Venky’s writing on Krishi.System is consistently the clearest thinking I’ve seen on systems-level questions in food and agriculture; the post is worth the subscription if those questions are your work too.