Field signals, farmer judgement: designing agricultural decision supportField signals, farmer judgement: designing agricultural decision support

India has about 146 million operational farm holdings, and roughly 86% of them are small or marginal, under two hectares, according to the Agriculture Census figures cited in Krishyak's briefing. On a farm that size, one bad decision about seed, a spray or a sale can wipe out a season's margin.

Those decisions are rarely made with poor judgement. They are made with poor information, scattered across too many places. This note sets out the design principles we followed building Krishyak, and why we think agricultural software should support a farmer's judgement rather than replace it.

Five decisions, five different sources

Every season, a small farmer answers roughly the same questions:

Decision	Where answers often come from	What goes wrong
What to sow, how much to spend	Input dealers, last season's habit	Spend is fixed before risk is understood
Is this leaf diseased?	Neighbours, shopkeepers, helplines	Late or generic diagnosis, the wrong spray
When will rain, heat or pests arrive?	TV, generic weather apps	Forecasts never become field actions
When and where to sell	Local traders, word of mouth	Little view of mandi prices or MSP
Which scheme applies	Portals, CSC operators, agents	Eligibility and documents are unclear

The farmer is the only place where all five answers meet. Most digital tools address one row of this table. The design question we started from was different: what if the farm record were the place they met?

Principle 1: one dated record per field

Everything in Krishyak hangs off a field and a crop season: the boundary or area, the crop and sowing date, notes, photos, soil readings, weather, satellite indices and what the farmer actually did. Every item keeps its date and source.

This sounds administrative, but it changes what advice can be given. A falling moisture index means something different in week 2 than in week 12, and something different again if the farmer recorded an irrigation yesterday.

Principle 2: a signal is a reason to look, not a verdict

Satellite indices such as NDVI and NDMI are genuinely useful at 10-metre resolution, but they are not diagnoses. Cloud cover, a changed boundary or an uneven sowing pattern can all move the numbers. Krishyak only raises an inspection prompt when enough clear pixels support it, and it phrases the prompt as a reason to walk to the corner of the field, not as a conclusion.

The same applies to the crop-photo classifier. On public datasets it reached 93.71% accuracy on a held-out test set of 8,566 images across 38 classes and 14 crops. That is a laboratory-style result, not field accuracy, and the model is built to behave accordingly: it abstains below 0.75 confidence or when the photo does not match the selected crop, it gives no pesticide doses, and it always says who to confirm with before spraying.

Principle 3: show the range, not a single number

Farming outcomes are uncertain, and software that hides the uncertainty is not being helpful. Krishyak's season simulator runs 500 Monte Carlo scenarios per plan, varying rainfall, pest pressure, fertilizer response and price, and shows current, optimised and worst-case plans side by side. In the illustrative 2-hectare example from the briefing, the worst case was a loss. That is exactly the information a farmer needs before committing money.

Principle 4: language is infrastructure

A decision tool that only works in English excludes most of the people it is meant for. Krishyak ships language packs for all 22 scheduled Indian languages plus English, as static versioned packs that work offline, with right-to-left support for Urdu, Kashmiri and Sindhi and native scripts for Manipuri and Santali. Speech in and out is available with consent and a written fallback. Native-speaker review of every pack is still on the roadmap, and we say so.

Principle 5: consent and deletion are features

Farm data is personal and economically sensitive. Purpose-based consent, data export and verified deletion are built into the product rather than added later, and aggregated reporting for institutions suppresses small groups so individual farms cannot be singled out.

What has not happened yet

Krishyak is pilot-ready, not field-proven. Independent field validation, a funded pilot, institutional partnerships and revenue have not happened yet. The next honest step is a carefully consented pilot with an FPO or NGO, with agronomist review and a field benchmark on Indian farm photos.

We think that order matters: build the product, test it, verify the pipelines live, then earn the right to claim impact in the field.

If you run a programme for farmers and want to explore a pilot, we would like to hear from you.