Built for a low-end phone and a weak signal.

Agricultural AI that requires a stable connection, a modern handset and English literacy has excluded most of the people it was ostensibly built for.

01 / What is hard here

The problem underneath the problem.

The technical constraints here are unusual and they dominate everything else: intermittent connectivity, entry-level devices, multiple languages and scripts, and users for whom a wrong recommendation has consequences measured in a season's income.

Ground-truth data is also genuinely scarce and expensive. Yield outcomes arrive once a year, disease labels require an agronomist, and the conditions vary enough between districts that a model trained in one region transfers poorly to another.

02 / What we build

Work we take on in agriculture.

Crop and disease identification: On-device image models that run offline on entry-level hardware, with explicit abstention when the image is unclear rather than a confident guess.
Yield estimation: Satellite and weather data combined with plot-level history for area and yield forecasting at district and aggregator scale.
Advisory systems: Irrigation, input and timing guidance in local languages, designed to be useful when delivered by voice or SMS rather than assuming an app.
Supply chain and grading: Quality assessment, sorting support and demand forecasting for aggregators, processors and FPOs.
Weather-linked risk: Modelling for crop insurance and lending, where the underwriting decision depends on both the forecast and its uncertainty.

03 / Constraints

What shapes the build.

These are design constraints agreed at the start, not compliance work bolted on before launch. Retrofitting them is expensive and usually incomplete.

Offline first: Core functionality works with no connection and reconciles later. This is a requirement, not a degraded mode.
Device reality: Models sized for entry-level Android hardware, with capture guidance that assumes an ordinary camera and imperfect conditions.
Language and literacy: Local languages, and voice or pictorial interfaces where reading cannot be assumed.
Cost of being wrong: A confident wrong recommendation can cost a season. Abstention thresholds are set conservatively and deliberately.

04 / What we advise against

Where we would say no.

We would advise against any advisory product that cannot function offline, and against yield or disease models deployed outside the agro-climatic zones they were validated in without revalidation.

We would rather lose the work than build something that creates a liability for you or a harm for the people on the other side of it. If that rules out what you had in mind, it is better established now than at delivery.

05 / Other sectors

Adjacent work.

Different domains, the same delivery discipline — a measured baseline, a real evaluation set, and monitoring that catches decay early.

Next step

Working in agriculture?

Tell us which process is slow, expensive or error-prone. The diagnostic will tell you whether this is a problem worth solving with machine learning — including if it is not.

Request a diagnostic All industries