Making Machine Learning Usable for Farmers: Lessons from the IWOFA Advisory System

A model is not yet a product

IWOFA, the Intelligent Web-based Outcome-driven Farm Advisory System, started as a research project: a model that predicts crop yield from information such as location, planting date and weather. A good model on its own does not help a farmer. People need a tool they can open on a phone, give simple inputs to and trust. Our job was to bridge that gap. IWOFA is a research prototype and its predictions are advisory, not a guarantee.

The architecture in brief

The system has two parts. A Python service built with FastAPI loads the trained models, one per crop, and exposes endpoints for forecasts, scenarios, history and administration. A React front end offers four areas: Forecast, Scenario, History and Admin. Farmers can use it as an installable progressive web app, so it can be added to the home screen and keep working in a limited way when the connection is poor.

Honest uncertainty

The most important design choice was to show uncertainty instead of hiding it. Every prediction comes with an interval, and the interval is calibrated so that it behaves honestly. If a user provides only a few details, the interval becomes wider. If they supply more detail, it narrows. We call this the tiered-input principle: the system says plainly how much it knows given what it has been told. A wide range is not a failure; it is the model being truthful that it cannot be precise without more information.

Asking for only what people know

Farmers will not fill in long forms. The interface asks for the essentials first and offers an optional section to add detail for a narrower prediction. Missing fields are filled sensibly in the background, and the explanation shown to the user describes what that means for confidence, in plain language.

Scenario and history tools

A single number is less useful than the ability to compare options. The Scenario tab lets users change inputs, such as planting date, and see how the forecast moves. The History tab keeps a record of earlier forecasts so users can review them against what actually happened.

Validation with stakeholders

Because IWOFA supports a research study, it includes tools for collecting feedback from stakeholders in a structured way, with participants identified only by codes. An administrator area manages the workflow, including rebuilding and promoting models as a separate, deliberate step.

Lessons for anyone shipping machine learning

• Design the interface around decisions the user actually makes.
• Communicate uncertainty clearly and in plain language.
• Make the system degrade gracefully when information or connectivity is limited.
• Keep model training, serving and administration separate and auditable.
• Say what the tool is and is not, including that predictions are advisory.

Build something similar

If you have a model, a spreadsheet of expertise or a research prototype that needs to reach real users, we can help turn it into a dependable web or mobile tool. See IWOFA in our portfolio.