Small Models, Real Impact: ML for Smallholder Farms

You don't need a giant model to help a farmer decide when to plant. You need good data, honest evaluation and something that runs on modest hardware.

Data first

Satellite vegetation indices and local weather explained more variance than any architecture change we tried.

Evaluate like a skeptic

We backtested by season, never with a random split, because random splits leak the future into training.

Track everything

MLflow keeps every run, parameter and metric, so we can explain exactly which model produced a forecast.

Keep humans in the loop

Forecasts go to extension officers first. Their feedback has improved the model more than any hyperparameter search.