Rice Yield Forecasting for East Java Cooperatives
A forecasting service and dashboard that tells cooperative managers how much rice to expect, three weeks before harvest.
7% median error · 40 cooperatives
Problem
Cooperatives sold rice forward without knowing what the harvest would bring, so they either over-committed or left money on the table.
My role
I owned the system end to end: the schema, the ingestion jobs, the API, the dashboard and the model in production.
What I built
Nightly ingestion of rainfall and planting records into PostgreSQL
A gradient-boosted baseline, then a small PyTorch model that beat it by 18% MAPE
A FastAPI service with versioned models and a Next.js dashboard for managers
Outcome
Forecasts arrive three weeks before harvest with a median error of 7%. Forty cooperatives use it every week.
Want something like this built? Let's talk →