Agri-Fincaster
A crop-yield forecasting system built on public government agricultural data — made for the CBSE Science Exhibition 2024–25.
What it is
Farmers plan crops largely on experience and intuition, even though decades of yield, rainfall, and soil data sit in public government datasets. Agri-Fincaster turns that data into something usable: give it a region and a crop, and it predicts the expected yield and suggests a planting window.
The goal was never to replace a farmer's judgement — it was to put the data on the table next to it.
How it works
- Public agricultural datasets collected and cleaned with pandas — by far the largest share of the work.
- Feature engineering across region, season, rainfall, and soil variables.
- Regression models in scikit-learn — we compared linear models against a random forest and kept the forest.
- A simple query interface: region + crop in, prediction + planting window out.
What broke, what I learned
The first model looked brilliant and was lying: it had memorised the training data. Fixing that meant learning train/test discipline properly — hold-out sets, cross-validation, and being suspicious of accuracy that seems too good.
The other lesson came at the exhibition itself: explaining a model to non-technical judges taught me to talk about features ("rainfall matters this much") instead of algorithms. If you can't explain the prediction, it isn't finished.