A gradient boosting model on match data, built as a study in calibration rather than a chase for accuracy.
A gradient boosting model predicting IPL match outcomes from historical data, taken end to end: collection, cleaning, and feature engineering.
The interesting result was not the accuracy number. Sports data is noisy and low-signal, and a model that looks confident on it is usually wrong in a way that matters. So this became a study in calibration and in the limits of the data, comparing feature sets against each other rather than optimising for a headline figure.