Feature Drift Detection

Production models degrade silently when the data they see drifts away from the data they were trained on. This project is an end-to-end monitoring pipeline that compares a reference dataset against new production batches and flags the features whose distributions have shifted.

Drift is measured feature by feature with the two-sample Kolmogorov–Smirnov statistic — the largest gap between two empirical CDFs — which I implemented from scratch with vectorized NumPy instead of calling a library. Per-feature thresholds let important features alert at smaller shifts.

The pipeline covers data preparation, feature engineering, a baseline model, and monitoring experiments on a real retail dataset, with an interactive Streamlit dashboard for visualizing distributions and reporting drift scores.