Starbucks offer targeting: customer segmentation (K-Means, k=4), a send-time model of offer completion rebuilt after finding leakage in the first version (0.994 AUC fell to 0.865 on unseen customers, well calibrated), causal inference, and a rule-based recommender with +7.9% lift.
How it knows: A test deletes every event after each offer send and fails if any model feature changes, and the leakage ladder measures how much each leak in the first model was worth.
Nicholas Smith - all work