Subscriber Churn Prediction Engine

Owned an end-to-end churn modelling system for a subscription business with 3.4M active users. Engineered 180+ behavioural features from event logs in Spark, trained and calibrated a gradient-boosted ensemble, and shipped nightly batch scoring into the CRM. Built SHAP-based explanations so the retention team could see why each account was flagged, and ran a two-arm holdout to prove incremental lift rather than correlation. Wired drift monitors on the top features so silent data shifts page us before the model rots.