Built an end-to-end ML pipeline on Google Cloud Platform to predict customer churn for a telecom provider losing ~26% of customers annually. The platform ingests customer events in real-time via Cloud Pub/Sub, engineers churn-predictive features in BigQuery using defensive ELT patterns, and serves predictions through Cloud Functions loaded with a trained XGBoost classifier (ROC-AUC: 0.84+). Key engineering decisions: - Pub/Sub fan-out architecture — single topic feeds BigQuery, Dataflow layer, and Cloud Functions independently - Free-tier inference pattern — Cloud Functions + joblib artifact replaces Vertex AI endpoints at $0 cost - Feature parity between BigQuery SQL views and Python training pipeline ensures consistent scoring