Churn prediction model for a subscription meal-kit service

Every month roughly 7% of our meal-kit subscribers cancel, and today we only find out when they hit the cancel button. We want a weekly churn-risk score per subscriber so our retention team can intervene earlier with pause offers or box swaps. All behavioral data (orders, skips, support contacts, delivery issues) sits in a reasonably clean PostgreSQL warehouse going back three years. Deliverables are concrete: an exploratory analysis writeup identifying the strongest churn signals, a trained model (we lean gradient boosting for interpretability, open to arguments), a scoring script that runs weekly and writes scores back to a table, and a one-page explanation of the top risk factors our non-technical retention lead can act on. No production ML infrastructure needed - a scheduled Python job is fine for now. We are a 30-person company in Denver, mostly async, Slack for questions. Our data analyst can pair with you on warehouse quirks during your first week.