Citibike Demand Prediction System

Design a system that predicts Citi Bike demand for every station in NYC for the next 4 hours, updated every 15 minutes. The prediction target is the number of bikes that will be checked out from each station in 15-minute buckets. The system must ingest real-time trip data, weather data, and event calendars; engineer temporal, spatial, weather, and network features; train and serve a regression model at scale; and expose both batch predictions (for rebalancing trucks) and low-latency single-station queries (for the mobile app). Discuss data ingestion, feature store, model training pipeline, serving infrastructure, monitoring for drift, and how you would handle cold-start stations and special events like concerts or snow storms. Assume 1,000+ stations, 150k daily trips, and a peak of 1k trips/minute.