Citi Bike demand forecasting, end to end

Predict how many rides start at each Citi Bike station in the next six hours, and keep the prediction fresh every day without anyone running a notebook.

• ›Feature pipeline: raw monthly trip files from the public S3 bucket, cleaned and aggregated per station and time bucket, written to a Hopsworks feature group on a schedule.
• ›Training pipeline: LightGBM on lagged demand features, tracked with MLflow, with the winning model pushed to the Hopsworks model registry.
• ›Inference pipeline: pulls the latest features, writes predictions back to a feature group, and a Streamlit monitor plots error by hour against what actually happened.
• ›All three run as chained GitHub Actions workflows, so the whole thing is reproducible from a clean checkout.

Mean absolute error went from 139 rides for a naive baseline, to 45 for a last-four-weeks average, to 34.8 for LightGBM, to 34.4 after tuning. The tuning step mattered less than the feature work.

This started from a course template for NYC taxi demand. The adaptation to a different dataset, the station-level features, the zone geometry and the monitoring are mine; the pipeline shape is the course's.