CitiBike Demand Forecaster:Recursive ML & MLOps Pipeline
End-to-end ML system for 24-hour City Bike demand forecasting using LightGBM with a recursive multi-step prediction engine, automated Champion/Challenger MLOps pipeline, and a full-stack Next.js dashboard.
Core Impact
"Engineered a high-precision recursive forecasting engine with automated drift detection and Champion/Challenger model promotion."
Architecture Breakdown
Developed a Recursive Bridge Forecasting Engine using LightGBM to provide live 24-hour bike demand predictions, implementing a custom iterative logic that autonomously bridges 20-day historical data gaps by feeding sequential hourly model outputs as inputs for real-time inference.
Architected an Automated MLOps Pipeline orchestrated via GitHub Actions, featuring a Champion/Challenger model promotion strategy, Evidently AI for data drift monitoring, and MLflow for experiment tracking, ensuring only performance-validated models are promoted to production.
Engineered a Scalable Streaming Feature Pipeline that sequentially processes 12+ months of raw City Bike data (millions of records), optimized for resource-constrained environments by implementing a sliding-window download/process/delete cycle and calculating 28-hour feature lags for time-series context.
Implemented a Full-Stack ML Deployment utilizing Next.js and Tailwind CSS to visualize model intelligence, surfacing pre-computed S3 Parquet datasets and geospatial station demand (Leaflet) while providing deep-link transparency into the model's recursive logic and pipeline health for stakeholders.
Systems Analysis Concluded