CalvinLeo.

Deployed an end-to-end Time-Series Demand Forecasting System using Scikit-learn, shifting logistics from static volume forecasting to real-time, container-level prediction. Engineered a custom, reusable ContainerFeatureTransformer with cyclic trigonometric encoding (Sine/Cosine), securing 0% data leakage via strict Walk-Forward Validation. Developed a 'Dual Inventory Analysis' framework (All vs. Recent 90-day MTA) to identify and filter systemic Ghost Inventory, optimizing automated SURPLUS/DEFICIT logistics signals. Modernized legacy PO/WO systems into a modular React & TypeScript architecture, optimizing PostgreSQL databases to secure sub-second query latency for high-concurrency operations. Built a scraping pipeline with FastAPI and LLMs (Kobolt/Gorq APIs) for automated business news summaries, and integrated Leaflet.js with BMKG XML weather feeds for vessel risk mitigation. Established production-grade monitoring using Kolmogorov-Smirnov (KS) tests for live Feature Drift Detection, orchestrating all services via Docker and Nginx reverse proxies.

Engineered a robust credit risk classifier using LightGBM; applied SMOTE to resolve severe class imbalance (target ratio 11:39), boosting minority class recall by 18% with zero data leakage. Built an interactive Streamlit UI with SHAP and LIME to interpret risk drivers (credit history, loan amounts), ensuring regulatory compliance and audit-readiness for real-world deployment.

Designed a 3NF normalized schema across 4 tables and engineered a Python ETL pipeline to clean, validate, and migrate over 540,000 transaction records with optimized batch processing.