Built a 42-day sales forecasting pipeline over 1M+ records (1,115 stores), cutting error to 8.63% MAPE against an 18.56% seasonal baseline (53% reduction), with generalization confirmed on a sealed test window and a Kaggle private-leaderboard submission (0.1148 RMSPE, within 1.6% of the local estimate). Engineered 41 horizon-safe features to predict all 42 days in a single pass without recursive feedback, and ran MLflow ablation and seed studies to separate real feature contribution from run-to-run noise. Guarded correctness with a two-path leakage test and a train/serve equality check that caught a silent categorical-encoding skew returning wrong forecasts under a 200 status; deployed as a containerized FastAPI service on Cloud Run (~120ms latency).

Built a Selenium/undetected-chromedriver scraper collecting 4,200+ film and series records from a Cloudflare-protected, JavaScript-rendered site, with a canvas-based workaround for a poster CDN blocking direct requests. Designed a reproducible 5-stage cleaning pipeline (4,200 to 1,648 records) and trained a Random Forest rating classifier with leakage-aware feature engineering and stratified 10-fold cross-validation (ROC-AUC 0.87), plus K-means clustering and a Mann-Whitney U test.

Built a RAG pipeline over PDFs using recursive text splitting, text-embedding-ada-002 embeddings, and PineconeDB for semantic retrieval. Implemented chat history condensation for query rewriting prior to retrieval, preserving context across multi-turn sessions.

Mapped 25,000 Spanish pharmacies into an interactive tool for client segmentation: a Python scraper captured location, contact, and operational metadata from government sources with dynamic content handling and validation, feeding a MapBox GL JS interface with clustering and filtering over the geospatial distribution.