I build production-grade analysis that ends in a decision — quantitative risk, NLP, fraud detection, and the MLOps that keeps models honest in production.
Three Value-at-Risk methodologies on an 8-asset portfolio, stress-tested through the COVID crash and the 2022 rate shock — ending in a concrete rebalancing recommendation.
A 5-layer dbt + DuckDB warehouse turning raw seeds into a business-facing metrics layer, with automated testing baked into every model and a published lineage DAG.
A spaCy → TF-IDF → XGBoost pipeline that separates fabricated from real news articles, packaged as a paste-an-article web app for live inference.
Five model/resampling combinations benchmarked on a 0.17%-fraud dataset by PR-AUC, then a business cost function that turns a decision threshold into dollars per day.
A simulated GDPR/CCPA audit on 50k records, deliberately seeded with violations so every detector, scanner, and dbt masking model can be proven to catch them.
A drift-detection pipeline that caught a distribution shift weeks before any performance metric moved — PSI + KS tests, an Airflow loop, and a live Streamlit dashboard.
A full-stack finance app on Plaid Production: a TypeScript cleaning pipeline, a 7-detector alerts engine, and an explainable 0–100 financial health score, synced daily.