🎓 MS in Data Science, UMass Dartmouth (August 2026) | 🔭 Open to Data Scientist / ML Engineer / Data Analyst roles
I build end-to-end machine learning systems — from raw data and feature engineering through deployment, serving, and monitoring. I care about models that are honest, reproducible, and defensible, not just high-scoring.
Languages & Core: Python (Pandas, NumPy, Scikit-learn, Statsmodels), SQL, Unix/Linux, Git, Azure DevOps, GitHub Copilot, Claude Code
ML / Modeling: LightGBM, XGBoost, CatBoost, Isolation Forest, K-Means, SHAP, Optuna
Experimentation: A/B Testing, Hypothesis Testing, Power Analysis, Two-Proportion Z-Test
LLM / AI: OpenAI API, LangChain, RAG, FAISS, Text-to-SQL, ReAct Agents, Prompt Engineering
MLOps: MLflow, Docker, FastAPI, GitHub Actions (CI/CD), Evidently (drift monitoring)
Data & Pipelines: PostgreSQL, DuckDB, PySpark, Kafka, Parquet
Cloud: AWS (EC2, S3, IAM), Azure Blob Storage, Hugging Face Spaces
Visualization: Streamlit, Tableau, Power BI, Matplotlib, Seaborn, Excel
End-to-end platform on 6M+ real deliveries (Cainiao LaDe, 5 cities). Two models running side by side: a supervised LightGBM disruption classifier (PR-AUC 0.22, 12.2x lift over baseline) and an unsupervised Isolation Forest anomaly detector. Includes a Kafka streaming pipeline, a FastAPI serving layer with a feature store, a hybrid Text-to-SQL + FAISS RAG assistant, a simulated A/B test (two-proportion z-test), Evidently drift monitoring, and a live operational dashboard deployed on AWS EC2. Also published as 3 Tableau dashboards and a 3-page Power BI report.
Key finding: 30 couriers (0.6% of the network) generated ~20% of all anomalies — turning 66K raw alerts into a 30-courier action list. Caught and corrected a rolling-window data leak that had inflated PR-AUC to 0.29, traced to 117 test-only couriers with impossible feature values.
LightGBM churn model on 970K users × 101 features (time-based validation, AUC-ROC 0.9481) trained on 392M+ listening events processed out-of-core in DuckDB. K-Means RFM segmentation into 5 behavioral cohorts, SHAP explainability, and ~100M TWD in CLV-weighted revenue at risk quantified. Deployed as a 6-page Streamlit app with GPT-4o-generated retention recommendations. Also published as 3 Tableau dashboards.
LightGBM delay classifier on 18.2M US domestic flights from 4 federal datasets, with strict temporal validation (AUC 0.8629, MAE 16.82 min). SHAP analysis identified aircraft-tail cascade features as top predictors, nearly doubling model lift (AUC 0.80→0.86); documented a 58% structural unpredictability ceiling and a $2.78B annualized savings estimate. FastAPI + Docker + GitHub Actions CI/CD. Also published as 3 Tableau dashboards.
LightGBM global forecasting model (Tweedie objective, WRMSSE 0.4656) across 30,490 item-store series with 51 engineered features, beating AutoARIMA (0.7156). PostgreSQL-backed business insights and a LangChain + GPT-4o-mini ReAct agent enabling natural-language queries without writing SQL. Also published as 2 Tableau dashboards.