Rebuilt my master's thesis regression model as a Graph Attention Network on the full 49K-observation NASA satellite dataset. On a true 20% spatial holdout, the GAT held R² 0.32 while the thesis's best baseline (GWR) collapsed to R² -85, because its original score came from leave-one-out CV that never tested new locations.

CatBoost vs. XGBoost vs. LightGBM on UCI Credit Card Default data, optimized for recall on the minority class. SHAP explainability and EU AI Act compliance documentation built in.

Most marketing analytics measures who converted, not whether prior contact made the difference. Prior contact is associated with higher conversion on the UCI Bank Marketing dataset. The association holds after adjusting for measured covariates, with segmentation and SHAP explainability.

A softball coach told me my whole life not to swing on a 2-0 count. Capstone thesis project using Markov Chains on MLB data to test whether he was right. He was. This is also how I ended up in data science.