Projects

A short selection of technical projects. The repositories carry the details.

Yahtzee agent built with JAX/Flax/RLax, trained through self-play with stochastic MuZero-style MCTS and explicit chance nodes for dice rolls.

JAX · Flax · RLax · MCTS · self-play

81.1% win rate vs heuristic baseline over 512 games

Solar irradiance forecasting as a multi-horizon task, comparing GRU, LSTM, seasonal naive, and persistence baselines.

GRU · LSTM · time series · baselines

GRU reduced RMSE by ~34% at 4h and ~21% at 24h vs strongest naive baseline

Fraud detection model for highly imbalanced credit-card transactions, focused on precision/recall tradeoffs and useful threshold selection.

XGBoost · imbalanced classification · evaluation

Precision: 0.95, Recall: 0.87

Classification pipeline for customer churn prediction with feature engineering, model comparison, and FastAPI deployment.

scikit-learn · XGBoost · FastAPI

Recall: 0.711, ROC AUC: 0.842