Projects

Framed PET random coincidence correction as binary edge classification on detector hit graphs. Designed a 3-stage GNN (linear encoder → 2-layer GATv2 → MLP classifier) trained on 3.25M labeled graphs from GATE Monte Carlo simulations. Achieved 92.9% precision, 92% recall, F0.5 = 0.927 on test set.

Implemented a reinforcement learning agent using Monte Carlo Counterfactual Regret Minimization (MCCFR) to learn optimal strategies in imperfect-information environments. Built iterative simulation and evaluation pipelines to analyze convergence toward Nash equilibrium strategies.

Implemented game engine logic, state management, and responsive front-end behavior for a dynamic 5-shoe Blackjack simulator with card-counting logic, probability calculations, and interactive UI.