A Gymnasium-style environment where a DQN agent learns to climb procedurally generated walls. Action masking keeps the agent from wasting moves on unreachable holds, curriculum learning ramps difficulty as it improves, and every generated wall is BFS-verified to be solvable before the agent ever sees it — a Pygame renderer replays the climbs. Reached roughly a 95% success rate after iterating on reward shaping.

Research extending classical Kalman / EKF / UKF observer baselines on bioprocess models, most recently with a back-and-forth EKF for microbial growth state estimation using analytical Jacobians and forward-backward smoothing. Also built the surrounding OD data pipeline, from raw plate reads to fitted kinetic parameters, and explored where RL can adapt filter behavior under weak identifiability.

An AI@MIT project training a reinforcement learning agent to design antibodies, using ESM-IF structure-conditioned protein representations to guide the search over sequence space toward candidates with favorable predicted structural properties.

MIT UROP work on combinatorial models for Grothendieck polynomials, computing them for permutations in S₄ via isobaric divided difference operators and, more recently, investigating "ghost" Kohnert moves. Computational experiments run in SageMath, with formal write-ups in LaTeX.