An attack-and-defense competition on prompt injection and data exfiltration against tool-using LLM agents. I built both attack payloads and a defense that drove exfiltration to zero on the hidden set, and published a public working note walking through the approach.
The abstract-reasoning benchmark. I am working on test-time training and program-synthesis approaches on top of small open models, inside the offline compute budget the competition enforces.
Tracking cells through 3D microscopy volumes over time: detection followed by linking across frames, with a careful cross-validation protocol to avoid leakage between neighbouring frames.
Deep learning on medical scans for an RSNA challenge, a kernels-only competition with a staged baseline and an identity-verification gate before submission.
A two-player farming-simulation agent ladder. My bot mixes hand-built heuristics with replay of strong recorded games; against the top runtime agents, silver-tier strength is the honest ceiling and the target.
A bot-ladder competition for the Pokemon Trading Card Game. I built a Monte Carlo Tree Search agent tuned around the scoring quirks of the ladder and its search-depth trade-offs.