John Otters

With NuzlockeAI, I directed Claude Code agents through a generate, verify, and adjudicate pipeline to develop a C++17 solver for real-time battles. It uses AND–OR search to prove whether a position is winnable against every opponent choice and RNG outcome.

I also audited the solver’s formal assumptions against a 1,500-matchup corpus. A crit-damage dominance property held for only about 4% of moves, so I replaced it with a weaker property that held and kept the original as a monitored fast path.

For my M.S. thesis, I created RootSegmenter, a human-in-the-loop root phenotyping tool combining a U-Net with a classical computer-vision pipeline. After training on twelve corrected images, it reached about 0.98 Dice; I also documented the evaluation’s ground-truth contamination risk.

My other projects include a PowerPC game-mod toolchain, a high-performance board-game solver, and responsible security research that led to a patched Zoom Windows client vulnerability. I’ve also taught AP Computer Science and robotics at Code.Ed.