Hi, I’m Kyle. I work as a machine learning engineer at Melange building automated patent search. Before that I studied math at UT Austin and Williams College. I’m into mechanistic interpretability, prediction markets, and NBA analytics. I’m from Austin, Texas.
LLMs often commit to an answer before chain-of-thought, then rationalize it. We can detect these answers with linear probes and switch them via steering.
Improving the uncertainty calibration of LLMs by sampling over semantically equivalent perturbations of the same prompt. [AAAI 2025]
Massey-style team ratings, player plus-minus ratings, game predictions, Monte Carlo season simulations, and player career projections.
Reach out! My email is
{firstname}{lastname}2000 at gmail dot com.