I spent five years as an applied ML engineer at Deepgram building production speech and LLM evaluation infrastructure, which taught me to start from deployment failure and ask what the benchmark missed. Now I work independently on measurement validity for alignment-relevant evals: whether the benchmarks we use to catch sycophancy, dishonesty, and spec-violations actually measure what they claim, and whether we'd know if they didn't. My MSCS capstone (defended July 2026) is a run-level reproducibility audit of published sycophancy benchmarks, with an arXiv preprint in progress. I ship concrete evals and public, reproducible harnesses rather than position papers.

Based in Chiang Mai.