AI Safety Research Engineer
Li Bearden
I build measurement infrastructure for AI safety.
LLM evaluation methodology and sycophancy reproducibility, grounded in 5 years of applied ML at Deepgram.
The work
Research
Run-level reproducibility of published sycophancy benchmarks — whether single-run numbers survive being re-run. PARROT reproduces cleanly as the positive control (Spearman 0.95–0.99); SycEval is the open falsification test.
Consulting
The LLM Evaluation Audit — applied evaluation engagements for teams shipping models to production. Eval design, reproducibility checks, and the infrastructure to run them.
Bench ↗ fieldbuilding
A deliberate-practice tool for engineers ramping toward frontier safety research — a daily Prep → Ingest → Produce → Assess → Connect pipeline that surfaces one recommended next action and grades timed attempts against a rubric.
Writing
Working notes and essays on evaluation methodology, eval validity, and the institutional dynamics of frontier AI safety review.