Raizen developer profile — @nadia_okonkwo
Staff ML Engineer, Evaluation at Anthropic
ML engineer who treats evaluation as a product. I build grader harnesses, slice reports, and agent traces that a hiring team can defend — not a vibe score.
Location: San Francisco, CA
Ranked #1 on the Raizen verified-developer leaderboard.
Website: https://github.com/nadiaokonkwo
2023-03-01 – Present
Owns the human+model eval loop for assistant quality. Cut disagreement on fairness slices by 31% and shipped a trace reviewer used in weekly launch reviews.
2021-06-01 – 2023-02-28
Risk ranking and calibration for dispute models. Production PyTorch + Spark.
M.S., Machine Learning
2019-08-01 – 2021-05-15
Open eval kit that scores graders on demographic and domain slices, with bootstrap CIs and a failure gallery. Used internally to block a launch that looked fine on mean accuracy.
Parses tool-calling traces into a graded rubric: goal completion, unnecessary steps, and unsafe actions. Python + PyTorch pairwise head.
dbt + Delta Lake package that proves which rows entered an eval split. Catches leakage before it hits production.