Antaripa Saha. Engineer at Quotient AI working on agent systems, with a focus on making AI evaluation a first-class part of harness engineering rather than a post-deployment quality check. Her published work integrates the Hamel Husain and Shreya Shankar AI Evals course frameworks with practical experience building and debugging agent harnesses in production.
Operating themes
- Evals as feedback, not audit Evaluation systems belong inside the improvement loop, not at the end of it.
- Harness-first debugging Agent failures trace to harness decisions; traces and eval cases localize which harness element failed.
- Criteria drift as a feature What counts as good agent output evolves as teams observe more behavior; evaluation workflows should expect and accommodate revision.