Building AI Evals
A practical guide to designing and using evidence in AI evaluation.
A forthcoming book on designing, grading, and validating AI evaluations, and deciding what their results can support.
About the book
Building AI Evals is a forthcoming book by Saeideh Bakhshi about designing evaluations that produce useful evidence for AI product decisions. It begins with three linked questions: what are we evaluating, why are we evaluating it, and how will we know?
The book brings together ideas from software testing, machine learning evaluation, measurement, experimentation, human judgment, and user research. It follows the choices involved in building an evaluation through to interpreting its results and deciding what the evidence supports.
Who it’s for
For practitioners who design or assess the evidence behind AI product decisions: researchers, software and AI engineers, product managers, data scientists, designers, and domain experts. The core reasoning is accessible without code, with technical terms explained where they become useful.
About the author
Saeideh Bakhshi, PhD, is a quantitative UX research leader and computational social scientist. She currently leads user research for model behavior at OpenAI.
What it covers
- Frame the evaluation. Choose the goal, define the system or interaction being evaluated, and specify what good performance means.
- Build the evidence. Design tasks and cases, plan sampling and coverage, and decide what observations each requirement needs.
- Grade and validate. Write rubrics, combine automated checks, model graders, and human judgment, and examine whether the evaluation measures what it should.
- Use the results. Analyze uncertainty and failures, explain what conclusions are justified, and maintain the evaluation as systems and their use change.
Examples include scheduling and customer-support agents, connecting evaluation design to actions, context, and outcomes beyond a final response.
The book is in development. Subscribe above for updates, or read the author’s AI evaluation essays.