# Clean Evals

> Evals and diagnostics for AI agents.

Test cases are based on user requests, agent traces, and known failures, with additional cases for difficult tasks and new capabilities. Automated grading is calibrated against product experts’ assessments and validated on a separate set of cases. Reports show which tasks work, examples of failures, and priorities for improvement. Cases and scoring are updated as the product and its usage change.

Co-founders: Max Petriev and Eli Polterovich.

## Services

- Audit: 1 week, $1,500. An eval roadmap with priority workflows and test cases, success criteria, missing data or context, and an implementation plan covering setup, test coverage, and acceptance criteria.
- Implementation: 8 weeks, $20,000. Deliver an eval harness and runner, calibrated graders, hill-climb and regression evals, baseline reports, and documentation.
- Continuous eval improvement: $5,000 per month. A dedicated eval engineer, updated eval cases and success criteria based on production traces, and grader recalibration when needed. Includes weekly 30-minute check-ins and monthly reports on performance, coverage gaps, and completed work.

Audit needs access to agent traces and about two hours with a product expert. Implementation needs 2–3 hours a week from a product expert. Code, data, and evals stay in the client's infrastructure.

Clean Evals handles evals and diagnostics. The client provides domain expertise and improves its agent. Agent traces, datasets, annotations, and results stay in client infrastructure or approved accounts. Project deliverables are placed in client repositories and workspaces as they are built.

## Contact

- [Book a working session](https://cal.com/max-petriev-rvyvdf/discuss-your-project)

Website: https://cleanevals.com/