Teamday
← People

Rayan Krishnan

CEO at Vals AI

agentsevaluationenterprisevals-ai

About Rayan Krishnan

Rayan Krishnan is the CEO of Vals AI, a company that builds independent benchmarks and custom evaluations for AI models and agents. Vals' public benchmarks, including the Vals Index and the Finance Agent Benchmark, aim to measure economically valuable work in domains such as finance, law, coding and healthcare, and are cited by model labs in release announcements. Vals raised a $40 million Series A led by Andreessen Horowitz in August 2026.

Vals also sells Vals Smith, which turns a company's merged GitHub pull requests into a private coding benchmark. Krishnan describes it as the product version of internal tooling Vals built to rein in its own token spend.

Career Highlights

  • CEO of Vals AI, maker of the Vals Index and Finance Agent Benchmark
  • Vals raised a $40 million Series A, led by Andreessen Horowitz, in August 2026
  • Vals has run more than six trillion tokens of model evaluation

Notable Positions

On Cost Per Task

List prices per token mislead. On a Vals tax benchmark, a model with lower token prices spent more tokens, cost more per task and scored 3% lower than the frontier model, so buying the cheaper model meant paying more for worse results.

On the Jagged Frontier

Rankings fragment as tasks get more specific. On Vals' Excel modeling benchmark, the LBO subset has a lower frontier and a different leading model than the overall score, so firms should benchmark the sub-tasks they actually do.

On Building Evals from Work History

Krishnan argues a company's history of completed work, such as merged pull requests, client email replies, SharePoint and data warehouse dumps, holds its standard of excellence. Evals built from it let a company right-size usage, negotiate pricing, route between models and fine-tune.

Key Quotes

  • "agents solve the test not the problem" (on reward hacking in evals)
  • "For that month, we spent $1.5 million worth of tokens." (on Vals' token-maxing experiment)
  • "I think we're spending 90% less, but still able to get the same output." (on cutting spend with evals)