Rayan Krishnan
Conference affiliation: Co-Founder & CEO · Vals AI · 2026
Rayan Krishnan is the co-founder and CEO of Vals AI, which evaluates AI systems on professional work. With co-founder Langston Nashold, he has built a company around independent measurement: working with practitioners to test models on tasks that lawyers, financial analysts, software engineers and other specialists actually need to complete. His work asks how benchmark results translate into a purchasing decision—especially when a model’s low token price conceals a higher cost per satisfactory result.
On this page
Research before Vals
Rayan studied computer science at Stanford, with minors in mathematics and philosophy. His earlier experience included research and work at the Stanford Artificial Intelligence Laboratory, Palantir and Valar Labs, alongside a software engineering internship at Microsoft. He also serves as a Summit STEM Fellowship mentor.
His collaboration with Langston predates Vals. For a Stanford machine-learning project, they compared a seasonal statistical model with an LSTM neural network for forecasting CPU demand using Azure data. Their results depended on the forecasting horizon: the statistical model performed better over the longer horizon, while the neural network did better over the shorter one and depended less on seasonal assumptions.
Rayan subsequently contributed equally with Mark Endo to CXR-RePaiR, a 2021 radiology-report system developed with Viswesh Krishna, Andrew Ng and Pranav Rajpurkar. The team approached report generation through retrieval, using learned image and text representations to match a chest X-ray with existing reports or report sentences. They evaluated clinical findings as well as similarity of wording: a report could resemble the reference text while describing the wrong diagnosis. Their compressed retrieval system achieved comparable performance while producing reports faster than the generative model used for comparison.
Tests built from professional work
At Vals, Rayan and Langston made evaluation the business. The team works with practitioners to construct representative tasks and scoring methods, while keeping test sets private to reduce opportunities for training on the answers. Finance Agent, for example, tests a recurring analyst workflow: researching public companies through their SEC filings. Vals also released Valkyrie, open-source infrastructure for running reproducible agent evaluations on hosted infrastructure or a local machine.
Rayan connects this work to several distinctions that matter when organizations choose models:
- Cost per completed task. A cheap token rate can conceal an expensive workflow if a model needs more tokens to finish. In his presentation on model cost and quality, Rayan uses tax work and leveraged-buyout spreadsheets to explain why even a broad finance or Excel ranking can mislead a buyer. The comparison needs to reflect the particular work being assigned and the total cost of producing an acceptable result.
- Useful drafts versus finished work. Rayan distinguishes average rubric credit from the proportion of answers satisfying every requirement. A professional editing a draft can benefit from partial success; a firm sending output directly to a client needs a stricter measure. He also describes a scoring failure in which an agent offers several possible answers to increase its chances of including the expected one. The score can reward an output that still leaves the user to resolve the uncertainty.
- A team’s own history as a test set. Vals Smith turns suitable merged pull requests into evaluation tasks with hidden tests. An agent receives the original issue and the repository before the fix. The tests must fail on that starting state and pass on the accepted solution, giving the evaluation a check that it can distinguish the original defect from completed work. This makes model comparisons specific to a team’s code and standards, including whether a change preserves existing behavior.
In his August 2026 funding announcement, Rayan argued that model development was advancing faster than the supply of meaningful tests: benchmarks saturate, questions enter training data, and developers often evaluate their own products. Vals announced a $40 million Series A and the general availability of Vals Smith. His proposed answer is continuing independent evaluation—developing harder, relevant tests as models improve, while helping organizations identify which improvements matter to their work.
1 conference talk
Loading more talks and highlights…
