Epistemic Labs
Teaching models the judgment that was never written down.
We source and build proprietary expert-decision data for AI training and evaluation.
We find it inside operating companies, clear the rights, and turn it into datasets models can learn from.
About
The bottleneck has moved from text to judgment.
Frontier models have learned from most of the public web. What they rarely see is the record of trained people making consequential decisions, and what happened next: the adjuster's call and the claim outcome, the quote and the lost deal, the technician's diagnosis and the repair.
That record lives inside established companies, unpublished and hard to reproduce. Epistemic Labs is an applied AI lab built to bring it out: we find it, secure the rights to use it, and turn it into data that models can learn from and be measured against.
What we build
Outcome-linked expert data
Real decisions paired with what actually happened, de-identified where needed and documented, for post-training and preference learning.
Evaluation sets
Cases with known results, so a model can be scored against real outcomes rather than opinions about them.
Reinforcement learning environments
Multi-step workflows reconstructed from real operations, with the tools, intermediate states, and outcome-based rewards that make them trainable.
Expert-authored data
Where real records cannot be licensed: synthetic matters written and adjudicated by practicing professionals, such as dispute cases scored by attorneys. We already do this work: our team writes expert legal evaluations for AI training programs.
Where the data comes from
From companies that never thought of themselves as data vendors.
Accounting firms, laboratories, logistics operators, inspection and support teams: years of decisions and outcomes, recorded in the systems they already run. Through our data opportunities service, owners learn what their records could be worth, keep ownership, and license only on terms they approve.
Explore data opportunitiesTeam
We build AI that makes expert decisions, and the data that keeps it honest.
We built and operate a production AI system that issues reasoned decisions from evidentiary records, along with the evaluation work around it: test sets drawn from de-identified real cases, adversarial and prompt-injection testing, citation audits, and checks that results hold steady when irrelevant details change.
Shunhe WangCo-founder, CEOBuilds AI systems that make and explain expert decisions, and writes expert legal evaluations for AI training and evaluation programs. Leads the team behind our production AI decision system and its evaluation program. Previously a litigator at Kellogg Hansen and a law clerk on the U.S. Court of Appeals for the Ninth Circuit. B.A. summa cum laude and J.D. from Yale, where he served on the Yale Law Journal.
Jason ChenCo-founder, CPOBuilds the systems behind our production AI decision system, from structured case records to settlement infrastructure. Previously a trader at Optiver, working on quantitative decision-making and high-stakes financial infrastructure. Studied mathematics and computer science at Yale.