Evalium exists because of a pattern we saw from the inside: human data programs fail when evaluation is an afterthought. Data gets produced at volume, quality gets asserted rather than measured, and models learn what the pipeline made easy — not what they actually needed. We built the company the other way around. Every engagement starts with evaluation, every dataset ships with its own eval, and the people producing the work are selected for one skill above all: knowing what expert-level looks like in their field.
Labs have the compute and the methods; progress is now rate-limited by coverage — the tasks, workflows, and expert judgment that were never on the public internet.
As models improve, the data that closes the last gaps in a capability is worth more than everything that came before it.
"High-quality data" is a claim. Rubrics, verifiers, and held-out evals are evidence. We only deal in evidence.
The team that finds a model's failure modes is the team best equipped to build the data that fixes them. Splitting the two is how the industry got its quality problem.
Evalium was founded by two operators who built their careers designing, selling, and delivering the programs that train the world's leading AI models.
Over a decade in software engineering, now working at the frontier of generative AI. At Turing, Vahid led programming-data generation for the world's leading model builders — directing 200+ engineers, shipping 160K+ premium training datapoints, and delivering programs that unlocked $10M+ in value. Earlier, he was a researcher at JetBrains, building tools to measure developer impact. His research on AI-assisted software engineering is published at ICSE, ESEC/FSE, and MSR, where he also serves as a reviewer. PhD candidate in Computer Engineering at Bilkent University.
A mathematician and data scientist who turns deep technical fluency into partnerships. At Turing, Iliya worked on post-training data and evaluation for frontier models within the AGI program, and previously spent time on generative AI at NVIDIA. He is data science faculty at UC Berkeley — head teaching assistant and course coordinator in the College of Computing, Data Science, and Society — and has published on scalable models for data science consulting. BA from UC Berkeley.
A small core team of researchers and operators, and a selective network of domain experts around it. If you'd rather build the standard than follow one, we want to hear from you.