We build
RL environments.
Production systems. Expert curated verifiers.
Research
We acquire proprietary environments and measure where and why frontier models lag acceptable expert performance benchmarks.
COBOLBench: Evaluating Coding Agents on Enterprise COBOL
A benchmark from a real banking system measuring whether AI coding agents can work in production COBOL.
FigmaBench: Evaluating Design-to-Code Agents on Production Figma
A benchmark measuring AI agents on production Figma-to-code conversion through API interaction, design hierarchy extraction, and iterative deployment.
Metaphi Simhub
Interactive simulation environments that move agent evaluation beyond static, non-interactive benchmarks.
From Static Benchmarks to Dynamic Worlds
How principles from autonomous vehicle simulation can transform code generation agent training and evaluation.
Get in touch
Tell us about your use case.