Clavia is studying how traces of work can become training data and learning signals for AI agents.
We build evaluations, environments, and post-training tools for domains where frontier model performance remains uneven. We work with AI labs and companies exploring these problems.
Writing
Calibrating a judge for legal work
How we built a legal evaluation environment and calibrated its LLM judge against lawyer preferences.
Our work
Tardigrade
A framework to build your agent harness as components over an immutable event log. Each component is a typed state machine over the event log, using EffectTS.
About us
Get in touch at contact@clavia.ai.