Axiomatic
Evaluation and observability for production language models.
The company
- Founded
- 2022
- First check
- Pre-seed
- Headquarters
- San Francisco, CA
- Sector
- AI
- Partnered since
- 2022
Every team shipping a language model into production eventually hits the same wall: they cannot tell whether the new version is better than the old one. Offline benchmarks say yes. Users say no. Axiomatic closes that gap.
The platform captures real production traffic, builds evaluation sets from the failures that actually occurred, and scores every candidate release against them before it reaches a customer. Teams get a regression suite for behaviour that was previously impossible to test.
Axiomatic now sits in the release path of engineering organisations at four public companies and more than two hundred startups, grading millions of model responses each week.
Why we invested
We had spent a year meeting companies building model evaluation and none of them started from production traffic. Amara and Ravi had run this problem inside a large search team and knew that the only evaluation data worth trusting is the data your own users generated. We wrote the first check three weeks after meeting them.
-
Tessellate
Synthetic data generation for computer vision teams.
-
Halyard
Deployment infrastructure for regulated engineering teams.
-
Portstack
Ephemeral preview environments for every pull request.