Auxerta Labs.
A small research lab building foundation models from the ground up.
The story so far
Auxerta started in 2025. We spent the first year building a domain annotation service — cleaning and labeling training data for clients in finance, law, and a few other regulated fields.
By early 2026 we'd shipped about a million labeled samples and learned where the real limits were. Better data only takes you so far when the underlying training objective is wrong.
Today Auxerta is a foundation model lab. We're working on pretraining methods that train representations independently of how the model produces tokens — the technical case is in the research note.
Founders
Philip Abao
Co-Founder
Engineering and operations. Writes most of the training infrastructure code; runs the company side — sales, hiring, finance.
Soraya Johnson
Co-Founder
Patent and scientific research. Handles partnerships with subject-matter experts and the regulatory side of the work.
Locations
USA
Headquarters. Distributed across the US with founding operations and core engineering.
Japan
Yokohama. Expanding operations across the Asia-Pacific region.
How we work
Small team, no handoff
The same people design the architecture, write the training loop, and run the evals.
Long runs
A pretraining run takes weeks. Most of that time, there's nothing publishable. That's how the work goes.
Papers, not press releases
We share results when there's a paper. Anything earlier is marketing.
Want to work with us?
We're always looking for people and organizations who take data seriously.