codeinterp

Interpretability · Control · Continuous learning

Make the machine legible.

CodeInterp is building infrastructure for organizations that run their own language models: to see what a model has learned, change it with precision, and let it keep learning, with a record of every change.

What we work on

Understanding where knowledge lives inside a model, and what it takes to change it.

  • Interpretability

    Reading what a model knows: where its knowledge and behavior live, described in terms a person can inspect.

  • Precise control

    Changing a model one thing at a time, locally, predictably and reversibly, rather than retraining the whole.

  • Continuous learning

    Models that keep learning after deployment, with what they learn kept as records that can be inspected, consolidated and rolled back.

Why it matters

Models are becoming infrastructure. They should be as inspectable as infrastructure.

More organizations now run and adapt their own open-weight language models, and those models increasingly inform consequential decisions. The people responsible for them need to know what a model has learned, correct it when it is wrong, keep it current, and be able to show their work.

The tools to do that precisely are still missing. That is what we are building.

Legible.Local.Reversible.

Founder

Srinidhi Ramamurthy

Founder & CEO

Srinidhi works on interpretability, spectral and random-matrix methods, and the internals of large language models, drawing on a decade of building applied-AI products and teams.

LinkedIn

Contact

Get in touch.

We would like to hear from teams that run their own models, and from researchers working on related problems.

Company

CodeInterp Inc.
710 Lakeway Drive
Suite 200 PMB#17606554
Sunnyvale, CA 94085