Training & MLOps

Trains models across many devices without the data ever leaving them.

Federated learning *is* the direction's core question made into a product — what it costs to send your data somewhere else — and they are the open-source default for it, sitting on four separate documentation roots with no single door in.

A feature store that keeps training data and production data computed the same way.

The feature store that predates the current wave and is now re-aiming itself at agents — a useful control on whether "AI infrastructure" is a new stack or the old data stack under new lighting.

Turns your live LLM traffic into fine-tuned models that are cheaper and better.

The "your production traffic is your training set" bet, priced at $50 a month — the clearest small-team attempt to make fine-tuned task models cheaper to run than the frontier model you are calling today.

Runs ML training pipelines on any hardware and records how each model was made.

A ten-year-old, ~29-person ML platform that runs on your own hardware and has never leaned on funding news — the closest thing on the roster to a real £500/mo buyer rather than a logo.