We started with the hardest audio, and the people who handle it.

Speech models were failing quietly. A transcript read well, a digit in an account number was wrong, and nobody noticed until three systems downstream. Clean read-speech corpora could not catch it, and crowd raters had no idea what a good call sounded like. So we built a bench of people who did — transcribers and aligners who work to a written guide and pass a test before they touch a paid file — and put it on a platform we run ourselves, Foundry.

That bench has now delivered more than four thousand hours of speech across nine locales, at fifty hours a working day, with every batch shipping the numbers it was produced at. The same machinery — recruiting to a locale, consent captured on the file, a calibration gate, QA sampling, a quality report — does not care whether the file is an audio clip, a photograph, a paragraph or a first-person video of someone folding a towel.

So the next step is the obvious one. Video, image and text collection to spec, annotation of any data type on the same bench, and a Physical Intelligence line — egocentric video, depth, teleoperation and handheld-gripper demonstrations — for teams training robots. We say plainly which of those has shipped and which is being scoped, because a number nobody measured is worth nothing to the person buying it.

Perit is operated by Zyno AI Inc and backed by Y Combinator. No frontier lab owns a piece of it, and none sits on the cap table — your held-out data does not end up next to a competitor's.