Perit builds AI and ML training data for frontier labs. We work across audio, image, video, and reinforcement learning environments, producing the high-quality, human-generated datasets that the next generation of models is trained on.
The work
You listen to recorded conversations between a person and an AI voice agent, and rate every agent turn on two things what it said, and how it sounded each with a short written justification in your own words.
What we're looking for
- A good ear for American English. You notice when a voice slips: a robotic patch mid-sentence, a shift in voice identity between turns, pacing that's a beat off, stress landing on the wrong word.
- Real listening comprehension. You can follow rapid, casual, overlapping speech without replaying it three times.
- Opinions you can defend.
Most of this job is writing one or two sentences on why a response was weak — specifically. "It was okay" is worth nothing to us.
- Your own writing. AI-generated justifications are obvious and we reject them. This is a writing job as much as a listening one.
- The nerve to call something bad. The most common failure we see is someone noticing a problem and then giving it a generous score anyway.
Not for you if you were planning to run the task through ChatGPT, or you want something to half-watch while doing other work. The bar is high, and it's checked.
Terms
- Pay: 25000/month, depending on output
- Fixed Term Contract
- Commitment: 8 hours a day
- No prior experience required — we train you
- Fully remote