What the work actually involves

You are the second pass on someone else's labels. Clips arrive already annotated — an action label like "person transfers object from left hand to right hand," plus keypoint or hand-pose data — and your job is to decide whether the annotation matches what is visible on screen, whether it matches the guideline's definition of that action, and whether it is consistent with how similar clips were handled. When it isn't, you fix it and write a short note explaining the discrepancy. Expect frame-by-frame scrubbing, disagreements over boundary cases (when exactly does a reach become a grasp?), and a guideline document you will come to know better than the people who wrote it.

The volume is real. This is large-scale dataset QA, not curation of a handful of interesting examples. Throughput matters, but so does not drifting: the failure mode reviewers are watched for is loosening their own standard by hour four. Written feedback is part of the deliverable, not an extra — clients use reviewer notes to patch ambiguous guidelines, so "wrong label" is a worse contribution than "label assumes single-handed grasp; guideline §4 doesn't cover bimanual handoff, recommend explicit case."

What the screen looks for

micro1 runs an AI-led interview. For this role it probes concrete tool exposure (which annotation platforms, what task types, how many hours), English reading and writing at a level sufficient to parse and critique guideline prose, and — most heavily — your judgment on ambiguous cases. Expect to be asked what you do when a label is defensible but not what you'd have chosen, or when the guideline contradicts itself. Answers that resolve everything by personal preference score badly; so do answers that escalate everything. No AI or machine-learning background is required or expected.

Logistics

  • Remote, contractor engagement, asynchronous within project deadlines.
  • Hours are typically flexible and project-scoped rather than fixed shifts; sustained availability across a batch matters more than specific times of day.
  • Requires a computer capable of smooth video scrubbing and a stable connection — buffering makes frame-level review impossible.
  • Pay band observed at $15–20/hr for this listing; rates vary by project and are set by the client, not guaranteed.