How it works

How we turn skilled work into training data.

There’s no secret to collecting egocentric data — the difference is in how carefully each step is executed. Here is our entire process, in the open.

  1. Step 1

    Spec together.

    Every program starts from your model’s needs, not our template. In a scoping call we define the task list, environments, camera setup, demonstration protocol and annotation taxonomy — then write it into a capture spec you sign off before a single minute is recorded.

  2. Step 2

    Professionals, not crowds.

    We recruit from established service companies: professional cleaners, kitchen staff, maintenance technicians — people at the top of a craft that is far harder than it looks. Each operator is trained on the capture protocol: keep hands in frame, work at natural speed, cover the full task from first touch to final check, and treat mistakes and corrections as first-class data, because your model learns most from recovery, not perfection.

    [Media] Operator prep · 30–60 s

    A professional cleaner gets ready: head-camera fitted, protocol checklist, hands calibrated into frame.

  3. Step 3

    Capture in real environments.

    Recording happens in genuine, lived-in European homes and workplaces — not studio mock-ups. Head-mounted cameras capture high-resolution, stabilized first-person video optimized for model training: consistent framing, controlled lighting where needed, full-task continuity. Every session runs on a documented consent chain: the worker’s, the household’s, and GDPR-compliant handling of anything the camera sees.

    Outside view

    The cleaner at work, third person.

    POV

    The same moment, first person.

    [Media] Split image / video — the same instant from both viewpoints.
  4. Step 4

    Quality gate.

    Raw footage is screened before annotation: framing, blur, hand visibility, task completeness, privacy compliance. What doesn’t meet spec doesn’t reach you — and doesn’t get billed.

    • Framing
    • Blur
    • Hand visibility
    • Task completeness
    • Privacy
  5. Step 5

    Annotation to training-ready.

    Our in-house team segments every session into temporal action units, labels objects and interactions, and attaches task metadata — in your taxonomy and format (or ours, if you’re starting from scratch). You receive data your pipeline can ingest on day one.

    [Media] Annotation tool screen recording

    Timeline segmentation and labels visible as the annotator works.

  6. Step 6

    Scale on demand.

    Once the pilot validates the spec, we scale: more operators, more environments, more hours — to the thousands of hours deep models need — with the same QA gate on every batch and a feedback loop that tightens the spec as your training runs reveal what the model actually needs.

No black box, no surprises — just a disciplined pipeline you can audit at every step.

Book a scoping call See what lands in your pipeline