Data capture for AI training. Real work, recorded the way a model needs it.
Some things a model needs to learn have never been written down. They happen on a job site, in a kitchen, at a bedside, inside a spreadsheet. eeko puts a camera, a screen recorder, or a teleoperation rig on the people who do that work, records it to a written protocol with consent from everyone involved, reviews every episode, and delivers it licensed with a signed record on each file.
What is data capture for AI training?
Data capture for AI training is the deliberate recording of real activity so a model can learn from it: a plumber's hands and eyes as they sweat a joint, an accountant's clicks as they close the month, an operator's commands as a robot arm picks a part, a nurse's voice as they hand over a shift. The recording is paired with labels that mark what was happening and when, and with consent from everyone who appears in it.
It differs from scraping or buying stock footage in every way that matters. The task is chosen in advance to match the capability being trained. The people are the ones who actually do the work, on their normal shift, not performers on a set. Variation comes from real conditions and from a protocol that asks for it. And because we arrange the recording, we can prove where every frame came from and that the person in it agreed.
eeko runs capture because the agents and robotics systems we build keep needing it. Egocentric footage teaches embodied models how a task unfolds. Screen recordings teach computer-use agents how work moves between applications. Teleoperation episodes are the raw material of robotics training data. None of it exists on the open web, and none of it should be taken without asking. The recordings come from our paid capture programs and from the specialists in the eeko human data network.
Six capture methods, each run by the people who do the work.
Every method follows a written protocol, is checked against it before volume, and is reviewed before it reaches you.
- Worker's-eye video, labeled by task and phase
- Hands and tools annotated, release on file
- From the capture programs
- Video plus a click, keystroke, and window log
- Task edges marked, personal fields masked
- See the screen-work program
- Joint and end-effector trajectories with commands
- Wrist and scene cameras, pass or fail per episode
- Detail under teleoperation data
- Audio with verbatim transcripts
- Speaker, language, accent, and environment tags
- More under speech and audio datasets
- Demonstrations, preference pairs, rubric scores
- Evaluation sets, second-expert reviewed
- Join via apply as an expert
- Depth maps, point clouds, meshes
- Force, tactile, and telemetry aligned to video
- See 3D and CAD datasets
Screen recording training data for computer-use agents.
An agent that has to work inside an ERP, a ticketing queue, or a design tool learns fastest from people who already do. We record them, carefully.
What a screen session gives the model
- The route — which application opens first, what gets copied where, which report is checked before the entry is posted.
- The events — every click, keystroke, scroll, and window switch, timestamped against the frames.
- The boundaries — where one task ends and the next begins, marked by the worker and confirmed in review.
- The exceptions — the moment a value does not match and a person goes looking, which is the behavior an agent most needs to copy.
Programs run for bookkeepers and office workers, customer support agents, and designers and editors. The result feeds the same computer-use work we do under enterprise workflows.
Twelve electricians, four weeks, one protocol.
An illustrative run of an egocentric capture project for a model learning residential electrical work. Every episode on the record.
How a capture project runs, start to finish.
Four steps, fixed scope, fixed price, with a pilot before volume. A data lead replies to every request inside one business day.
Real workers, paid for real hours.
Sixteen capture programs cover the jobs a model most needs to watch. Each pays per accepted hour, supplies the kit, and handles consent.
Three ways in
- Join a program — workers in warehouses, factories, farms, stores, salons, veterinary clinics, labs, and surgical and dental practices apply through get paid to train AI.
- Open your home — homeowners in Minneapolis, Chicago, and Miami get a free home service in exchange for letting the tradesperson record the visit.
- Bring your company — staff enroll together, each hour is paid, and you can take an exclusive licence back on the result. Already sitting on years of recordings? Sell your company data instead.
Data capture questions.
Who does the recording in a data capture project?
Working people, not actors. Camera and screen footage comes from our capture programs: electricians, cooks, nurses, mechanics, warehouse staff, bookkeepers, and others, paid by the accepted hour, recording a shift they were working anyway. Teleoperation episodes come from trained operators. Expert demonstrations are written by specialists from the eeko expert network, each verified in their field before they see a task.
Can we control exactly what gets recorded?
Yes. Each project opens with a written capture protocol: which tasks, which viewpoint, which variations to cover, what the frame or screen must include, and what counts as a complete episode. Our team grades the opening sessions against that protocol before volume recording starts, and you approve sample episodes before the project scales. Changes to the protocol are versioned so later episodes can be told apart from earlier ones.
What equipment is used for egocentric and screen capture?
For first-person video we ship a compact head- or chest-mounted camera, and add a depth sensor when the protocol needs geometry. Screen work uses our recorder, which captures only the applications the worker has chosen and logs clicks, keystrokes, and window changes alongside the video. Teleoperation runs on your robot hardware or on ours, decided during scoping. Workers never pay for kit.
How are customers, patients, and bystanders handled in recordings?
Consent comes before capture, from the worker, the employer, and anyone likely to appear, including customers on site. Before review, faces and any screens or paperwork unrelated to the task are blurred, personal fields in screen video are masked, and location metadata is removed at intake. Clinical and care settings are recorded only where the facility and every person involved has agreed in writing. Recordings showing identifiable people are offered only with that consent on file.
Can eeko run a capture project inside our own company?
Yes. Your staff enroll together. eeko manages consent for each participant and pays every accepted hour, and you can take an exclusive licence on the result or, with your permission, let other buyers license it for a share of the fee. If your company already holds years of recordings, call logs, or records, see our sell your data page instead; that route needs no new recording at all.
What does a finished data capture project deliver?
Accepted episodes with the labels in your spec, such as task and phase boundaries, hand and tool annotations, or click and keystroke events, together with technical metadata, the consent references, and a signed provenance record for each file. It ships as WebDataset shards, Parquet, or JSON Lines pointing at the media, with a Croissant dataset card describing the set and a signed receipt listing every file by hash.
How much does a data capture project cost?
It depends on the method, the number of accepted hours, how specialised the people are, the labeling depth, and the licence. Worker pay runs from twelve to thirty dollars per accepted hour by program and is passed through transparently, with protocol design, review, masking, labeling, and delivery priced on top. Every project is quoted fixed scope and fixed price against the written protocol, and a small pilot batch comes before volume so you can check the footage in your own pipeline first.
Explore the data practice.
Show us the task. We will go and record it.
Five minutes to send a request. A data lead replies inside one business day with a first draft of the protocol, a sourcing plan, and a fixed price for a pilot batch. No deck. A protocol and a date.
Experienced within
Markets served.
As an enterprise AI agency, eeko systems delivers production AI systems remote-first across the United States and internationally — including these markets: