We build training environments, expert tasks and physical-world data with real businesses and real experts. The kind of work that never made it onto the internet.
Why we exist
AI has read the internet. What it hasn't seen is how a business actually runs, what an expert knows, and how skilled hands do the job. That knowledge lives inside companies and people, so that's where we go.
Inside real workplaces
15 industries so far, and whatever a lab needs next.
What we build
01 Training environments & business data
Real companies,
rebuilt for AI agents.
We work with operating businesses and rebuild their software, documents and daily tasks as safe copies. Agents do the work inside them and get scored on every step. We also clean and license the business data itself.
Environments
- Copies of real companies' tools, with fake but realistic data
- Hundreds of everyday tasks, each with a pass-or-fail check
- Local apps and languages, not just English
Cleaned business data
- Real records from businesses that never put them online
- Names, numbers and private details removed or replaced
- Licensed with clear rights, ready to train on
02 Expert tasks & expert network
Hard problems, written and checked by people who do the work.
Accountants, lawyers, engineers, doctors, language examiners and olympiad medallists write the questions today's best models get wrong, with answers verified before delivery. Need the experts themselves? We find, vet and manage them for you.
We deliver tasks
Ready-made sets of expert questions, step-by-step answers and grading rubrics, built to your spec.
We find your experts
Vetted professionals for your own projects: evaluations, feedback, red-teaming. We handle hiring, quality and pay.
Domains
03 Physical-world data
Skilled hands,
captured in full.
Plain head-camera video is cheap now. Robots need more: what the worker sees, what the hands do and how hard they press, all in sync and labelled step by step.
- < 30 ms
- glove-to-video sync, so every touch lines up
- 100 Hz+
- fingertip force, from 0.01 to 5 N
- 55% → 88%
- robot task success once first-person human data is added (EgoMimic, 2024)
- 10×
- more demonstrations per hour than robot teleoperation (EgoMimic, 2024)
Why DataCentric
Data from the parts of the economy
that never went online.
Access, not scraping
We work inside real businesses under signed data-rights deals. None of this data was ever public.
Real professionals
Vetted professionals write and check the work, not anonymous click-workers.
Beyond English
Tasks built natively in local languages, from Kazakh, Uzbek and Arabic to Turkish and many more, wherever the work happens.
Clean by default
Private details removed, every item reviewed, and a full record of where each file came from.
How we work
From shop floor to training run.
Raw footage never leaves the country it was filmed in. Only cleaned, consented, labelled data goes out.
- 01
You send the task
A task list, the robot or model it's for, and how many hours you need.
- 02
We get consent
A signed form from every worker, in their own language, and a data-rights deal with every business.
- 03
We capture on site
Trained crews film scripted tasks in real workplaces, including slips, drops and second tries.
- 04
We clean, label and check
Faces, badges and screens are blurred. Every step is labelled and reviewed. Weak footage is thrown out, and you only pay for hours that pass.
- 05
You train
LeRobot, MCAP, HDF5 or WebDataset, with a datasheet and a full custody log. Free sample first, then a pilot, then a licence.
We never film
Minors, customers who haven't agreed, or private homes without the resident's signature.
Every file has a trail
Who filmed it, where, the consent ID and every processing step.
Exclusive if you need it
Non-exclusive by default. Exclusive sets for a fixed period on request.
Get started
Tell us what your
model needs to learn.
For AI labs, robotics teams and data buyers. We'll reply with a plan and a free sample.