Data built to a spec you write.
You know the capability you're missing better than any vendor does. Our job is to turn that into a specification, find the people who can satisfy it, and prove that they did.
Start small, prove it, then scale
Nobody should commit to standing capacity before seeing a graded batch. Every engagement starts as a pilot that is designed to be falsifiable.
Scoping session
A working session with your researchers, not a sales call. We leave with a capability target and a draft definition of done.
Spec and graders
We write the task specs and the graders, and send them to you for review before anything is commissioned. You approve what "correct" means.
Pilot batch
A small graded batch delivered in your format, with the full manifest and the rejected items included so you can see the bar.
Standing capacity
If the pilot holds up, we scale the network on that spec and deliver on a cadence you set, with quality reporting each cycle.
The commercial parts, said plainly
These are our defaults. All of them are negotiable, and we would rather argue about them now than after a contract exists.
- Ownership
- Work commissioned for you is yours. Contributors assign rights before payment, and you receive a documented chain of title with delivery.
- Exclusivity
- Exclusive by default. We do not resell commissioned work, and it does not appear in any general corpus. Where you want a shared dataset at a lower price, we'll say so explicitly in the contract.
- Pricing
- Per verified item, priced by domain and difficulty. You are not billed for rejected work — our review costs are ours, which keeps our incentives pointed at the bar rather than at volume.
- Acceptance
- Delivery is against the criteria agreed at Gate 01. Items that fail your acceptance check are re-run at our cost.
- Confidentiality
- Mutual NDA before scoping. Your identity is not disclosed to the expert network unless you want it to be, and tasks are constructed so that it doesn't need to be.
- Residency
- Managed workloads run in a region you nominate. Where data cannot leave your estate, the platform deploys into your own cloud instead.
What teams ask us first
Usually in this order.
How do we know the experts are real?
Credentials are verified before anyone receives a brief — employment history, professional registration where the domain has one, and a paid trial task graded against the same bar as production work. We report the credential class on every delivered record, so you can audit the composition of a batch rather than trusting a claim about it.
Can you work in a domain you haven't listed?
Often, but we'll tell you honestly whether we're starting from a standing bench or from a recruiting problem, and the timeline reflects the difference. We would rather decline a domain than run a thin network in it and pretend otherwise.
What happens when the grader and the expert disagree?
It escalates to a senior adjudicator, and the item is flagged in the manifest. In practice a persistent pattern of disagreement usually means the specification is wrong rather than the expert — so those items feed back into Gate 01 and the spec gets revised.
Will this data leak into a competitor's model?
Not from us. Commissioned work is exclusive by default and is not resold, and we track exposure events per item in the registry. If you want a contractual remedy attached to that, we'll write one.
Can we use our own experts?
Yes — that's the deployed configuration of the platform. You bring practitioners you already employ, we provide the workbench, grader runtime, review system and support, and it all runs inside your cloud. More on deployment.
How fast is a pilot?
It depends almost entirely on how quickly the specification converges, not on how fast we can recruit. Teams that arrive with a sharp definition of done move fast. We'll give you a real timeline after the scoping session rather than a number now that we'd have to walk back.
Tell us what your model can't do yet
Send the failure mode. We'll tell you whether it's a data problem, and say so if it isn't.