The operational record, permissioned.
A de-identified collection with timestamps and cross-references intact where the license allows: the thread, the ticket, the document and the CRM record that belong to the same piece of work.
For AI teams
Tell us the capability gap.
We spec and source permissioned real-world data for teams that train and evaluate models on how work actually gets done. Companies are approached only with their written consent.
The path
Request, spec, source, pilot sample, scale.
A buyer engagement starts with the gap, not with a folder of files. Companies are approached only with their written consent.
- 01
Request
The capability you are training or evaluating, where current data falls short, and what good looks like.
- 02
Spec
Domains, systems, years, outcome signals, exclusions and license scope, written so a sourcing effort can be judged.
- 03
Source
Permissioned supply, approached with the company's written consent. Preparation of a company's data is a separate engagement.
- 04
Pilot sample
A small, documented sample checked against your acceptance criteria, once a company has agreed in writing.
- 05
Scale
If the sample holds, the same standard extends: provenance, exclusions, and checksums.
Deliverables
From records to tasks, environments and evals.
These are different objects. We say which one a request is asking for, and we build it only on licensed, de-identified or synthetic seed data.
Units of work, not a pile of text.
Context, the actions taken, and the outcome. Useful when a model needs to practice a job.
World, state, tasks, tools, verifiers.
A written specification a team can build against: the world, the state, the tasks, the tools and the verifiers.
Held-out judgment, not fluency.
Tasks and rubrics aimed at exceptions, handoffs and incomplete information, kept separate from material used to shape training.
A person in the loop, on purpose.
When the domain needs it, we design the task, the rubric and the review. Seed material is licensed, de-identified or synthetic.
A package you can audit.
Manifest, data dictionary, provenance record and checksums. The same standard as a preparation engagement, under the buyer's own agreement.
What to send
The request, in your own words.
Tell us these things. We turn them into a spec and a sourcing plan.
Send this when you are ready.
- Capability
- What you are training or evaluating.
- Domain
- The work. A profession or an operation, described specifically.
- Systems
- email, chat, CRM, tickets, documents, ERP, or others you name. Generic names are enough.
- Years
- How much history the task actually needs.
- Outcome signals
- What shows the work succeeded, failed, or was handed off.
- Exclusions
- What must never enter the spec.
- License scope
- Who may use it, for what, and for how long.
- Acceptance
- How you will judge a pilot sample.
Questions from buyers
Do you represent an AI lab?
No. Steamwire Labs is independent. We are not an agent of any AI lab or data vendor.
What should we send?
The capability, the domain, the systems of record, how many years of history the task needs, the outcome signals, the exclusions, the license scope, and your acceptance criteria.
Can you both introduce a company and prepare its data?
Not on the same company. Introductions and data preparation are separate engagements. For any given company, we do one or the other.