Traindeck is an expert data company. Credentialed specialists produce the judgment your model learns from — then the evaluations that prove it moved, the trajectories that show it how, and the environments where it practises.
traindeck · noun
the stack of expert judgment, evaluations and environments a model is trained against — assembled, verified, versioned, and dealt one hard hand at a time.
INVENTORY ACROSS LICENSED CORPORA AND PARTNER ORGANISATIONS · UPDATED MONTHLY
Data partnerships
Labs will pay for proprietary, real-world data they cannot scrape: how your engineers actually write code, how your operations team actually closes a ticket, how your analysts actually build a model. If your organisation holds that, licence it to us — you keep ownership, we handle the cleaning, structuring and sale.
What we licence
Your submission is ready. Copy it and send it to data@traindeck.ai — we reply within two working days.
What we ship
One supply chain of credentialed human judgment, sold four ways. Every line lands in your repo with its verification, its provenance and its known failure cases attached. Scope is bespoke; this is the starting shape.
The core product. Credentialed specialists author, grade, rank and adversarially probe inside their own domain — verified before they start, calibrated on paid sample work, then audited against gold sets for as long as they stay on the project.
Private, held-out task sets authored from scratch for your research question, with a grader harness you run yourself and baselines you can reproduce on day one.
Where static data stops being enough. Experts work a task end to end while we record why each step was taken, which tool call was wrong, and how they recovered after we deliberately broke something mid-task.
Where the model practises. CRMs, EHR front-ends, spreadsheets, ticketing systems, internal admin tools — running in a container, reset deterministically, with goals written by people who use them for a living.
your schema · Parquet or JSONLProcurement detail
"Environment" means different things to different vendors. Here is ours, written so your infra lead can argue with it before anyone signs anything.
| Line | Deliverable | Verification | Format |
|---|---|---|---|
| Expert data | Authored, graded or preference-ranked items with annotator credentials attached | Calibration gate, rolling gold-set audit, agreement thresholds | your schema; Parquet or JSONL |
| Evaluations | Held-out task set, grader harness, baseline run, per-item difficulty metadata | Canary strings, contamination screen, blind double-grade | task set + harness + baseline report |
| Trajectories | Demonstrations with tool calls, step rationales, error and recovery segments | Second-pass expert review, disagreement adjudication, per-step labels | JSONL, tool-call schema, per-row provenance |
| Environments | Software replica, task suite, seeds, expected-state fixtures, known-shortcut list | State-diff and unit assertions per task; rubric fallback only where the goal is open-ended | OCI image + env.yaml + verifier/*.py |
| Every line | Weekly batch notes: throughput, rejection rate, failure taxonomy, open questions | Named reviewers, audit trail from task to annotator to payment | dashboard + signed batch manifest |
How it works
We start from the capability you're trying to move, not a task-count quote. Out of that: a task taxonomy, a definition of done, and the failure modes worth paying for.
Sourcing against the exact profile — licence, seniority, tooling, language. Credentials verified, then a paid calibration task before anyone touches production data.
A small, deliberately hard batch delivered fast, with the verifier code in the open. You run it against your own models before we scale a single seat.
A red team of specialists tries to pass each task the wrong way. Every shortcut they find is either closed in the verifier or shipped to you as a documented known hack.
Throughput ramps while gold audits, agreement thresholds and rejection rates stay visible. If quality slips we cut volume before you find it in a training run.
Domains
Depth over breadth: each domain is a standing pool with its own reviewers and gold sets, not a job posting we open when a contract lands.
Quality & security
A weak verifier is worse than no data: the model learns the shortcut. We write the check first, try to break it, and tell you what we couldn't close.
Eval items are authored, never harvested. Canary strings, public-corpus screening, and strict separation between what we train on and what we test with.
Every item traces to a named contributor under a work-for-hire agreement with rights assigned to you. No scraped content entering a training set by accident.
Least-privilege access, per-project isolation, NDAs down to the annotator, and deployment into your VPC or on-prem where the work requires it.
Specialists are paid market rates for their field, on time, with a named reviewer and an appeal route. It's why the good ones take a second batch.
Throughput, rejection rate, agreement, cost per accepted item — live, per batch. Bad weeks are reported, not smoothed.
For experts
Frontier labs need people who can tell a plausible answer from a correct one. If that's your day job, teaching a model to do it is worth your evenings.
Get in touch
A pilot is small, paid and fast — one hard batch with the verification open, so you can judge the work instead of the pitch. Exploring a data partnership or joining as an expert? Same inbox rules: we reply within two working days.
partnerships@traindeck.aidata@traindeck.aiexperts@traindeck.ai