Closing your company? Your data can still earn. License it to AI labs before you shut down.See what it's worth

Evaluations, expert data and environments for real work.

We find where your model breaks, build the data that closes the gap, and give it a place to practice.

Request a sampleSee how it works
01 · EVALUATE

See where your model breaks, step by step.

Practitioners write long tasks from their own jobs. We grade every step, so failures show up as a map, not a single score.

  • Real multi-step work, not quiz questions
  • Expert rubrics plus automatic end-state checks
  • Re-run on every model, checkpoint and vendor
Failure map · 1,200 tasksSteps 1 to 8
12345678Month-end closeBank reconciliationVendor onboardingContract reviewRefund ticketsCRM cleanup
PassesSometimesFails
Late steps that tie numbers together62% fail
Spotting duplicates and exceptions48% fail
Example output
02 · CURATE

Expert judgment, captured as data.

Verified professionals solve tasks, compare answers and write rubrics, aimed at the gaps the evaluation found.

  • Reasoning traces for SFT
  • Preference pairs and rubrics for RLHF
  • Computer-use trajectories from our sandbox
Preference review · Senior CAB preferred
PROMPT

A vendor was paid twice for invoice INV-20931. What should the accountant do?

Response AModel
Reverse the second payment in the ledger and close the ticket.
Response BChosen
Ask the vendor for a refund or credit note, record it against INV-20931, then re-tie the AP sub-ledger.
Correct accounting treatment5/5
Controls and audit trail4/5
Clear next steps5/5
03 · TRAIN

RL environments that teach the job.

A sandboxed copy of real business software. The agent acts, the world changes, a verifier scores the result, and the reward teaches it.

  • ERP, CRM, ticketing, spreadsheets, IDEs and browsers
  • Verifiers check the end state and the path
  • Docker or hosted, over API and MCP
How an episode runsRunning
AgentModel being trainedToolspost_journal, approve_invoiceWorld stateLedger, vendors, ticketsVerifierReward 0.0 to 1.0Thousands of episodes per task
04 · SOURCE

Built from how real companies work.

Companies license their records to us. We remove personal details and turn real workflows into tasks and environments.

  • Signed license with every data owner
  • Meetings, tickets, docs, CRM and code
  • Every source listed in the data card
For companies with data
GGong call transcriptsJJira ticketsNNotion SOPsSSlack threads
De-identifyNames removedCustomers removedEmails, phones removedOwner sign-off done
kestrel-ops-erp2,400 tasks14 toolsReal workflows, no real people
WHAT YOU RECEIVE

One package, ready for your pipeline.

Every delivery comes with lineage: who made each item, how it was graded, and where the data came from.

SFTRLHFDPORL with verifiable rewardsRubric rewardsComputer use
delivery_2026_11/Signed
trajectories/2,000 tasks
  close_books.jsonl312 MB
preferences/8,400 pairs
rubrics/146 rubrics
environment/Docker + MCP
  Dockerfile
eval_harness.pyRe-run any model
data_card.mdSources and licenses
lineage.csvExpert per item

Start with a free sample.

Scoping callFree sample in 5 daysPilot of ~200 tasksProduction
Request a sample