Expert AI trainers, proven on real tasks before they touch yours.
Cupel tests specialists in coding, scientific computing and evaluation on practical assessments, scores every result with two blind reviewers, and delivers the experts who pass as a managed pod with a reviewer on every task.
New and taking our first audit and pilot clients now. We publish our rubrics so you can see exactly what we test; we never publish the answer keys. Are you an expert? Take the assessment.
Two blind reviews on every expert's assessment.
First test scored against a fixed answer key.
No client files kept after delivery.
Replies within one working day.
What a pod delivers
Four kinds of expert work, each with a written definition of done. Your own checks decide acceptance; ours run first.
Why teams start with Cupel
Most ways of buying expert work ask you to trust a CV. Cupel shows you the evidence first and lets you start small.
| Question | Cupel | Open expert marketplace | Hiring in-house |
|---|---|---|---|
| How experts are proven | A practical test scored against an answer key, two blind reviews and a call on their own work | Varies by vendor | Your own hiring loop |
| Review before you see work | Every task | Varies by vendor | Your team's time |
| Smallest first step | A $750 founding audit of up to 20 tasks | Varies by vendor | A full hire |
| Time to start | Within a week of the scoping call | Varies by vendor | Months |
| Your data | Stays in your systems; nothing kept after delivery | Varies by vendor | Yours |
Start small: a task quality audit
Send up to 20 of your tasks, evaluation items or rubrics. Within five working days you get every defect that would fail a lab's review, why it fails, and the fix.
- You send
- Up to 20 tasks with their tests, reference solutions or answer keys, under NDA.
- You get
- Findings for each task, the patterns across the set, and the fixes ranked by impact.
- Time
- Five working days from receipt.
- Price
- $1,500 fixed. Founding price of $750 for our first three clients. If the report gives you nothing you would act on, you do not pay.
- Your data
- Read only by named experts under NDA, and deleted when the report is delivered.
What a finding looks like
- Task 07, hidden testsFails review. The test compares floating-point output with ==, so a correct solution fails on different hardware. Fix: compare with a stated relative tolerance and put it in the instruction.
- Task 12, reference solutionFails review. The reference fails 2 of 14 hidden tests on a clean run, so the task cannot be solved as written. Fix: run the reference inside the grading image before shipping.
- Rubric 5, criterion 3Inconsistent grading. "Clear explanation" has no observable anchor, so two reviewers gave the same response 2 and 5. Fix: tie each level to something visible in the response.
See a full sample report
Illustrative: five coding tasks written for this sample, not client data.
| Task | Verdict | Findings |
|---|---|---|
| T1 | Ship after fixes | 1 minor |
| T2 | Ship | None |
| T3 | Rework | 2 fail review, 1 minor |
| T4 | Ship after fixes | 1 fails review |
| T5 | Rework | 2 fail review |
Patterns across the set
Three of five tasks keep expected outputs in files the agent can read. Two take longer than a 10-minute grading budget on a clean machine.
- T3, hidden testsFails review. Expected outputs sit in a fixtures folder inside the agent's workspace, so an agent can copy them instead of solving the task. Fix: move fixtures into the grader image, outside the workspace.
- T5, environmentFails review. The build compiles a large dependency from source and takes 14 minutes on a clean grader that stops at 10. Fix: pin a prebuilt package.
- T1, instructionMinor. "Should handle errors gracefully" is vague and untested. Fix: name the error cases and test their exit codes.
Every report ends with the fixes ranked by how many tasks each one unblocks.
Already agreed a price? Pay securely with PayPal.
What we check
Every item is read in full, run from a clean environment and attacked the way a model would attack it. These are the checks behind every finding.
Coding tasks and RL environments
- Every behaviour the tests check is stated in the instruction
- Tests judge behaviour, never private implementation details
- Time, randomness, ordering and floating point are controlled
- Edge and negative cases, plus a positive case for every feature
- Three shortcut attempts, including reading expected outputs and hardcoding
- Builds from clean with pinned versions, no network, inside the time budget
- The reference solution passes inside the grading image
Scientific-computing tasks
- Units and conventions stated once and used consistently
- Reference values recomputed by an independent method
- Tolerances justified by an error estimate, relative where scale varies
- Degenerate and ill-conditioned cases handled
- Invariants, analytic limits and convergence order tested
- Seeds, versions and compute budget stated
Evaluation questions
- One defensible answer, confirmed by a second method
- Self-contained: every number, figure and convention included
- The answer format cannot give the answer away
- Hard because of reasoning, not obscurity or trick wording
- Not answerable by search or recall of a famous result
Rubrics and output reviews
- Each criterion can be observed in the response
- Every score level has an anchor reviewers can apply
- Weights match what matters for your training goal
- Three sample responses scored twice to test agreement
- Must-catch errors listed and checked
How an audit runs
Day 0: scope
A 15-minute call, an NDA, and access to the tasks in your environment or as files.
Days 1 to 3: review
Every item read, run clean and attacked with shortcut attempts; numbers recomputed.
Day 4: confirm
Every finding that would fail review is re-run and checked a second time.
Day 5: report
Findings with evidence and fixes, patterns across the set, and a 30-minute walkthrough.
Next 14 days
Questions answered and up to five fixed items re-checked at no charge.
What you receive
- Findings for every item: severity, evidence and the exact fix
- Patterns across the set, with fixes ranked by how many items each unblocks
- Every finding as a CSV you can import into your tracker
- A 30-minute walkthrough call and 14 days of follow-up
Estimate your audit
Set the size and type of your set. The written quote after the scoping call is final.
- Standard price
- $1,500
- Founding price
- $750, for our first three clients
- Report in
- 5 working days
- Included
- Walkthrough call, CSV of findings, 14 days of follow-up and re-checks of up to 5 fixed items
How every expert is vetted
CVs and quizzes say little about whether someone can build training data. So we test the work itself, the same way for everyone.
Apply
One role, with their field, experience and the hours they can give.
Practical assessment
Two to three hours: find the planted defects in a flawed task, then design an original one.
Two blind reviews
Two reviewers score the work against a weighted rubric without seeing the name. Splits of two points are settled before any decision.
Follow-up call
A short call about their own submission. People who did not write the work cannot defend it.
Passport
Scores per criterion, number of reviewers, date and level, recorded and shared with you on the pod sheet.
Part 1 of every assessment is scored against a fixed answer key of planted defects, so the first number on every passport is objective. Part 2 is scored by two reviewers whose agreement we track.
What a passport is
A record that an expert passed a Cupel assessment in one domain. It has a public code (CP- followed by ten characters), the domain, the level, the status (active, suspended, expired or revoked) and the dates it was issued and expires. The expert and the teams they work with also see the scores per criterion and the number of reviewers.
Who can check it
Anyone, on the passport check page, with the code. The check shows the status, domain, level and dates. It never shows the expert's name, email or submission.
How long it lasts
Twelve months, in every domain. After that it shows as expired until the expert passes a fresh assessment.
Assessment integrity
Experts may use the references and tools they would use at work. They may not paste the materials into an AI assistant, submit AI-written text, or use anything from client, employer or platform work they must keep confidential. Every submission is followed by a short live call about the expert's own choices.
How we check
Part 1 is scored against a fixed answer key. Part 2 is scored by two reviewers who never see the expert's name, and splits are settled before any decision. A passport is suspended while a concern is looked into, and revoked if the work turns out not to be the expert's own.
The two-week pilot
Small enough to decide quickly, real enough to measure. Here is what you get.
- Team
- A pod of passed experts matched to your work type, plus a Cupel reviewer on every task.
- Start
- Within a week of a 30-minute scoping call where you set the acceptance criteria.
- Quality
- Every task passes our review before it reaches you. Any task that fails your acceptance check is reworked or replaced at no charge.
- Security
- NDA with every expert. No AI-written submissions. Your task content stays in your systems; we keep only our own labels and outcomes.
- Price
- From $125 per expert hour with review included, or a fixed price per accepted task. Agreed in writing before work starts.
- Reporting
- Weekly: tasks submitted and accepted, your acceptance rate, the defects your checks found and what we changed.
Request an audit or a pilot
Tell us what you need. We reply within one working day.
How a pilot runs
Week 0: set up
A scoping call, your acceptance criteria in writing, and a pod sheet with each expert's passport scores.
Week 1: calibrate
Five tasks per expert, reviewed against your criteria. You approve the bar before full volume.
Week 2: full volume
Every task reviewed by Cupel before it reaches you, with a weekly report on your acceptance rate.
Close: decide
An acceptance summary and a clear choice to scale, adjust or stop. No lock-in.
Built into every pilot
- Any expert you are not happy with is replaced at no cost.
- Stop at any weekly checkpoint and pay only for work delivered to that point.
- Tasks that fail your acceptance check are reworked or replaced at no charge.
Pay for agreed work
Pay here once we have agreed the scope and price in writing. PayPal processes the payment, so Cupel never sees your card details.
- Founding audit
- $750, for our first three clients
- Task audit
- $1,500 for up to 20 tasks
- Pilots
- As quoted. For larger amounts we send a PayPal invoice.
Who pays, and when
- Teams
- Pay Cupel. Audits are paid once the scope and price are agreed in writing. If the report gives you nothing you would act on, we refund it in full.
- Pilots
- From $125 per expert hour with review included, or a fixed price per accepted task, set in a written quote before work starts.
- Experts
- Never pay to apply or to be assessed. They are paid their own quoted rate for accepted work, by PayPal or Wise, within 7 days of the client accepting it.
- Disputes
- If a team and an expert disagree on whether work passes the acceptance check, a second Cupel reviewer decides within 5 working days. Work that fails gets free rework.
How to pay
- Enter the exact amount from your written quote.
- Add your organisation name and quote reference in the PayPal details.
- Pay with PayPal or a debit or credit card.
For experts
If you can build hard, checkable work in your field, we want to see it. Choose one role. The assessment is the application.
Paid, with the rate shown first
Every pilot has a published rate, per hour or per accepted task. You see it before you accept the work. Accepted work is paid within 7 days of the client accepting it, by PayPal or Wise.
Your passport is yours
Pass, and you receive a Cupel passport with your scores per criterion, the number of reviewers and the date. Share it wherever you like.
Rules
Take the assessment
Read the brief, do the work in your own time, and paste your answers below. Your draft is kept in this browser until you finish.
Who runs Cupel
Cupel is run by Peter Ndungu, a working AI data contractor. He builds coding and scientific-computing tasks for AI training and reviews other authors' work, so he knows first-hand why tasks get rejected: hidden requirements, flaky tests, open shortcuts and reference values nobody re-checked.
Cupel is new, and we say so plainly. It is operated by Peter Ndungu in Nairobi, Kenya, and company registration is in progress. We have no client case studies to show yet: the founding expert cohort is forming now, the first three audits are offered at a founding price, and every audit in the first months is reviewed by Peter personally. Client names stay confidential, including those of the platforms Peter works with. A custom domain and role-based email addresses are next; until then, the contact form below is the way to reach us.
Talk to Peter directly
Use the contact form on this page. Peter replies within one working day, by email to the address you give.
What we promise in writing
A fixed price before work starts, a reviewer on every task, free rework for anything that fails your acceptance check, and no client content kept after delivery.
Contact Cupel
Questions, privacy or deletion requests, payments and refunds. Your message goes to Peter, who replies within one working day by email to the address you give.
Privacy and assessment terms
Plain-language terms for experts who apply and teams who send us work.
- What we collect from experts: your name, email, the links you share, your submission and our reviewers' scores.
- Why: to assess you, match you to paid work and issue your passport if you pass.
- Who sees it: Peter and invited Cupel reviewers. Reviewers see a submission code, not your name. Buyers see your passport under an anonymous code unless you agree otherwise.
- How long we keep it: submissions that do not pass are deleted after 12 months, and anything is deleted sooner if you ask.
- Your work stays yours: you keep ownership of what you submit for the assessment. We never sell your data or use your submission as client work.
- Teams who send us work: task content stays in your systems; we keep only our own labels and outcomes. Files sent for an audit are deleted when the report is delivered.
- Payments: processed by PayPal. We receive the payer's name, email and amount, never card details.
- Questions or deletion requests: use the contact form and choose "A privacy or deletion request". We reply by email within one working day.
Questions
Who reviews the work?
Practitioners in the same field who passed the same assessment. Each submission gets two reviewers who score it independently, without the expert's name.
Can experts use AI tools?
For research, yes: documentation, papers, solvers and compilers. Submissions must be their own writing, and the follow-up call checks that they can explain every choice.
What happens to my submission?
We use it only to assess you. Reviewers see a submission code. Ask us at any time to delete it.
Do experts see client data?
Only the named experts on your pilot, under NDA, inside your systems. We keep our own task labels and outcomes, never your content.
How fast can a pilot start?
Within a week of the scoping call, once the pod is matched to your work type.
What if we do not like the work?
A task that fails your acceptance check is reworked or replaced at no charge, and the weekly report shows you the acceptance rate on your own checks.