03 - TALENT

The top 1% of global experts, verified

We run an admissions process rather than a hiring one. Around 1 in 50 applicants get in, and staying in means passing calibration every week.

Apply as an expert
Last admissions cycle
Applied
0
Credential screen−85.0%
0
Domain examination−60.0%
0
Paid work trial−50.0%
0
Admitted
0
Admission rate0%
The bar

Quality starts at the gate

We secure data fidelity by strictly vetting who gets to author your ground truth.

Last admissions cycle

Each tile is an applicant. The green ones cleared all four gates. Only they engineer our data.

Admission rate

0

of applicants admitted

200 admitted from 10,000 applications. Four gates sit between an application and a high-fidelity record, and the fourth one stays open for good.

Rejected at each gate

Credential screen−0
Domain examination−0
Paid work trial−0
ongoing, forever

Four gates

Stage 01

Credential check

We verify degrees, professional licenses, and work history directly with the organizations that issued them.

Stage 02

Domain testing

Candidates take strict, timed exams written and monitored by our senior experts.

Stage 03

Work trial

Candidates complete paid, real-world tasks. Their work is hidden-tested against perfect answers.

Stage 04

Continuous quality checks

We test our experts weekly. If their work drops below our high standards twice, they are removed from the team.

OUTPUT
What they do

Two ways experts move your model

Experts write data that does not exist yet, or grade what your model produces. Both end up in your training loop.

lh2 · expert output live
01 · Build
Creating what doesn't exist. When no dataset covers the capability you need, vetted specialists author it from scratch - the reasoning, the reference answers, and the tasks themselves.
Reasoning traceshow the answer is reached
observe · 3-week cough, night sweats
rule out · TB - no exposure, CXR clear
weigh · 20 pack-year hx shifts odds
conclude · CT chest, not repeat CXR
4 of 7 steps · derivationauthored
How a specialist actually gets to the answer
Golden Datasetreference · rubric-graded
goldStage IIIa NSCLC - chemoradiation
correctness4/4
citation3/4
reasoning2/4
✓graded against gold · 9 / 12
Reference answers with graded rubrics
Instruction dataauthored, not scraped
taskReconcile an FX-hedged invoice
  1. 01pull the hedge contract
  2. 02revalue at settlement rate
  3. 03book the variance to CTA
✎written by a controller · 11 yrs
Real procedures, written down by the person who does them
02 · Evaluate
Judging what models produce. The same specialists grade your model's output - ranking responses, scoring agent episodes, and probing for the failures only a practitioner would find.
Preference rankingpairwise judgments
A
vs
B
ranked by 3 specialistsκ 0.86
Expert judgments for RLHF and DPO
Trajectory gradingagent episodes
episodeRefund an out-of-policy order
✓03fetch_exceptions(tier)
✗04issue_refund() · no approval
✓05notify_customer(refund)
rubric 7.4 / 10 · docked at step 4
Rubric scores for agent episodes
Red-team reviewadversarial probes
probe setContraindication edge cases
✗10interaction · 2 CYP3A4 inhibitors
!1 of 18 probes breaks it
Adversarial probing by domain insiders
ENGAGE
Engagement models

Expert capacity, shaped to your pipeline

Some teams want a standing bench, others want a surge for one release, and some want graders inside the training loop. Pick the shape you need.

001

Dedicated expert pods

A named team of specialists in your workflow, with a calibration lead answerable for quality. You get the same people each week, so they keep the context instead of relearning it.

002

Surge capacity

Scale to thousands of expert-hours for a dataset push or a pre-launch eval sprint, then scale back down. You pay for the peak you used.

003

Experts-in-the-loop

Human feedback wired into your training runs and agent deployments. Graders sit in the loop and catch failures while the run is still going.

Get paid to write the data models learn from

If you are at the top of your field in engineering, medicine, finance, law or language, we pay you to prove it.

Apply as an expert