Human data for AI · Evaluation, video and speech

AI needs more experts in the loop. We’re building the pool.

Qualified AI evaluators are scarce: the same small crowds, mostly English-speaking, mostly generalists. We are opening a new, verified pool of licensed professionals, trained evaluators and video and voice contributors in Brazil to evaluate your models and capture the real-world data that trains them.

Licenses verified on public registriesCalibrated on gold tasksNDA and LGPD by default
The networkGrowing weekly
HealthcareLegalAccounting and taxEngineeringTrained evaluatorsOnboarding now
LicensedspecialistsCalibratedevaluatorsNationwidecoverage
Illustrative. Dot counts do not represent actual headcount.
Why now

Models are getting better. The pool that evaluates them isn’t growing.

As models move into medicine, law and finance, every team needs expert judgment at the same time, from the same few sources.

01 · ScarcityThe same small poolsA handful of vendors and crowds serve most of the industry. Long queues, repeated raters and little domain depth.
02 · DepthGeneralists grading specialistsA fluent answer about dosage or tax law needs a nurse or an accountant to judge it, not a general annotator.
03 · CoverageEnglish first, everyone else laterOther languages, regulations and cultures are under-represented in evals and training data, so models underperform there.
The result: evals that miss real-world errors, and training data that plateaus.Our answer ↓
How we expand the pool

We don’t rent a crowd. We build and qualify one.

Brazil has a large base of licensed professionals who have never been asked to work on AI. We recruit them nationwide, verify them and train them for evaluation and training tasks.

  1. ApplicantsStage 1SourceNationwide recruitment through Verakin, our contributor network, reaching professionals and everyday workers the usual crowd platforms never see.
  2. VerifiedStage 2VerifyIdentity checks and professional licenses confirmed on each council’s public registry: COREN, OAB, CRC, CREA, CRM.
  3. CalibratedStage 3Train and calibrateTask guidelines, practice rounds and gold questions. Only raters above the agreement threshold move on.
  4. ActiveStage 4Deliver and monitorMulti-rater work, lead review and continuous quality scores. Raters who drift are retrained or rotated out.
Tier 1Licensed specialistsNurses, doctors, lawyers, accountants and engineers with an active council registration. For domain-critical judgment.
Tier 2Trained evaluatorsDigitally fluent raters, many bilingual, calibrated for preference ranking, safety review and general quality.
Tier 3Video and voice contributorsPeople across Brazil, from nurses to mechanics, cooks and farmers, who record first-person video of real work and voice samples.
What the network does for you

Evaluate what you have. Capture what trains the next model.

Evaluation
Assistant evaluationAI in productionEvery answer graded for correctness, completeness and risk by experts in your domain.
Red teamingBefore launchStructured attempts to make the model fail, leak data or give dangerous advice.
Benchmarks and gold setsEvery releaseExpert-written prompts and reference answers to measure each model version.
Training
Supervised fine-tuning dataSFTExpert-authored demonstrations of how a great answer looks in your domain.
Preference and reward dataRLHFPairwise rankings and written rationales from calibrated raters, with agreement reported.
Real-world capture
Egocentric videoRobotics · embodied AIFirst-person video of real tasks, recorded with head-mounted phones, hands in frame, reviewed before delivery.
Speech dataVoice AI · ASRBrazilian Portuguese voice recordings across regions, ages and accents, with transcripts.
Custom collectionAny task, any regionWe brief, recruit and pay contributors for the exact scenarios your model is missing.
Egocentric video

Robots learn from hands at work. We record them.

Contributors across Brazil strap a phone to their head and do real tasks at a natural pace: cooking, fixing an engine, laying bricks, sewing, harvesting. You get first-person footage of the physical world, with the diversity robotics and embodied AI teams can’t find in lab data.

  • Wide-angle phone cameras, hands in frame, natural pace
  • Kitchens, workshops, construction sites, farms and homes
  • Every clip reviewed for light, framing and task completion
  • Signed consent per contributor; only the contributor appears
  • Task labels and metadata delivered with the footage
REC · POV00:14:32
TASK: CHOPPING VEGETABLESSETTING: HOME KITCHENHANDS: 2LIGHT: PASSREVIEW: APPROVED
Illustrative frame and metadata
The network behind it

Our contributor network runs on Verakin.

Verakin is where professionals and everyday workers across Brazil sign up, get verified and are paid every week via Pix for evaluation, video and voice projects. Portuguese-first, mobile-first and built so contributors trust it, which is why they show up and do good work.

EvaluationLicensed expertsVideoFirst-person tasksVoiceRegional accents
Visit verakin.com.br ↗
Why Brazil

An untapped bench of experts, in your time zone.

Depth of talentLarge, regulated professions with public license registries, mostly unreached by AI data vendors.
Hours that overlapOne to four hours from US time zones. Your team and ours work the same day.
Portuguese and EnglishNative Brazilian Portuguese for local launches, bilingual raters for English work.
Local rules, real casesHealth plans, consumer protection, tax and labor rules that generic raters don’t know.
Sample report

What lands on your desk.

Scores per rubric, every failure with a real example and a severity, the rater agreement behind it and what to fix first. Raw data in JSONL or CSV.

Evaluation report · IllustrativeMember support assistant · Health plan, Brazil
400 cases · 3 raters each
Inter-rater agreement 0.81
82%
Correct
74%
Complete
3.1%
High-risk answers
11%
Should escalate
FindingSeverityCases
Suggests dosage without asking patient weightHigh7
Quotes waiting periods that contradict ANS rulesMedium12
Does not hand off to a human on urgent symptomsMedium9
European Portuguese terms confuse membersLow21
Example · reviewer note“The answer is fluent but applies a 180-day waiting period to an emergency. Under ANS rules the limit is 24 hours.” Registered nurse, COREN verified
How a pilot works

A pilot in weeks, not quarters.

  1. Step 1ScopeYour use case, the risks that worry you and what a good answer looks like.
  2. Step 2Team and guidelinesWe staff a calibrated team for your domain and agree the rubric and case set with you.
  3. Step 3Evaluation or data workMultiple raters per item, reviewed by a domain lead, with agreement tracked daily.
  4. Step 4Report and scaleFindings walked through with your team, then a plan to scale volume.
Domains

Where the pool is deepest today.

Need a profession not listed? We recruit and verify for new domains on demand.

HealthcareCOREN · CRM
LegalOAB
Accounting and taxCRC
EngineeringCREA
Financial servicesVetted professionals
General evaluationTrained raters
Request a pilot

Tell us which experts your model is missing.

An English-speaking project lead replies quickly with a proposed team, rubric and timeline. Invoicing in USD or EUR.

  • Mutual NDA before any data is shared
  • Pilot sized to fit a single budget approval
  • Data processed under LGPD, DPA available
Are you in Brazil and want to contribute?Join the network on Verakin ↗

We use your details only to reply to this request.