Human data for AI · Brazil

Real work and real speech, recorded for AI.

First-person video of homes, worksites and skilled trades, and read and spontaneous Brazilian Portuguese speech. Collected to your spec, with consent and an AI-training license on every recording.

Consent recorded per contributorNative QA before deliveryNDA and LGPD by default
The networkGrowing weekly
Illustrative map of the contributor network
Skilled tradesHousehold tasksPartner companiesSpeech contributorsPT-BR evaluatorsOnboarding now
2,000+registered contributors100 hof speech per week37+video tasks, others on demand
Ships with every recording
Consent and licenseSigned by the contributor before recording, versioned and hashed.
Metadata from the sourceTask, setting, device and speaker profile, captured at recording time.
Native reviewEvery file checked by a Brazilian QA team before it reaches you.
Clear withdrawal ruleA withdrawal blocks future licensing. Datasets already licensed stay valid.
Three lines of work

Capture what is missing. Evaluate what you have. Train what comes next.

Most of our work is capture: video for robotics and physical AI, speech for voice AI. Native evaluation and training data run on the same network.
EvaluateModel and product teams
Response comparisonNative PT-BR raters judge which answer is better, and why.
Red teamingStructured attempts to make the model fail, leak data or give harmful advice in Portuguese.
Benchmarks and gold setsPrompts and reference answers written for Brazilian users.
Evaluation →
TrainFine-tuning teams
Prompt and answer writingDemonstrations of how a great answer looks for a Brazilian user.
Preference dataPairwise rankings and written rationales from calibrated raters, with agreement reported.
Training data →
What you can order

Deliverables, not a headcount.

Three example orders. Every field is adjustable; the spec is agreed in writing before collection starts.
Example order · CaptureEgocentric video dataset
Volume
100 to 200 h pilot
Clip length
2 to 30 min per recording, set per task
Video
1080p minimum · 30 fps · 0.5x lens
Viewpoint
First person, head-mounted phone, hands in frame. Stereo rigs on request
Settings
Homes, kitchens, workshops, construction sites, cleaning and facilities
Metadata
Task label, setting, device, light check, review status
Consent
Signed per contributor, site declaration included
Full spec →
Example order · CaptureSpeech dataset
Volume
50 to 500 h
Language
Brazilian Portuguese, read and spontaneous
Regions
5 of 5 regions
Speakers
18 to 65 · balanced on request
Audio
WAV PCM 16-bit mono, 48 kHz, unprocessed
Included
Transcripts, speaker and device metadata, consent ledger
Consent
Signed per contributor
Full spec →
Example order · EvaluatePT-BR evaluation run
Volume
500 to 5,000
Raters
3 per case, lead review
Profile
Native Brazilian Portuguese, calibrated on gold tasks
Rubric
Agreed with your team before the run
Report
Scores per rubric, failures with severity, rater agreement
Raw data
JSONL or CSV
Turnaround
2 to 3 weeks
Full spec →
How we build the network

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

Contributors join through Verakin, our network across all five regions of Brazil, and through partner companies that bring their own teams.
  1. Stage 1SourceRecruitment by profile on demand: region, trade, age range, device.
  2. Stage 2ConsentTerms read and accepted in the app before the first recording, with version and timestamp.
  3. Stage 3Brief and practiceTask guides, framing checks and practice rounds before paid work.
  4. Stage 4Review and deliverOnly approved hours are paid and delivered. Contributors who drift are retrained or rotated out.
The contributor side runs on Verakin: Portuguese-first, mobile-first, paid via Pix per approved hour.Quality and compliance in detail
Capture partners

Real work happens inside companies. We bring them in.

Construction firms, commercial kitchens, cleaning companies and workshops join as capture partners. Their teams record during normal shifts. Participation is voluntary, consent is individual and site rules are agreed in writing before anyone records.

Empresa no Brasil? Seja parceira (PT) →
Worksites, not studiosReal sites, real tools, real pace. Nothing staged for the camera.
Volume in one placeMany workers on one site, with a field supervisor nearby.
Site rules in writingOff-limits areas, screens and documents agreed with the company first.
Phone or stereoPhones by default. Stereo head rigs when your spec needs depth.
Why Brazil

The variety lab data doesn’t have, in your time zone.

Real-world varietyHomes, worksites, trades and accents that English-first crowds don’t contain.
Hours that overlapOne to four hours from US time zones. Your team and ours work the same day.
A large Portuguese marketNative Brazilian Portuguese for voice and assistant launches in the region.
Mobile-first contributorsPeople already record and upload on their own phones, in the browser.
US and EU procurement

Buying from Brazil without the paperwork pain.

USD or EUR invoicesInternational contracts in English.
Mutual NDA firstBefore any data or sample is shared.
DPA availableData processed under LGPD.
Priced per accepted hourQuoted once we see your spec.
Training rights in writingUsage scope defined per project.
Pilot in one approvalSized to fit a single budget sign-off.
How to start

From your spec to a paid pilot in weeks.

Step 1Share your specThe spec you buy against: tasks, settings, speaker mix, format and volume.
Step 2Sample under NDAA sample matched to your spec, its metadata and a quote per accepted hour.
Step 3Paid pilotA small batch to your exact spec. Weekly capacity is set from what the pilot shows.
Step 4ScaleVolume plan, delivery cadence and quality reporting.

Tell us what your model is missing.

Send the spec you buy against. We start with a sample or a small paid pilot. Or write to contact@productlab.tech.

Request a sample