For data, legal and procurement teams

How a recording becomes data you can train on.

Three questions decide whether a dataset is usable: is it good, is it clean, and do you have the rights. This page answers each one.

1 · Quality pipeline

Eight gates between a contributor and your bucket.

  1. 01RecruitmentNationwide, through Verakin, by profile and region.
  2. 02Identity and licenseIdentity checks. Licenses confirmed on COREN, OAB, CRC, CREA and CRM public registries.
  3. 03Task trainingGuidelines and practice rounds for the specific task.
  4. 04Gold questionsOnly contributors above the agreement threshold move on.
  5. 05Automated checksVideo: light, framing, hands, blur. Audio: clipping, silence, SNR.
  6. 06Human reviewMulti-rater work and domain lead review.
  7. 07Metadata validationJSON schema check and SHA-256 hash on every file.
  8. 08Delivery and monitoringQuality scores per batch. Contributors who drift are retrained or rotated out.
95%+usable items per delivery, or we replace them
100%of items reviewed by a person before delivery
0.80minimum rater agreement (Krippendorff’s α) to ship an evaluation batch
100%of required metadata fields filled, checked by schema
2 · Privacy

What we keep out of video and audio.

Footage from homes and worksites can pick up things nobody consented to share. Each has a rule.

Other people and children
Only the contributor appears. Faces in focus are blurred. Clips with children are rejected.
Faces
Recognizable faces in mirrors, photos or reflections are blurred or cut.
Documents and screens
Readable screens and documents are blurred or cut.
Addresses and plates
House numbers and plates are blurred.
Voice as biometric data
Consent signed per contributor. Voice is never used to identify the speaker. IDs are pseudonymized per delivery.
Your prompts and outputs
Mutual NDA before any data is shared. Processed under LGPD, DPA available.
3 · Rights

What you are allowed to do with the data.

Set per project and written into the contract, backed by the agreement each contributor signs.

Model training rights
Perpetual, worldwide training and evaluation rights
Commercial use
Use in commercial models and products
Exclusivity
Exclusive for custom collections, non-exclusive for catalog data
Contributor agreement
Full transfer of approved recordings, likeness used for AI only, no advertising. Template shared under NDA
Consent withdrawal
Contributors can withdraw at any time. It blocks future licensing and does not affect datasets already licensed
Retention and deletion
Raw files kept 12 months for re-delivery, then deleted unless the contract says otherwise
Geographic restrictions
None. Collection happens in Brazil only
Procurement

What your vendor form will ask.

Invoicing currency
USD or EUR
Contract language
English
NDA
Mutual, before data
DPA
Available
Data protection law
LGPD
Project lead
English-speaking
Time zone
1 to 4 h from US zones
Legal entity
Brazilian limited company · CNPJ 47.684.137/0001-76
Data storage
Cloud providers outside Brazil, encrypted at rest, under LGPD transfer safeguards
Security documentation
Security overview and DPA template, on request

Check the quality yourself.

Ask for a sample with its metadata and the contributor agreement template. Shared under NDA once we have your spec.

Request a sample