AI Data and Evaluation

Data and evaluation for AI systems in production.

Commission multilingual collection, annotation, validation, human review, linguistic QA, and model evaluation as one managed service line.

150+ languages Collection to evaluation EEA processing available
Discuss your dataset Review delivery controls
Coverage matrix 5 × 4
Modality Collection Annotation Validation Audit
Speech and audio Consent by scope Human QA IRR + bias Manifest by scope
Image and video Consent by scope Human QA IRR + bias Manifest by scope
LiDAR / 3D Consent by scope Human QA IRR + bias Manifest by scope
Text and NLP Consent by scope Human QA IRR + bias Manifest by scope
Sensor Consent by scope Human QA IRR + bias Manifest by scope
illustrative coverage controls set by scope not client evidence

Coverage across

  • Multilingual speech
  • Image and video
  • Text and NLP
  • LiDAR and sensor data
  • Expert evaluation
  • Linguistic QA
150+
Languages covered
210,000+
Contributors
50+
Countries
Human QA
Applied according to the agreed acceptance and sampling plan.

The data liability

Data quality, provenance, and evaluation affect deployment decisions.

Product, ML, procurement, and legal teams need the same project records to understand source quality, delivery controls, and remaining risk.

Regulatory non-conformance

For high-risk systems, incomplete representativeness, error, and traceability records can complicate the customer's conformity assessment.

Jurisdictional exposure

A delivery chain involving non-EEA entities may require additional transfer, sub-processor, and contractual review.

Audit-trail gaps

Missing provenance, consent, and quality records can delay procurement review and leave evidence gaps for the customer to resolve.

Operating services

Choose the managed data or evaluation work you need.

Data Collection

Source speech, text, image, video, and multimodal material across 150+ languages through a 210,000+ contributor network. Collection plans define purpose, contributor controls, consent or other applicable legal basis, and delivery evidence.

Explore collection

Annotation

Human annotation and review with task-specific guidelines, quality gates, and consensus methods where useful. Contract terms and Data Processing Agreement requirements are matched to the engagement.

Explore annotation

Validation

Evaluate datasets and model outputs for accuracy, representativeness, and defined failure modes. The resulting records can support Article 10 documentation where applicable; they do not certify the customer's system.

Explore validation

Ethical Framework

Define privacy, provenance, retention, deletion, and delivery controls for the engagement. EEA residency is the default, and the standard end-of-contract erasure commitment is 30 days.

Explore framework

How it works

From ingestion to delivery, every step has a record.

  1. 01

    Ingest and taxonomize

    Define modalities, languages, jurisdictions, quality criteria, and any applicable regulatory evidence before work begins.

  2. 02

    Route to suitable contributors

    Match tasks to relevant language, domain, and demographic requirements. Identity and consent controls are applied where the project requires them.

  3. 03

    AI-assisted tooling

    Pre-labeling and active learning can reduce human time on routine cases. Tooling and data location are agreed for the project.

  4. 04

    QA and consensus

    Human QA, sampling, inter-rater agreement, and representativeness checks are configured against the acceptance criteria.

  5. 05

    Secure delivery

    Deliver the agreed dataset, evaluation outputs, provenance, QA records, and transfer documentation to the customer's approved destination.

Delivery controls

Evidence supports the engagement.

Governance, privacy, and procurement requirements are translated into project controls and evidence where they apply. They support the service rather than define it.

Statutory framework

  • EU AI Act Article 10

    Data and data governance for high-risk AI systems

    Project evidence can include representativeness checks, error reports, sampling methods, and bias assessments that support the customer's Article 10 documentation.

  • GDPR Article 7

    Conditions for consent

    Where consent is the applicable basis, the project records purpose, scope, timestamp, and withdrawal handling for the relevant contributors.

  • GDPR Article 28

    Processor obligations and DPAs

    Standard Article 28 terms are available where YPAI acts as a processor. Roles, instructions, sub-processors, transfers, and audit provisions are agreed in the contract.

  • DORA + MiFID II

    ICT third-party risk and algorithmic trading

    Project controls can support third-party risk review and provenance requirements in regulated financial workflows.

  • US CLOUD Act

    Extraterritorial data demand exposure

    YPAI is a Norwegian entity and is a Norwegian company; data residency, subprocessors and transfer controls are defined per project. European infrastructure is the default; any customer-directed transfers and sub-processors are documented for the engagement.

Available project artifacts

  • Per-contributor consent register (GDPR Art. 7)
  • Dataset lineage manifest (EU AI Act Art. 10)
  • QA sampling and IRR record per dataset
  • Dataset version and change record
  • Sub-processor and transfer record where applicable
  • End-of-contract deletion record
Request a sample evidence schedule

What we deliver

Three illustrative engagement patterns.

Automotive and mobility

Project need A perception model needs stronger low-light edge-case coverage

Delivery approach Multi-camera and LiDAR annotation with a night-condition sampling plan

Evidence label Illustrative engagement pattern, not client evidence

Healthcare and life sciences

Project need A clinical NLP team needs Nordic ambient-speech material

Delivery approach A project-specific collection and review plan for sensitive health data

Evidence label Illustrative engagement pattern, not client evidence

Financial services

Project need A document-AI team needs multilingual suitability-review data

Delivery approach Structured text annotation with defined provenance and QA records

Evidence label Illustrative engagement pattern, not client evidence

Integrations

Delivers into your existing AI stack.

Supported delivery targets

  • Customer object storage
  • S3-compatible
  • Secure file transfer
  • Customer cloud environment
  • Warehouse export
  • ML metadata export
  • Customer API endpoint
  • Offline delivery

Delivery destinations, residency, access controls, and transfer requirements are agreed for each project.

Illustrative manifest structure 

// Explanatory structure, not client evidence
{
  "delivery_id": "agreed project identifier",
  "provenance": ["scope-specific records"],
  "quality": ["acceptance and review logs"],
  "transfer": "customer-approved destination"
}

Procurement FAQ

What procurement, legal, and security ask first.

How do you prevent foreign government access to our proprietary training data?

YPAI is a Norwegian company and is a Norwegian company; data residency, subprocessors and transfer controls are defined per project. European infrastructure is the default, while transfers and sub-processors are reviewed and documented for each engagement.

How do you document consent for data subjects involved in our datasets?

Where consent is the applicable legal basis, the collection plan records purpose, scope, timestamp, and withdrawal handling for the relevant contributors. The evidence package is defined for the project.

Are your processing contracts compliant with current European data protection laws?

Standard GDPR Article 28 terms are available where YPAI acts as a processor. The signed agreement defines instructions, roles, sub-processors, transfers, and any audit provisions.

Can we use your services to train models on special category personal data?

Potential projects involving special-category data require a documented lawful basis, purpose, safeguards, access model, and retention plan. Feasibility and roles are assessed before collection begins.

How do your datasets support high-risk AI system compliance?

Validation protocols can assess demographic and functional distributions, defined error modes, and representativeness. The resulting records support the customer's documentation; they do not certify the AI system.

Can we audit your data governance and annotation processes?

The engagement can include reviewable consent, provenance, QA, and deletion records. Contractual audit rights, response procedures, and evidence access are agreed during scoping.

Beyond data

Data and evaluation can stand alone or connect to implementation.

Enterprise Services shows how both independently purchasable service lines connect through one managed operating model.

Explore Enterprise Services

DATA PROJECT INTAKE

Scope a data project.

Bring modality, volume, jurisdictions, evaluation needs, and any regulatory context. A named project lead replies within one business day. NDA-first review on request.

  • EEA residency by default
  • Standard GDPR Article 28 terms where applicable
  • Project-specific provenance, QA, and delivery records
  • Named project lead, one business day reply
Modalities (optional)

GDPR Article 7 · GDPR Article 28 · EU AI Act Article 10