AI data annotation
Every frame passes under a human eye.
Image, video, text, audio, LiDAR and sensor-fusion annotation under one quality regime. Marked by people, countermarked where it matters, released against your acceptance plan, delivered with the record. Governed in the EEA.
Why annotation fails in training
Labels look right until the model says otherwise.
- 01
The hard 20 percent hides in frames nobody opened.
Sampled QA finds the errors you would have caught anyway. Occlusion, low light, crowds and mixed scripts hide in the frames a sample never reaches. Every frame is opened by a person.
- 02
Two annotators who agree are not proof.
Agreement is a number, scored per task against a reference set your team accepts. Where two independent markers disagree, a named reviewer decides. Nothing is averaged.
- 03
Labels without their record do not pass assessment.
EU AI Act Article 10 asks for representativeness, error examination, bias examination and provenance. Each arrives as a document with the labels, not as a promise.
The six routes
Six kinds of raw material. One table.
Perception work leans on image, video, LiDAR and fusion. Conversational work leans on text and audio. Each modality is its own service with its own guidelines, reviewers and acceptance metric. The name opens the route.
- Image annotation
Bounding boxes, polygons and keypoints on a real scene
- Video annotation
Multi-object tracks held across frames
- Semantic segmentation
Per-pixel masks over a real scene
- Text annotation
Entity spans held across mixed scripts
- Audio and speech
Diarized speaker lanes with transcript
- LiDAR and sensor fusion
3D cuboids on a point cloud, projected to camera
Data you do not hold yet is a collection, not an annotation. It is captured under the same quality regime and comes back onto this table. Scope a data collection
Working in a modality not on the table? We scope custom annotation protocols. Scope a project
Under the loupe
Where cheap labeling lets go, the mark holds.
Cheap labeling collapses on the hard 20 percent. These are the failure modes our guidelines and QA are written for.
- Occlusion
- The box is held through partial visibility. It does not shrink to what is visible.
- Low light
- The mask edge follows the object, not the light.
- Crowd density
- Every instance keeps its own identity. No two figures share a box.
- Mixed scripts
- Entity spans hold across Latin, Cyrillic and Arabic in one line.
Production work runs against your raw data under your engagement DPA.
The second marker
Two markers on one frame, and a number for how often they agree.
Quality is not a promise. Dispute-prone slices are annotated twice, independently. Agreement is scored with the method chosen for the task, against a reference set your team accepts. Disagreement goes to adjudication, not to averaging.
Every defect is classified by type (boundary, class, attribute, miss) and severity, and reported per class as a confusion matrix, so the next guideline revision has somewhere precise to begin.
Calibration comes first The production team clears the task-specific reference set and acceptance threshold before a single production item is marked.
Written in the margin
EU AI Act Article 10, satisfied at the label.
Governed from Norway.Processed in the EEA.A DPA on every engagement.
Article 10 obligations cascade to the annotation partner. Each clause maps to a concrete deliverable you can hand your conformity assessor.
- Article 10 representativeness Ontology and sampling design
- Demographic and segment distribution report
- Article 10 error-freeness Kappa-gated QA and gold sets
- Per-class agreement and defect report
- Article 10 bias examination Label-level bias audit
- Bias-examination notes per dataset
- Article 10 provenance Per-item provenance logging
- Dataset datasheet and provenance log
- GDPR Articles 9 and 25 Lawful basis, minimization, data protection by design
- Signed DPA, 30-day erasure SLA, sub-processor list
Compliance is evidenced through EEA jurisdiction, named regulations and audit-ready documentation, not third-party certification badges.
Scale on proof, not promises
You commit to one sheet before you commit to scale.
Scope first: objectives, modalities, taxonomies, quality targets, risk level and DPA, defined with your team. Then a measured pilot with iterative guideline refinement and reporting against the agreed metrics. Production only after the exit gate: metric targets met and a stable, repeatable process, with SLAs on throughput and defect ceilings. Then continuous QA, relabeling campaigns and the annotation-to-model feedback loop.
The exit gate is named before scale: metric targets plus a stable process, not a single number.
Scope your pilot
Put your data on the table.
We scope the modalities, taxonomies and quality targets your deployment needs. If YPAI is not the right fit, we will say so directly. A managed European partner absorbs the QA, PM and compliance overhead a marketplace pushes back onto you.
FAQ
Frequently asked questions
What types of AI data annotation does YPAI support?
YPAI supports image annotation (bounding boxes, polygons, semantic segmentation, keypoints), video annotation (object tracking, event labelling), NLP annotation (named-entity recognition, intent, sentiment), audio and speech annotation (transcription, diarisation, prosody), and 3D / LiDAR point-cloud annotation for autonomous-vehicle perception stacks. All modalities run inside a single quality and compliance regime so multimodal projects share the same review, provenance, and audit-trail pipeline. See AI data labeling for the labelling-versus-annotation distinction.
How does YPAI ensure annotation quality?
YPAI runs a multi-stage review pipeline: spec calibration, double-pass annotation on dispute-prone slices, expert adjudication, and a final QA sweep against task-specific acceptance criteria. Each batch ships with an inter-annotator-agreement (IAA) report and a per-label confusion matrix so machine-learning teams can target re-work where it materially affects training. Sample sizes are statistically scoped to the project rather than fixed.
Is YPAI annotation GDPR-compliant?
Yes. Annotation pipelines are built under GDPR from the start: lawful basis, purpose limitation, data minimisation, and Article 35 DPIA scaffolding are part of the engagement, not a checklist that arrives at the end. Personal data is processed inside EU/EEA infrastructure under a signed Data Processing Agreement (DPA), and subject-rights requests (access, rectification, erasure) route through the data request form. See data ethical framework for the underlying governance.
Which languages does YPAI cover for text and speech annotation?
YPAI has delivered annotation across 150+ languages and dialects, with native-speaker reviewers concentrated in European, Nordic, and major Asian markets. Lower-resource languages are quoted on a per-project basis because the limiting factor is reviewer recruitment, not pipeline capacity. Multilingual projects route through a single delivery lead, so glossary, style guide, and IAA targets stay consistent across languages.
How is YPAI different from Scale AI, Labelbox, or Appen?
YPAI is an Oslo-headquartered EU/EEA-native operator: data lives under EU jurisdiction by corporate structure, not by contract terms. YPAI is not a US-domiciled provider; data residency, subprocessors and transfer controls are defined per project. Engagements are run by in-house teams with sector specialists (automotive OEM, healthcare ASR, financial documents), not a public crowd marketplace. Pricing is per-project after a feasibility scope, not per-task with hidden minimums.
Can YPAI annotate medical imaging data (DICOM, FHIR)?
Yes. YPAI handles DICOM, NIfTI, and FHIR-bundled imaging with clinician reviewers contracted for the modality. Data is processed inside EEA residency by default, and the pipeline supports GDPR Article 9 special-category handling. The exact controls and evidence artifacts are documented in the statement of work.
What is the minimum project size YPAI accepts?
YPAI prioritises engagements where the annotation work is non-trivial and the compliance posture matters. Below roughly a single-batch pilot (a few thousand items for image / NLP, a small audio corpus for speech) the engagement model rarely justifies the scoping cost on either side. Pilots are common and welcomed; the scoping call confirms whether the brief is a fit before any quote.
How fast does YPAI reply after a project inquiry?
YPAI replies inside one EU business day after a submission to /contact-us/ with a feasibility read, the next concrete step, and an estimated scope window. The first reply is from an engineer or delivery lead with context on the brief, not a generic confirmation email.