How our data annotation process works, from raw data to training data
We run your annotation work end to end: we write the guidelines with you, train a dedicated team on your edge cases, check every batch in layers, and deliver in your format on a schedule you can plan around.
Six steps from first call to steady production
The early steps are short, but they decide how accurate everything after them will be, so we don’t skip them.
- Days 1–3
Scope the work with you
A project lead and an annotation specialist meet your ML team to understand the model, what “correct” means for it and where it fails today. We review sample data, agree accuracy targets, throughput and tooling, and sign your NDA and DPA.
- Days 3–6
Write the guidelines together
We turn your intent into an illustrated guideline: every class defined, every edge case shown with a right and a wrong example, decision trees for occlusion and overlap, and a gold set for testing.
- Days 5–10
Run a pilot
A small team labels up to 500 items from your real data. You see accuracy, speed and our disagreement log before you commit, and most guideline updates happen here.
- Weeks 2–3
Build and certify your team
We assemble a team matched to your domain and language needs. Each annotator passes a certification test on your gold set before touching production data.
- Ongoing
Label at production scale
Work flows in daily or weekly batches. Model-assisted pre-labels handle the easy parts so people spend their time on judgement calls, with a shared channel to your team lead.
- Every batch
Deliver, measure, improve
Each batch ships with a QA report to S3, GCS, Azure Blob or your API, in COCO, YOLO, Pascal VOC, JSONL, CSV or a custom schema. Your feedback goes back into the guidelines.
What we annotate
The process is the same whatever the data. What changes is the tooling, the team’s background and the checks we run.
Images
Boxes, polygons, semantic and instance segmentation, keypoints and classification.
Image annotationVideo
Object tracking with consistent IDs, action and event tagging, temporal segmentation.
Video annotationText and NLP
Entities, intent, sentiment, relations and classification, including Swahili.
Text annotationAudio and speech
Transcription, speaker diarization, timestamps and sound events, by native speakers.
Audio transcriptionLiDAR and 3D
3D cuboids, point-cloud segmentation, sensor fusion, lane and road-edge polylines.
3D annotationDocuments and forms
Key-value extraction, table structure, layout regions and OCR correction.
Document extraction
Every label passes through four checks
Accuracy isn’t one review at the end. Errors are caught at several points, and each catch feeds back into training and guidelines.
Automated validation
The tool blocks impossible labels as they are drawn: boxes outside the frame, missing attributes, class combinations your rules forbid.
Consensus and agreement
A share of items is labeled by two or three people. We track inter-annotator agreement and IoU, and route disagreements to a reviewer.
Expert review
Senior reviewers sample each annotator’s work daily. Anyone trending below target gets targeted retraining the same week.
Gold-set audit
Hidden gold items are mixed into every batch. A batch ships only when it meets the accuracy you set, and the score is in its QA report.
Choose how you want to work with us
All three follow the six steps above. They differ in how the team is staffed and billed.
Best for one-off datasets
Managed project
A fixed scope, a fixed price per item and a delivery date. We plan and staff it; you approve the output.
- Priced per image, frame, minute or page
- Quote within 48 hours of the pilot
- Good for 10k to 5M items
Best for ongoing training
Dedicated team
A team that works only on your data, keeps its knowledge of your edge cases and scales with a month’s notice.
- Monthly per-seat pricing
- Named project lead and reviewers
- Your tools, your processes, our people
Best for model evaluation
LLM evaluation & RLHF
Trained reviewers rank, rate and write responses to evaluate and fine-tune generative models.
- Preference ranking and red-teaming
- English, Swahili and more
- Hourly or per-task pricing
Questions teams ask before starting
Can you work in our annotation tool?
Yes. Our teams work in Labelbox, CVAT, Encord, SuperAnnotate, V7, Label Studio and custom in-house tools. If you don’t have one, we can provide a platform.
How do you measure accuracy?
We agree a metric with you in step one: usually precision and recall per class, IoU for boxes and masks, word error rate for transcription, or F1 for text. Every batch’s score against hidden gold items is in its QA report.
What happens if a batch misses the target?
It doesn’t ship. We rework it at our cost, find the root cause (often a guideline gap or one annotator) and tell you what we changed.
How quickly can you scale?
We scale in phases: recruit, train, certify, then release new annotators into production. Most image and text projects can add capacity within two to three weeks; specialist domains take a little longer.
Do you use our data to train other models?
No. Your data and labels belong to you and are used only for your project. We return or delete copies at the end of the work, as agreed in your contract.
Send us a sample. We’ll show you the labels in about ten days.
Tell us about your data and your model. A project lead replies within one business day, sets up a scoping call within 48 hours and sends pilot results with accuracy and a quote.