Process & QA
A process you can
audit.
Quality in annotation isn’t a talent; it’s a system. Here is ours, step by step, including what you receive at each stage.
- 01Day 0–2
Scope
A working call with your ML lead. We review sample data, the classes you need, edge cases you already know about, volume and timeline.
You receive
- Data & class audit
- Written proposal within 2 days
- Mutual NDA
- 02Day 2–4
Guideline
We write the labelling guideline. Every class defined, every edge case illustrated with a correct and an incorrect example. You approve it before anyone labels.
You receive
- Versioned guideline document
- Edge-case gallery
- Gold set definition
- 03Day 4–9
Pilot
A fixed-fee pilot on a real slice of your data, credited against your first production order. You receive labels and a full QA report, so the decision rests on measured results.
You receive
- Labelled pilot batch
- QA report with metrics
- Fixed per-unit quote
- 04Ongoing
Production
A dedicated, trained team works in batches. Every task passes three quality layers, and every batch ships with its own report.
You receive
- Weekly batch deliveries
- Per-batch QA report
- Shared issue log
- 05Every batch
Feedback loop
Your model’s errors become our next guideline revision. We track rework, retrain annotators and version every rule change.
You receive
- Guideline changelog
- Error-driven retraining
- Monthly review
Quality model
Three layers.
One gold set.
Every task passes annotator self-check (L1) and full peer review (L2). A QA lead audits a blind sample against the gold set (L3). Batches below target are reworked before delivery, not after you find the problem.
Batch report
PRJ-ADAS · Batch 14
Frames
2,400
Objects
33,033
Mean IoU
0.931
Rework
2.1%
| Class | Objects | IoU | Recall | Rework |
|---|---|---|---|---|
| car | 18,412 | 0.947 | 0.991 | 1.2% |
| truck | 2,106 | 0.931 | 0.986 | 2.1% |
| pedestrian | 6,930 | 0.912 | 0.978 | 3.4% |
| cyclist | 1,204 | 0.904 | 0.972 | 3.9% |
| traffic_sign | 4,381 | 0.928 | 0.989 | 1.8% |
Metrics by task
The number we are held to.
Targets are agreed per project during the pilot and written into the statement of work.
Security
Your data is not our data.
Named people only
Only annotators assigned to your project can open your data, through role-based accounts with MFA.
Locked workstations
Managed machines with USB, screenshots and personal devices disabled on the production floor.
Your cloud, your tool
We can work entirely inside your platform or VPC so files never leave your environment.
NDA & DPA
We sign your NDA or data-processing agreement before the first file is shared.
De-identified by default
Medical and personal data is handled de-identified, under GDPR- and HIPAA-aligned procedures.
Deleted at close
All working copies are deleted when the project ends, with written confirmation.
Next step
Scope your project with
an annotation lead.
A 30-minute call with the person who would run your project. We review a sample of your data, recommend an approach and send a fixed-price proposal within two business days.