Two people see every label. Then we measure.
Quality isn’t a final check, it’s how the work is organised. Here is exactly what happens to each item, what we report to you, and how your data is protected.
The review loop
- Annotator labels. Working from a guideline with pictured edge cases, and a question log for anything unclear.
- Reviewer checks every item. A second, more senior annotator fixes or rejects. Rejections go back with a note.
- QA lead audits a sample. A random slice of each batch is re-checked independently. If it misses your target, the whole batch is reworked.
What you see in every batch report
Agreed with you during the pilot, so targets are measured the way you measure them.
- Accuracy
- Share of sampled items that pass the QA lead’s audit without changes.
- IoU for boxes and masks
- How tightly shapes match the audited version, reported as an average per class.
- Reviewer agreement
- How often annotator and reviewer agree on the class, flagging guideline gaps early.
- Rework rate
- Items sent back during review. A rising rate pauses the batch until we find out why.
Data security
Your data is the reason clients trust us. These rules apply to every project, not only on request.
NDA before data
Every team member signs a confidentiality agreement. We sign yours too.
Access on a need-to-know basis
Annotators see only the batches assigned to them, through your platform or ours.
Company-controlled devices
No personal USB drives or downloads. Work happens inside the labeling tool.
Named, trained people
No anonymous crowd. You can know exactly who has worked on your data.
Sensitive data handling
Face and licence-plate blurring, PII redaction and restricted teams on request.
Deletion on completion
Working copies are deleted at the end of a project and we confirm it in writing.