Free pilot batches for qualified AI teams. See how it works
Quality

How we get to 99%, and how you can check it

Our headline number is first-batch acceptance. Here is exactly what it means, how we achieve it, and what you receive with every delivery.

99%first-batch acceptance

What the number means

The share of first deliveries that clients accepted against their agreed acceptance criteria, without sending the batch back for rework. It is measured across our delivered projects.

It is a delivery measure, not a model score: it tells you how often our labels meet your standard the first time.

Five layers of quality on every project

  1. 01

    Guidelines and calibration

    Before any labelling, we write or refine guidelines with your team and run a calibration batch so every annotator labels the same way.

  2. 02

    Annotation with self-check

    Domain-trained annotators label to the guidelines and check their own work against a checklist before submitting.

  3. 03

    Independent review

    A second, senior reviewer checks every item in the pilot and a defined sample in production, correcting errors and logging why they happened.

  4. 04

    Edge-case log

    Unclear or unusual cases are logged, decided with your team, and added back into the guidelines so they are handled the same way next time.

  5. 05

    Delivery with a QA report

    Every batch ships with a report: accuracy, agreement, error types and the edge cases found.

What's in your QA report

Every batch ships with a short report, so you can judge the work yourself instead of taking our word for it.

  • Accuracy against your acceptance criteria
  • Inter-annotator agreement on the reviewed sample
  • Error types found and corrected
  • Edge cases logged and how each was decided
  • Guideline changes made during the batch
Get a QA report on a free pilot
Sample QA report Illustration: your report shows your real figures
Items deliverede.g. 2,000 images
Items independently reviewede.g. 100%
Accuracy vs. acceptance criteriae.g. 99.2%
Inter-annotator agreemente.g. 94%
Most common error typee.g. Occluded objects
Edge cases loggede.g. 37

Already have labelled data?

We can audit it, whoever labelled it, and give you the same report.

Data quality audits →

Start with a free pilot batch

We agree the batch size on a short call. If the quality isn't there, you've lost nothing.