What we do
The careful part of machine learning
Models learn from examples, and someone has to make the examples. We label, review and evaluate data for machine-learning teams, with people we train and employ, under a quality system you can audit.
Four kinds of work
| Service | Typical tasks | Share |
|---|---|---|
| Image and video | Bounding boxes, polygons, segmentation, keypoints, tracking, attributes | 58% |
| Text and language | Classification, entity tagging, transcription, translation review, document extraction | 22% |
| Sensitive-content review | Policy labelling and review for platforms, in tiered rooms | 12% |
| Model evaluation | Pairwise preference, rubric scoring, fact-checking, adversarial testing | 8% |
- Rungwell Tasks time records, 2025. See the Impact Report.
How a project runs
Scope
We agree the task, the volume, the deadline and what “correct” means, with examples. We ask how the data was collected and on what basis you hold it, as our ethics charter requires.
Guidelines
Our project lead turns your instructions into guidelines our labellers can follow, with edge cases drawn from your own data, and agrees them with you.
Pilot
A paid pilot of 2,000 to 10,000 items, usually over two weeks, scored against your gold set and ours.
Calibration
Labellers, reviewers and your team label the same sample and compare. Disagreements become new guideline examples.
Production
Work runs in Nairobi, Manila or both, in shifts that cover 21 hours a day. Reviewers check a sample of every batch before it leaves.
Acceptance
You accept or return each batch. We report acceptance monthly, and we only invoice for accepted items.
Who we work for
We worked for 46 clients in 2025. We name a client only with its permission, which most do not give, so our case studies are anonymous.
| Field | Examples of work |
|---|---|
| Agriculture | Crop, weed and disease segmentation from drone and tractor cameras |
| Retail and consumer goods | Shelf and product imagery, receipts, catalogue attributes |
| Mobility and robotics | Street scenes, warehouse robots, pedestrian intent |
| Language technology | Speech and text data and evaluation in East African and Philippine languages |
| Platforms | Trust and safety review under the platform’s own policies |
| Medical and clinical imaging | Clinical photography, first-person and wearable device imagery, under restricted-room conditions |
| Public safety and security | Crowd and scene understanding, within the limits of our ethics charter |
We decline work for weapons targeting, for identifying individuals without their knowledge, and for any client that will not tell us what a system is for. See the ethics charter.
Page last reviewed 28 September 2026