Toronto · Nairobi · Manila
Human judgment for machine learning.
Rungwell is a not-for-profit annotation company. 2,964 people we train and employ in Nairobi and Manila label images, video and text, review sensitive content and evaluate models for machine-learning teams on four continents. What the work earns goes back into their pay, their training and the next rung.
2025 in four numbers
From Rung by rung, our Impact Report for 2025. Figures at 31 December 2025.
- People who have worked at Rungwell since 2016
- 4,1181
- Items accepted by clients on first delivery
- 98.7%4
- Everyone employed by Rungwell Kenya Limited or Rungwell Philippines, Inc. in a delivery role between 14 March 2016 and 31 December 2025, counted once however many contracts they held.
- Median gross monthly earnings (base pay for scheduled shifts plus task earnings) of labellers who worked a full month, divided by the statutory minimum wage that applies to the work in Nairobi, and in Metro Manila for a 26-day month. A full month is 20 or more scheduled shifts with an average quality score of 95% or more (on site) or an average Focus Assurance score of 80 or more (Rungwell at Home). In a typical month in 2025, 58% of labellers worked a full month. Task rates vary by task.
- Promotions to senior labeller, quality reviewer, trainer, team lead or centre management since 2016. A person promoted twice is counted twice.
- Items accepted by the client at first delivery, as a share of items delivered, across all projects in 2025. Items returned by a client and corrected are counted as not accepted.
The work
Machine learning needs people to show it what things are. That work is repetitive, exacting and, done well, a skill. We do it in two centres and in labellers’ homes, with a quality system our clients can audit.
-
Image and video
Boxes, polygons, keypoints, segmentation and tracking, from retail shelves to farm drones to street scenes.
KESTREL · 4,800 items
-
Text and language
Classification, entity tagging, transcription and translation review in English, Kiswahili, Filipino and eleven more languages.
14 languages
-
Sensitive-content review
Trust and safety review for platforms, in tiered rooms with counsellors on site and the right to step away.
Tier S1–S3
-
Model evaluation
Rating, ranking and red-teaming of model outputs by trained reviewers who know when they do not know.
Pairwise · rubric
Two centres, twenty-one hours
Manila opens as Nairobi sleeps and Nairobi opens before Manila closes, so a batch sent at the end of a working day in Toronto can come back reviewed the next morning. 940 of our labellers work from home through Rungwell at Home, on the same hours.
- Manila
- 06:00–22:00 PHT
- Nairobi
- 06:00–22:00 EAT
- Toronto
- Client services, 08:30–17:30 ET
- Labelling
- 21 hours a day, Monday to Saturday; closed 19:00 to 22:00 UTC
A labeller’s story
At sea you check the same gauges every watch, and you learn that the boring check is the one that matters. Nobody told me that was a qualification. Here it is.
Arnel Gatdula, quality reviewer, Manila
Arnel spent eleven years as an engine oiler on bulk carriers before a back injury brought him home to San Mateo, Rizal. He joined the Manila centre’s third cohort in 2019 and has reviewed agricultural and retail imagery since 2022. We asked him for an object instead of a photograph. He brought his record book, and the reading glasses he bought with his first pay.
-
· People
Thank you, Faith
Faith Odera, who joined our first Nairobi cohort in 2016, has left Rungwell.
-
· Recognition
Rungwell receives the Ethical Supply Chain Award
The Linwood Responsible Business Awards named Rungwell the winner in digital services, citing our impact reporting, our wellbeing program and Focus Assurance.
-
· Statement
Ethics charter, third edition
The board has adopted a third edition of our charter. Every client must now tell us, in writing, where the data it sends us came from.
-
· Statement
An independent review of our wellbeing program
As promised in our impact report, an independent panel will review how the wellbeing program works for labellers in practice. It will report by the end of the year.
For machine-learning teams
Most projects begin with a two-week paid pilot on 2,000 to 10,000 items, with your guidelines, a calibration round and a quality report you can check against your own gold set. We quote per accepted item, so rework is ours to pay for.
- 98.7% first-pass acceptance across all projects in 20254
- ISO/IEC 27001 certified information security, restricted rooms for confidential work
- No subcontracting. Every item is labelled by a Rungwell employee
- Working with us