Home Case study
The class where the kids came too.
Around thirty single mothers in Hong Kong were recruited and trained into paid data-annotation work, fitting two to three focused hours a day around childcare. This is how the pilot cohort was built, what it taught us about quality, and why the same model now underpins every project we run.
Impact Innovation Lab · Cohort 4 · Backed by the SIE Fund
single mothers recruited and trained
flexible hours a day, around childcare
remote-friendly, done from home
image and video annotation, the starting specialism
Two shortages, sitting in the same city.
On one side, AI teams could not get enough well-labelled training data, and the labelling that existed was largely done overseas by workers with no connection to the market the model would serve. On the other, single mothers in Hong Kong were shut out of the job market by a scheduling problem rather than a skills problem: school runs, medical appointments and school holidays do not fit a nine-to-six contract, and part-time work that does fit is usually physical, poorly paid and offers nowhere to go.
The insight came from the community itself. As founder Anthony described it to MakerBay, a member asked why the work could not be given to someone else, and the team discovered an entire international market of online platforms where people take on data-labelling work based on their language, ability and expertise. The demand was real, remote and already flowing past Hong Kong. What was missing was a bridge: recruitment, training, management and quality assurance that a serious buyer could rely on.
“Data labelling used to be in-house, but it could be an industry on its own.”Anthony, founder, Label Less
Build the workforce, then build the pipeline.
Recruit through the people who already know them
Partner NGOs and a social welfare centre helped implement the programme and screen participants. Around thirty single mothers joined the training cohort. Going through organisations that already held the trust of these families did more for recruitment than any advertisement could have, and it meant support was in place from day one rather than bolted on later.
Train against a real spec, not a demo
Participants learned annotation software and worked through written guidelines and edge cases on practice data before touching anything live. The training was deliberately unglamorous: what counts as an occluded object, where a box ends, what to do when the guideline does not answer the question. That last skill, knowing when to flag rather than guess, is what separates an annotator from a clicker.
Design the schedule around the actual constraint
The work was structured as two to three focused hours a day, taken between childcare and housekeeping, done from home. This was not a concession to make the numbers look kind. Annotation quality degrades badly with fatigue, so short, self-chosen, high-attention sessions produce better labels than long shifts. The constraint and the quality requirement turned out to point the same way.
Put the quality system in front of the client
Every batch runs through multi-pass review and consensus checks before sign-off, and workflows are auditable so a buyer can verify quality rather than take it on trust. The social mission never appears in the acceptance criteria. Data is bought because it is good, and the impact is what buying it also achieves.
Childcare was the training programme.
A social welfare centre provided childcare so the mothers could attend the sessions. In practice the boundary blurred, and the children kept turning up in class next to their mothers. It is the kind of detail that gets edited out of a corporate case study, and it is the most important thing in this one: the programme worked because it was designed around how these families actually live, not around how a training course is supposed to look.
The other correction was about capability, not logistics. Many full-time mothers arrive believing they have fallen behind, that the working world has moved on without them. They have not. Sustained attention, patience with repetitive detail, and the discipline to follow a specification exactly are the core competencies of this job, and they are competencies these women have been exercising daily for years. The training did not install them. It named them.
“Kids would come find their mums, so it ended up that they took the class together.”Anthony, founder, Label Less
“Many full-time mums think that they are no longer up-to-date in the society, but in fact, they have their own strengths.”Anthony, founder, Label Less
The impact model is also the quality model.
Buyers reasonably assume there is a trade-off, that a social enterprise asks you to accept slightly worse data in exchange for a good story. The pilot cohort pointed the other way. A stable, trained, locally-supported team that works short high-attention sessions and is invested in staying is a better annotation workforce than an anonymous, high-churn pool paid by the click, because annotation quality is mostly a function of retention, training and care.
It also proved the local advantage is real. Annotators who live in Hong Kong read the Traditional Chinese signage, recognise the districts and estates, and hear the Cantonese in the audio the way the end user will. That knowledge cannot be bought at the bottom of an offshore price list, and it is now a specialism rather than a side effect.
What carried forward into every project
- Training against a written spec before any labels ship.
- Short, self-scheduled sessions that protect accuracy.
- Multi-pass review and consensus on every batch.
- Flagging over guessing, with an edge-case ledger for the client.
- NGO and social-worker support as part of the operation.
- Auditable workflows under signed confidentiality agreements.
Same bridge, more people crossing it.
The pilot was always meant to be a proof, not a ceiling. The programme is extending to people with disabilities, for whom remote, flexible, screen-based work removes barriers that a commute and a fixed office never will, and to ethnic minority communities in Hong Kong, whose languages and cultural knowledge are a direct asset in multilingual annotation rather than something to be worked around. The specialisms have widened too, from image and video into Cantonese and Traditional Chinese language work, audio and speech, robotics and teleoperation data, and egocentric collection for embodied AI.
The name is the promise. Label Less exists to put fewer labels on people. Single parent. Disabled. Patient. Those are the labels we are trying to remove, by labelling the data that powers AI. Every project a client sends us is another set of hours of dignified, flexible, skilled work for someone the job market wrote off, and a dataset the client can actually verify.
Cohort figures, working hours and quotations in this case study are drawn from MakerBay's account of the programme, published by the Impact Innovation Lab. Named individuals other than the founder have been left out to protect participants' privacy.
Read MakerBay's original story ↗Cantonese & Traditional Chinese annotation
Transcription, entities, code-switching, sentiment and moderation with real Hong Kong context.
VISImage & video labeling
Boxes, polygons, segmentation, keypoints and multi-frame tracking for computer vision.
ROBORobotics & teleoperation data
LiDAR, trajectories, action segments and teleoperation demonstrations for embodied AI.
Have data that needs labeling?
Tell us your formats, volume and timeline. We'll scope a pilot and quote with clear quality targets.
Request a quote →Want to back the mission?
Join the program, partner with us, or help us reach more people across Hong Kong.
Get involved →