Introduction
Rosa Felder knew the floor plan of an office building before she knew what the people there were building during the day. At night she pushed a cleaning cart through corridors where whiteboards were still full of arrows. In the stairwell, during breaks, she listened to lectures on her phone.
Rosa lives in Zurich, studies information science and now leads a small data annotation team. She speaks about model quality with the precision of someone who has spent hours labelling images that were supposed to be simple.
Story of the Path into AI
Her first annotation job sounded harmless: assign categories, mark edge cases, move on. It was easier to schedule than extra cleaning shifts. Soon Rosa understood that the supposedly mechanical clicks later became part of a model’s view of the world.
The images were not the only problem. The instructions were often unclear. Was opened packaging damaged, or merely photographed? Was a reflection an object? Did a regional term count as a mistake? Rosa wrote down such cases even when nobody had asked her to. In one early project, shiny metal parts were repeatedly classified as jewellery because the training examples were too clean. That became her first glossary: examples, counterexamples, permitted doubts.
Current Work
Today Rosa organizes annotation work for images, short texts and product descriptions. She pushes for instructions that annotators can actually use and for feedback loops with developers. In one retail project, she made the team document damaged packaging, shadows and local terms separately instead of blaming individual workers for inconsistent labels.
The effect is unspectacular but valuable: less rework, fewer hidden assumptions, clearer conversations between people who prepare data and people who train models. Rosa does not believe data work is neutral. Whoever writes the rules helps decide which world a system learns to see.
Personal Advice
“Always ask who sees the edge cases,” Rosa says. For her, AI competence often begins where an example does not fit the category. Anyone entering data work should not hide those moments. There is usually more knowledge in the awkward examples than in the smooth ones.
Key Facts
Age and place: 27, Zurich.
Background: working-class family, study, night work.
Entry into AI: improved annotation instructions for ambiguous retail images.
Focus today: data quality and annotation teams.
Typical tools: annotation platforms, bias documentation, quality guidelines.
Werkstattnotiz
Rosa’s glossary contains a column called “why this was hard.” It was meant as an internal note and became the most useful part of the document. She is still testing how to make uncertainty visible without turning every annotation task into a debate nobody has time to finish.