Introduction
The depot smelled of dust, cardboard and cold stone. Sofia Bellini held a photograph up to the desk lamp because the scanner was drawing stripes again. On the back someone had written only “bowl, old”; nobody knew more. For an art history student that was not a minor nuisance, but the start of a long search.
Sofia lives in Vienna and now works on digital collections. She does not like calling herself an AI expert. She would rather say she builds sorting aids that force museums to write down their uncertainties more visibly.
Story of the Path into AI
Sofia had actually wanted to become a restorer. Then a student job left her sorting badly labelled archive photographs. An image model suggested an era for a ceramic fragment that sounded plausible and was still wrong: it had recognized the form but missed the local context. That kind of mistake interested her.
Between art historians and technicians Sofia often had to translate. One side feared flat interpretation; the other underestimated provenance, material and collection history. Sofia learned basic Python, computer vision and the discipline of writing error memos. In her first prototype the system only marked possible categories. The decision remained open. Every marker had an uncertainty field and a note: Which source is missing?
Current Work
Today Sofia develops workflows for small museums whose depots are full of analogue slips. In one collection of ceramic fragments, a model helped group similar shapes faster. The questions of origin and meaning stayed with curators, lab analyses and historical sources. Sofia insisted that the wrong suggestions should not be deleted. They became part of the documentation.
That saves search time in some institutions, but it changes above all the tone of digitization. Not every object suddenly receives a clean story. Some receive only a more honest gap. Sofia finds that more valuable than an interface that pretends to be complete.
Personal Advice
“Let AI sort first, then contradict it carefully,” Sofia says. Her advice is aimed especially at people in cultural professions. Rejecting the machine outright wastes a tool. Believing it too quickly loses the real competence: looking slowly, asking precisely, documenting doubt.
Key Facts
Age and place: 22, Vienna.
Background: art history, depot work, digital collections.
Entry into AI: image sorting for archive photographs with visible uncertainty.
Focus today: AI workflows for small museums and collections.
Typical tools: computer vision, metadata work, basic Python.
Werkstattnotiz
In Sofia’s notes, one fragment still says “bowl?” with a question mark and a pencil line. The model offered an elegant category; a restorer later found a fracture that shifted everything. Sofia collects such cases like warning signs. Her next test is less about hit rate than about whether mistakes remain visible enough.