“Summer in Berlin” should mean different things to different people. Styled starts with a taste graph: silhouettes, colours, brands, references, and the price I’ll actually pay. It includes the things I’d never wear.

Each new item gets a compact description in the same format, so the system can compare the item with my taste before asking a model to reason about it.

Rules remove the obvious misses: sold out, wrong category, a colour I won’t touch. Image and text embeddings narrow the rest by visual fit, brand, season, and novelty. A diversity pass keeps the shortlist from becoming twenty versions of the same jacket.

Only then does a vision model look at the remaining images, my taste graph, and the last twenty things I saved. It picks the final set and explains why. The expensive reasoning comes at the end.

I can tell within seconds whether I’d wear something. That gave me a way to test the filtering architecture on thousands of inputs, with a result I could judge myself.

It tests personal taste matching. Whether two people would get along is a different question, and still an open one.