Building Style.it: every fit, on your face
Most fashion AI tells you what you would probably like. Style.it renders the fit on your actual face and lets your eyes decide. Here is the architecture behind showing instead of telling.
Style.it's whole thesis fits in five words: every fit, on your face. Fashion AI has mostly meant recommendation, a system that studies clicks and tells you what you would probably like. We are building the opposite of telling. The app renders real outfits on your actual face, and judgment stays where it has always belonged, with you.
The gap we are aiming at is older than the internet but the internet made it worse. A product photo is evidence about the model wearing it. The question you actually have, what does this look like on me, is the one photo no store has ever had. Shopping online is arguing from someone else's evidence. That gap is the product.
The pipeline behind the mirror
On the surface it is a phone app, Expo and React Native. Behind it sits a Node and Express API that owns the load-bearing boring things: Supabase for data and auth, payments, storage, and the AI calls themselves, routed across OpenAI, Google GenAI, and Replicate depending on the job. A fit render is a pipeline, not a single model call, and treating providers as interchangeable parts of that pipeline is what keeps any one of them from becoming a landlord.
The multi-provider shape is not indecision. Detection, embedding, and generation are three different jobs, and no vendor is best at all three in the same quarter; the rankings reshuffle constantly. So the seams between stages are the stable part of the architecture, and the models plugged into them are the disposable part. When a better renderer ships, and one always ships, swapping it in should be a config change, not a rewrite.
Before anything can go on a face, the system has to understand clothes, and that gets its own service: a Python endpoint running YOLOS to detect garments inside images and FashionCLIP to embed them, so a piece can be found, named, compared, and searched. Generation is the part people see. Detection and understanding are the parts that decide whether generation has anything true to work with.
The third piece never appears in a demo: an offline pipeline that turns raw store inventory into structured catalogue items. You cannot try on a mess. The pipeline runs offline on purpose, far from the request path, because users get milliseconds and pipelines get hours, and confusing those two budgets is how apps get slow.
A mirror, scrolling forward
The payoff surface is the explore feed: outfits already rendered on you, browsable. Not a grid of strangers wearing things you might buy. A mirror, scrolling forward. Once the render exists, browsing and deciding collapse into the same motion, which is the entire point.
None of this gives the AI taste, and that is deliberate. The machine produces evidence; your eyes make the call. I distrust products where a model's preferences quietly replace the user's. The AI's job here is not to have opinions about you. It is to make your own opinions testable before any money moves.
What is next is shortening the loop from seeing a fit to owning it. The mirror had to come first. Everything else is walking distance from a mirror that works.