Building Style.it: an AI that actually has taste
Most fashion AI recommends. We wanted one that styles: opinionated, agentic, and willing to say no. Here's the architecture behind it.
Recommendation engines are everywhere, and almost all of them are timid. They surface what's safe: the median of what everyone else clicked. Style.it started from the opposite premise: a stylist has a point of view, and a good one will tell you that the jacket doesn't work.
From recommender to stylist
The shift sounds small but changes the whole system. A recommender ranks items. A stylist reasons over a goal (occasion, body, weather, budget) and assembles a coherent outfit, then defends it.
“Taste is just consistency under constraints, applied confidently.”
A principle we kept coming back to
The pipeline
Behind the surface sits a three-stage content pipeline that keeps a clean product graph flowing into production:
- Scrape: pull raw catalogue from partner sources.
- Reconcile: dedupe, normalise, and enrich into a typed product graph.
- Approve: a human-in-the-loop gate before anything reaches users.
const outfit = await stylist.assemble({
occasion: "dinner",
weather: "cool",
palette: wardrobe.dominantTones(),
});What's next
We're pushing the agent to explain its choices in plain language, not because users demand it, but because a stylist who can't explain a look isn't one yet.