Personal · Never launchedBack to projects
Shut down · 2026
Wardro8e: AI Fashion Discovery
An AI fashion discovery app. It learned your taste from how you used it rather than from what you typed, ranked outfits by visual similarity using CLIP embeddings, and monetised through affiliate links. Built, shipped, run with real users, and shut down.
Shut down May 2026. Codebase archived.
Talk about a project like this- Year
- 2025
- Duration
- 2025 to 2026
- Category
- Product
- Status
- Shut down · 2026
Outcome
Shipped to real users, then shut down in May 2026 when the affiliate thesis stopped holding.
Stack
React NativeExpoSupabaseFastAPIPythonCLIPHuggingFace
Scope
What I built
- Onboarding style quiz.
- Personalised swipe feed ranked on CLIP embeddings.
- Collections, saves and a trending view.
Why it shut down
The idea was Pinterest for outfits, but smarter. An app that learned your taste from what you interacted with rather than from what you typed. Style quiz, personalised feed, visual similarity over CLIP embeddings, collections, trending. It got built, it got shipped, and it got real users onto it.
What killed it was a spreadsheet, not a bug. I pulled the numbers on a better funded competitor doing discovery-led affiliate fashion and worked through their unit economics. Affiliate rates on fashion are thin enough that the model only closes at volume, and you cannot buy that volume until the model closes. There is no version of that where a small player gets out first.
So the thesis was wrong, and it was wrong in a way no amount of better engineering would fix. I shut it down in May 2026. The honest version is that I knew before I admitted it, and spent a while looking for a reason not to.
By July I had started Scri8e, and the first thing I did was ten customer discovery conversations before writing any code. That is the whole lesson: I built the hard part of Wardro8e first and checked whether the business worked far too late.
The engineering still stands up. I would build the recommendation pipeline the same way again, and the CLIP-plus-pgvector setup is the part of this page I am most willing to defend.
