Wheerli

The problem

You see a great outfit in a photo and there is no tag, no link and no way to describe it well enough to search. Generic image search returns lookalikes you cannot buy, or stock that sold out weeks ago.

How I approached it

Split the problem into detection, retrieval and verification. Find each garment in the photo, embed the selected crop with a fashion specific model, search an index built only from retailers whose terms permit it, then live check price and stock before showing anything. If nothing verifiable is found, say so honestly rather than fill the screen.

What I built

• Native SwiftUI iOS app with a Share Extension, and a Flutter Android app with crop correction, history, saves and Story sharing
• Versioned search API on Cloudflare Workers: garment detection, FashionSigLIP embeddings in Vectorize, reranking and live result verification
• Scale to zero, HMAC authenticated embedding service on Cloud Run
• Clothing only catalogue ingestion from JSON-LD and sitemaps, gated on robots and retailer terms
• Hard monthly cost ceiling enforced by atomic spend reservations in D1
• Marketing site, App Store listings and international launch preparation

Where it stands

Search pipeline, both native apps and wheerli.com are built, with the UK App Store candidate prepared and catalogue coverage gated before launch.

What it taught me

In visual search the model is the easy part. Trust comes from the catalogue: a perfect match you cannot buy is worse than an honest empty result.