How the engine is built.

Notes on the decisions underneath the product — what we chose, and what it cost.

We used Postgres full-text search instead of embeddings

The in-app assistant answers from a knowledge base. Everyone reaches for a vector index. For a corpus this size it was the wrong tool.

A shopper hands you a picture of their body to try on a jacket. Retention should be a property of the system, not a promise in a policy.

Cheapest-model-wins is the obvious way to route generations. It is also how you ship a bad try-on to a real shopper.