DUIBL

Stilez takes a photo of an outfit and gives it an AI rating and feedback. People expect an answer in a couple of seconds, each rating has to be cheap enough to give away a lot of them, and two similar outfits should get similar scores.

I handle all of it: the React Native/TypeScript client, the Postgres backend, thirteen serverless Deno functions, the AI pipelines, payments, security and compliance, through App Store review and launch.

The first ceiling we hit was an output-token rate limit on the model provider: past a certain load, ratings started failing. Instead of making people wait in a queue, I split the pipeline across two models. Each part of a rating goes to whichever model does it well for the least money, and a failed call gets rerouted to the other one.

That doubled how many ratings we can sustain and cut the cost of each by about 30%.

Image generation was the other big cost. Many requests are near-duplicates, so I built a pgvector similarity cache with three tiers: exact match first, then embedding match, and only then generate.

The cache now serves 68% of production requests in under a second and cut image-generation spend by two-thirds.

Model providers update their models under you, and an LLM's idea of a "7/10" can drift with them. We keep a set of outfits that users have scored and check the model against it, so ratings stay calibrated when a provider ships an update.

We also release in short cycles to outside beta testers and change the app based on what they tell us.