DUIBL

Stilez takes a photo of an outfit and gives it an AI rating and feedback. It's a consumer iOS app, so the pipeline has to be fast enough to feel instant, cheap enough to run at scale, and consistent enough that two similar outfits get similar scores.

I own everything: the React Native/TypeScript client, the Postgres backend, thirteen serverless Deno functions, the AI pipelines, payments, security, and compliance, all the way 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. Rather than queue users, I split the pipeline across two models with feature-level routing, sending each part of the rating to the model best suited (and cheapest) for it, with automatic rerouting when a call fails.

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 user-labeled regulation set and check the LLM's scores against it so ratings stay calibrated across provider updates.

We ship in iterative release cycles with external beta testers and fold their feedback straight back into the product.

• Owned a production consumer app end to end: mobile client, backend, AI, payments, compliance, launch.
• Solved real scaling and cost problems with measurable results, not just features.
• Built the feedback loops (beta cycles, calibration set) that keep an AI product reliable over time.