An AI stylist that builds outfits from the clothes you already own.
Users build a virtual closet from photos, and the app suggests full outfits based on weather, plans, and personal style, then shows the look on the user's own photo before they commit to wearing it.
- 01Gemini
- 02Embeddings
- 03LLM Pipeline
- 04TypeScript
I built the AI system. Gemini tags each closet photo with attributes and an embedding. A background worker turns completed sessions into preference memories, deduplicated by embedding similarity. Planning enriches the prompt with those memories, expands it into per-type descriptions, retrieves the top four closet matches per type by vector similarity, and has Gemini compose explained outfits. Reprompts route through a tool-calling agent.
AI engineering is a lot more than writing a prompt. Getting consistent output meant constraining the model with structured responses and fixed attribute values instead of hoping it phrased things the same way twice.
I learned to split a problem into steps and decide which parts belong in code and which need a model. In StyleMe, code filters and ranks closet items by attributes and embedding similarity, and the LLM only handles the parts that need judgment, like expanding the prompt and picking the final outfit. That split made the pipeline faster, cheaper, and easier to debug.
Since models are stateless, I had to design memory and context myself: saving a trace of each outfit to support rerolls and turning accepted or rejected outfits into style notes. I also learned that a simple sequential workflow with a router is often a better choice than a full agent.