Problem
People used to hire a consultant who walked the shops and put together outfits for an event at different budgets. Capsule does that job with an AI agent.
Constraints
- The total must never exceed the user’s budget, every time.
- Every suggested item must link to a product that can actually be bought.
- Cost and latency per request are measured, not guessed.
Architecture
- Brief → style plan (structured output) → parallel shop searches → vision check that the photo matches the item → budget solver in plain code → reranker → explanation.
- Wardrobe and garment embeddings in PostgreSQL with pgvector; background jobs on Redis.
- Every prompt traced in Langfuse with its cost and latency.
Decisions
- [ADR-001: the LLM picks the style, code does the arithmetic]
- [ADR-002: product data from affiliate feeds instead of scraping]
Results
[Add eval results here: budget adherence, live links, look quality score, cost per look.]
What I would do next
[Fill in.]