Every AI demo looks magical. Then you ship it, and your image analysis returns "this is a photo" for 30% of uploads. Your embedding model invents categories that don't exist. Your users' searches return confidently wrong results.

When I was CTO of an AI startup, I built a content pipeline for a creative workspace used by design teams. It analyzed images, generated semantic embeddings, and powered natural-language search across visual libraries. It used Gemini for vision and embeddings, Claude for conversation, and a custom orchestration layer on Convex.

This talk covers what broke, what we fixed, and four lessons from shipping it:
- Why every AI response gets validated against a strict schema before it touches the database, and the production incident that taught us that
- Adaptive batching sized to real throughput limits, not the number you assumed
- A priority queue that keeps interactive features fast while background analysis runs behind
- Why retrieval with reciprocal rank fusion beat pure vector search for creative content

Real code, real failure modes, and real before-and-after search results. You leave with patterns you can use Monday morning.

Delivered at the AI Community Conference - New York (Microsoft, 11 Times Square), June 26, 2026. Adaptable to 20, 30, or 45 minutes. Target audience: engineers and technical leads who ship AI features to real users.