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

I'm the CTO of Ideate, an AI-powered creative workspace for design teams. We built a content intelligence pipeline that analyzes images, generates semantic embeddings, and powers natural language search across visual libraries — using Gemini for vision and embeddings, Claude for conversational AI, and a custom orchestration layer on Convex.

This talk is about what broke, what we fixed, and what we learned shipping AI features to real creative teams:

- Why we validate every AI response against a strict schema before it touches our database — and the production incident that taught us to
- How we built adaptive batching that processes 5-15 images at a time (not the 50 I accidentally told my team) based on real-world
throughput limits
- The priority queue system that keeps interactive features fast while background analysis runs behind — our "deli line"
architecture
- Why retrieval-augmented search with reciprocal rank fusion outperformed pure vector search by 3x for creative content

No slides full of architecture diagrams you'll forget. I'll show real code, real failure modes, and the actual before/after of search results. You'll walk away with patterns you can apply to your own AI features Monday morning.