Beyond Embeddings: Building an AI System That Understands Outfit Compatibility
About this session
Building an AI fashion recommendation system is more than finding visually similar items. In this talk, I’ll walk through the engineering and modeling challenges behind Loom, an end-to-end outfit recommendation system that combines vision-language embeddings with structured compatibility scoring. We’ll explore how to represent fashion items, retrieve candidates, model relationships between pieces, evaluate outfit quality, and deploy the system in practice. I’ll also share what worked, what failed—including why a seemingly more sophisticated learned model performed worse—and the lessons these failures reveal about building reliable AI products beyond benchmark performance.
Speaker
Key takeaways
- Designing AI systems beyond similarity search
- Why more complex models aren't always better
- Evaluating AI products in the real world