A Multi-Agent AI System

The system worked. It just couldn't scale.

Highway's design AI was consuming up to 180,000 tokens per scan — enough to exhaust a team member's weekly budget in a few runs. I re-architected the system with a deterministic preanalysis layer that routes intent before anything reaches an LLM, reducing token usage by up to 95%. With a system that could finally run, I added an orchestration layer and a quality reviewer that evaluates craft against Nielsen heuristics, UX laws, and custom design principles.