Build Your Own ML in JS: a Model that Spots Bad UI

Even Opus — loaded with skills, commands, and design-system rules — doesn't implement a layout pixel-perfect. The root of the problem is that the agent can't see its own output: it writes code blind, with no feedback from the actual render.

In this talk we'll close that loop by hand: live on stage, we'll train a local model in JS that looks at the rendered component and tells good markup from bad, then wire it into an MCP server that renders whatever the agent produced, runs it through the model, and returns not a vague "doesn't match the design" but a concrete score the agent can use to correct itself.