There was no training on existing apps. We launched the feature in May and we didn't start training on data until a couple months ago. We used off-the-shelf GPT4o for the feature with no fine tuning. How it works is we feed in a design system with premade components, and the LLM decides how those premade components are arranged, then populates them with data and chooses how to theme them. The issue was the underlying design system we contracted to have made had components that too closely matched other screens.
Ah, so that’s the technicality. Contractors built components based on existing apps and Figma used that for training. It’s one step removed, but, like, it didn’t not happen.
