Prompt2Design VLM System

Design a system that lets a user type a natural-language prompt (e.g., “a mobile banking app home screen with a card carousel and a bottom nav”) and automatically produces a ready-to-edit Figma file whose canvas contains a high-fidelity UI design that matches the description. Your end-to-end pipeline must (1) ingest arbitrary English prompts, (2) generate a structured visual layout, and (3) output a valid .fig file (or the equivalent REST payload) whose node tree can be opened directly in Figma with correct frames, components, text styles, colors, and auto-layout settings. Treat the service as a cloud micro-service that will initially serve 100 k requests/day with <5 s p99 latency and must support incremental fine-tuning as more (prompt, design) pairs are collected. Walk through the model choices, data strategy, training loop, evaluation metrics, serving stack, and how you guarantee the exported Figma schema is always valid.