GenUI answers the hackathon’s central question — “would this have been impossible as a chatbot?” — with a categorical yes. It’s a moonshot-track build: the entire user-facing interface is generated at runtime by Gemini 3 Flash. The user types any request and the model returns a JSON tree of components — cards, stats, tabs, charts, sliders, forms, buttons — that a 24-node React renderer paints in real time. Click a button or submit a form and the action plus current state goes back to the model; a fresh UI comes back. There is no routing, no dashboard, no settings page; the agent is the frontend. The DSL is A2UI-aligned where it matters (researched against the v0.9 spec) with a sandboxed-HTML escape hatch for the long tail, alias coercion that quietly heals model drift (success→positive, missing action→derived), Zod validation at the boundary, and an optional Google Search grounding toggle that swaps hallucinated data for live web results with citations. We tested all six landing-page demos end-to-end via Playwright: every one returned 25+ rich interactive components in ~10 seconds — Tokyo trip planners, side-by-side job-offer comparators, day-triage dashboards, real-time workout trackers — each generated from a single prompt.