YouTube Video
Project Description
GenUI is a moonshot-track entry that answers the hackathon’s central question — “would this have been impossible as a chatbot?” — with a categorical yes. The agent is the frontend. There are no pre-built pages, no routes, no dashboards. You type or speak a single request, and Gemini 3 Flash returns a fully-formed JSON tree of UI components — cards, stats, tabs, charts, sliders, forms, badges, modals — that the client renders in real time. Click a button or submit a form and the entire screen replaces itself with the next interface; there’s no chat scrollback because there’s no chat. The UI itself is the conversation.
Originality. This is the No Designer, No Problem track in its purest form. We didn’t wrap an LLM in a dashboard; we deleted the dashboard. Six pre-baked prompts on the landing page — trip planner, job-offer comparator, day-triage workspace, workout tracker, year-end recap, study dashboard — each generate a completely different interactive interface from one sentence in roughly ten seconds. Live Google Search grounding is on by default, so screens fill in with current real-world data and clickable citations rather than training-data fallbacks.
Technical execution. The DSL is A2UI-aligned (we researched the v0.9 spec directly) and validated on every turn so malformed responses fail loud instead of rendering broken UI. The action contract follows the AG-UI event/payload pattern; an HTML-in-iframe escape hatch covers the long tail. The whole thing is one server route, one renderer, and a deliberately lean stack — no CopilotKit/Postgres/Redis scaffolding. All six demos were tested end-to-end and consistently return 20+ rich interactive components per turn. The model doesn’t describe the answer — it renders it, and the answer is interactive, current, and different every time.