Distill by FusorIntelligence - Generative UI Global Hackathon: Agentic Interfaces
AI Tinkerers - San Francisco
Hackathon Showcase Finalist

Distill by FusorIntelligence

Team led by Will Stark, Parsons BFA and MIT-trained founder of X Fusor, specializing in industrial design, robotics, hardware prototyping, and autonomous systems.

1 member

Distill — turn complicated web pages or any info you looking for, into the single-purpose tool you came for.

Distill is a generative-UI workspace that goes well beyond text-based chat. You paste any URL into the CopilotKit chat shell. A Gemini 2.5 Flash agent extracts the page via Apify’s Website Content Crawler, infers the user’s job-to-be-done, picks one of six safe A2UI templates, and emits a typed JSON spec that the frontend validates and routes to local React renderers. The result is a working mini-app — a mortgage calculator with live sliders, a weighted decision matrix that re-ranks as you change weights, a persistent setup checklist, a stripped booking form, a runbook, or an approval card — saved as a card on a personal dashboard with source facts, evidence, confidence, and an expandable A2UI spec drawer. Same input shape (a URL), wildly different runtime-rendered interfaces. Cards are reusable: favorite, duplicate, refine, re-open the source.

Originality: most “generative UI” demos render markdown or static cards. Distill renders interactive, stateful tools grounded in a real source page — the agent output literally drives sliders, weights, and form state.

Technical execution: the agent does not generate executable React. It generates a typed A2UI-style JSON spec; the frontend validates the shape and dispatches to known typed renderers. That gives real interactivity without trusting agent-written code — safe-by-construction generative UI.

Interface ↔ agent output: AG-UI carries tool calls and state both ways. When the agent picks a template and emits a spec, the dashboard mounts the matching renderer; when the user interacts (favorites, refines, duplicates), state flows back to the agent for the next turn.

A2UI — safe declarative UI specs rendered by local typed React components (six templates: calculator AG-UI — agent ↔ frontend tool calls and stateful UI loop AI Tinkerers Apify Website Content Crawler — robust URL extraction from hostile pages CopilotKit CopilotKit — chat shell CopilotSidebar Gemini 2.5 Flash via Vertex AI — JTBD classification and A2UI spec generation GenUI starter base Google DeepMind LangChain LangGraph — agent orchestration (BFF on port 8133) LangSmith — optional trace visibility MCP Apps / mcp-use — Distill MCP server exposing template / spec / safety tools OpenAI approval card) checklist decision matrix form frontend tool registration runbook

Github repo

Summarizing URL...