Data Caddy - AI-powered data insights - Generative UI Global Hackathon: Agentic Interfaces
AI Tinkerers - Hong Kong
Hackathon Showcase

Data Caddy - AI-powered data insights

2 members Watch Demo

Data Caddy is the AI admin panel for vibe-coded apps. You paste a Postgres URL, get a dashboard, ask questions in plain English, and edit data without writing SQL.
The user we’re after: someone who shipped an app with Lovable, Bolt, or v0, ended up with a Supabase or Neon database, and has no idea how to look at their own data. Supabase Studio assumes SQL. Metabase assumes you’ll build the dashboard yourself. ChatGPT with a DB connection gives you markdown tables, not interactive UI. Nobody serves that user.
Working code today: connection-string onboarding, schema introspection, inferred-domain dashboard generation, plain-English chat over the data, and a preview-before-delete confirmation surface that renders the affected rows so you can deselect anything you don’t want gone. Demo runs against a seeded Postgres for an indie coffee shop (customers, subscriptions, orders, payments).
Stack:

  • Next.js + React with CopilotKit’s host SDK
  • AG-UI underneath for streaming agent events and tool calls (bundled with CopilotKit)
  • A2UI as the default rendering target: KPI tiles, charts, tables, edit forms, delete confirmations
  • Native React fallback via useCopilotAction({ render }) when A2UI can’t express a surface
  • Postgres MCP server for schema introspection, queries, and mutations
  • Claude as the agent; Gemini 2.5 Flash for the insight-batch generation pass
  • An optional Claude Skill the user runs locally in their own coding tool. It extracts the meanings the schema doesn’t carry (status enum mappings, soft-delete conventions) and sends only the structured understanding back. We never see the user’s source code.
    Generative UI matters here because no two databases look alike. Pre-building dashboards for every app shape is impossible, and asking a non-technical founder to configure a semantic layer defeats the entire point. The agent picks the shape per request, the frontend draws it.

None, we ideated and completed a very rough MVP during the duration of the hackathon (less than 3h)

A2UI AI Tinkerers Aspire House CopilotKit Google DeepMind Google Developer Group Cloud Hong Kong Google Developer Group Hong Kong Meta One Earth Alliance (OEA) Regal Hotels