Celestial - Generative UI Global Hackathon: Agentic Interfaces
AI Tinkerers - Hyderabad
Hackathon Showcase

Celestial

Team led by an Antz ai AI Engineer skilled in Python, FastAPI, React, Docker, and building voice-agent integrations with Twilio and ElevenLabs.

1 member

Celestial — Mortgage Servicing Without Pages.

Celestial replaces the traditional 30-page mortgage servicing portal with a single conversational surface where a Gemini-powered agent materializes the exact UI a borrower needs in the moment. There are no menus, no tabs, no dashboards. The borrower speaks (or types, in English or Hindi); the agent renders a live, interactive widget — an EMI recalculator with sliders, a CIBIL-impact simulator, a hardship-eligibility quiz, a PAN/Form-16 document uploader, an interactive payoff statement, or a successor-in-interest wizard for death/divorce/transfer scenarios — directly inside the workspace canvas.

Stack — generative UI, end to end:

  • CopilotKit (@copilotkit/react-core, @copilotkit/react-ui, @copilotkit/runtime) — the agent declares seven useCopilotAction tools, each one rendering a different React widget. The chat is the router; the canvas is the result.
  • Vertex AI Gemini 2.5 Flash — picks the right action and emits structured arguments for each widget.
  • Google Document AI — parses uploaded PAN, Form 16, and salary slips into typed fields the agent reasons over.
  • Google Cloud Speech-to-Text + Text-to-Speech + Translate — bilingual EN/HI voice loop.
  • Three.js + React Three Fiber — 3D speaking avatar synced to TTS state.
  • LangChain Vertex adapter — wires Gemini into CopilotKit’s runtime.
  • Recharts — agent-rendered amortization, CIBIL recovery, and payoff charts.

Why it scores on the rubric: every borrower flow that an Indian mortgage servicer normally serves through static pages — prepayment what-ifs, RBI moratorium eligibility, KYC capture, foreclosure quotes, succession on borrower death — is rendered at runtime by the agent, inline, in the user’s language, with voice. The interface is the agent’s output.

Starting point was a vanilla create-next-app scaffold (single Next.js boilerplate page). Everything else — CopilotKit integration, all seven generative-UI actions, all six interactive widgets, bilingual voice loop, Document AI parsing, 3D avatar, Cloud Run deploy script, RBI/CIBIL/EMI domain logic — was built during the hackathon.

AI Tinkerers CopilotKit Google Cloud Document AI Google Cloud Run Google Cloud Speech-to-Text Google Cloud Storage Google Cloud Text-to-Speech Google Cloud Translate Google Cloud Vertex AI Google DeepMind LangChain Next.js 16 React 19 React Three Fiber Recharts Tailwind CSS v4 Three.js

Source code (monorepo, celestial-mortgage/)

Summarizing URL...