RAISE Hackathon — AI Office Platform

3D office UI + multi-agent Google Meet assistant.

Architecture

apps/web  (Next.js UI: / + /meet)  ─┐
apps/api  (Hono REST stubs)        ─┼─►  backend/ (FastAPI)  ──WS──►  worker/ (Python)
                                    ─┘                                    │
                                                       Playwright/Chrome + Google Meet
                                                                        │
                                                        Virtual audio ⇄ Gradium STT/TTS / agents

• apps/web — 3D headquarters (/) + Meet deploy screen (/meet)
• apps/api — Hono REST stubs (chat, presence, meet)
• packages/shared — Shared TypeScript contracts
• backend/ — FastAPI: session REST + /worker WebSocket
• worker/ — Meet control, audio routing, speech pipeline
• speech/ — Gradium STT/TTS, wake-word routing, orchestrator
• agents/ — Cursor cloud agent SKILL.md files
• docs/AUDIO_SETUP.md — virtual audio setup

Prerequisites

• GRADIUM_API_KEY for STT + TTS
• CURSOR_API_KEY for cloud agent invocation (see below)
• Chrome with agent Google account in CHROME_USER_DATA_DIR
• Virtual audio devices (see docs/AUDIO_SETUP.md)
• Python 3.11+, Node 18+, PortAudio (brew install portaudio on macOS)

Run

# UI (headquarters + /meet deploy screen)
npm install && npm run dev          # http://localhost:3000

# Backend
cd backend && uv sync && uv run uvicorn main:app --host 0.0.0.0 --port 8000

# Worker
cd worker && uv sync && uv run playwright install chromium
cp .env.example .env && uv run python main.py

Flow: / → click agent → info popup → Meet → /meet → share link + Start Agent

Shared Meet link

All agents dial the same Meet room via SHARED_MEET_URL (see .env.example). Backend exposes it at GET /meet-url, worker uses it as the default target, and speech/agent_context.py injects it into every agent invocation.

Cursor Cloud Agents API

Set CURSOR_API_KEY and (optionally) override CURSOR_AGENTS_API_URL. The voice loop dispatches wake-word commands via speech/cursor_cloud_client.py, which wraps the Cursor Agents REST API with retry + try/except. Failures never crash the meeting loop.

Agent skills

Agent	Folder	Wake word
Meeting router	agents/meeting-router/SKILL.md	—
Angie (Incident orchestrator)	agents/angie/SKILL.md	Angie
Nikki (Code — local mock)	agents/nikki/SKILL.md	Nikki
Olaf (Computer-Use)	agents/olaf/SKILL.md	Olaf
Speech editor	agents/speech-editor/SKILL.md	—

Speech pipeline

Meet audio → Gradium STT → wake word → meeting-router
  → Angie delegates → Nikki (raw incident analysis)
  → speech-editor (conversational script)
  → Gradium TTS → Meet audio

Core loop: speech/meeting_session.py · Meet audio bridge: worker/speech_bridge.py

Mock incident codebase

Demo customer complaints map to a local bug — no Git repo access required.

mock-incident/
  INCIDENT.md           # customer write-up
  app/checkout.py       # planted bug (empty cart → KeyError)
  app/checkout_fixed.py # reference fix

Nikki reads these files via speech/nikki_client.py. Override path with MOCK_INCIDENT_ROOT in worker/.env.

Try in local sim: ROUTER_MODE=heuristic then say "Angie, a customer says checkout crashes on an empty cart."