CADRE·AI

Voice-controlled AI agent for architects, engineers, and business professionals.

Built for the Gemini Live Agent Challenge — March 2026

---

What It Does

Cadre-AI lets you talk to your building model, financial data, and the web — all through natural voice conversation. Ask "How many rooms on Level 1?" and hear the answer instantly while watching MCP tools execute in real-time. Say "Create a wall on Level 2" and see it appear in Revit.

This is the first voice-controlled BIM automation agent — nobody else has real-time Revit integration through voice.

---

Architecture

Local mode: All 3 MCP toolsets active, including Revit via named pipe. Cloud mode: Financial + Web Search active. Revit disabled (REVIT_ENABLED=false).

---

Features

Architecture & BIM (25+ tools)

• Query levels, rooms, walls, doors, windows, views, sheets
• Create walls, doors, windows, rooms
• Place views on sheets, add dimensions
• Run QA/QC validation and compliance checks
• Generate schedules and reports

Financial Intelligence (13 tools)

• Real-time stock quotes and market overview
• Technical analysis (RSI, MACD, moving averages)
• Fundamental analysis (P/E, revenue, earnings)
• Portfolio tracking and risk analysis
• News sentiment and Fear & Greed Index

Web Search & Weather

• Web search via DuckDuckGo (no API key required)
• Weather forecasts via Open-Meteo
• Building codes, material specs, industry news

---

Quick Start

Prerequisites

• Python 3.11+
• Google API Key with Gemini API enabled

Local Setup (with Revit)

# Clone
git clone https://github.com/WeberG619/cadre-ai.git
cd cadre-ai

# Install
pip install -r requirements.txt

# Configure
cp .env.template .env
# Edit .env → add your GOOGLE_API_KEY

# Generate SSL certs (required for browser mic access)
openssl req -x509 -newkey rsa:2048 -keyout key.pem -out cert.pem \
  -days 365 -nodes -subj '/CN=localhost'

# Run
python server.py
# Open https://localhost:8443

For Revit integration: install RevitMCPBridge2026 plugin in Revit 2026.

Local Setup (without Revit)

REVIT_ENABLED=false python server.py

---

Cloud Run Deployment

Quick deploy (source-based)

gcloud run deploy cadre-ai \
  --source . \
  --region us-central1 \
  --allow-unauthenticated \
  --set-env-vars "GOOGLE_API_KEY=your-key,REVIT_ENABLED=false,CADRE_MODEL=gemini-2.5-flash-native-audio-latest"

Terraform (IaC)

cd cloud/terraform
terraform init
terraform apply -var="project_id=your-project" -var="google_api_key=your-key"

---

Environment Variables

Variable	Required	Default	Description
GOOGLE_API_KEY	Yes	—	Gemini API key
CADRE_MODEL	No	gemini-2.5-flash	Model ID (gemini-2.5-flash-native-audio-latest for voice)
REVIT_ENABLED	No	true	Enable Revit MCP proxy
PORT	No	8443	Server port (Cloud Run sets this to 8080)
FINNHUB_API_KEY	No	—	Enhanced financial news
ALPHA_VANTAGE_KEY	No	—	Extended technical data
FRED_API_KEY	No	—	Federal Reserve economic data
GOOGLE_CSE_ID	No	—	Google Custom Search (falls back to DuckDuckGo)

---

Tech Stack

Component	Technology
AI Model	Gemini 2.5 Flash (Native Audio)
Agent Framework	Google ADK (Agent Development Kit)
Tool Protocol	MCP (Model Context Protocol) via stdio
Server	FastAPI + Uvicorn
Audio	WebSocket bidirectional streaming (16kHz in, 24kHz out)
BIM Bridge	Named pipes → RevitMCPBridge2026
Financial Data	yfinance, Finnhub, Alpha Vantage, FRED
Web Search	DuckDuckGo + Open-Meteo weather
Deployment	Cloud Run, Terraform
Frontend	Vanilla JS, Web Audio API, Canvas visualizer

---

Project Structure

cadre-ai/
├── cadre/
│   ├── __init__.py
│   └── agent.py            # Agent definition, MCP toolsets, system instruction
├── financial_mcp/
│   ├── __init__.py
│   └── server.py           # Financial MCP server (13 tools)
├── web_search_mcp/
│   ├── __init__.py
│   └── server.py           # Web search + weather MCP
├── revit_proxy_mcp/
│   └── server.py           # Revit named pipe proxy (25+ tools)
├── cloud/
│   └── terraform/
│       ├── main.tf          # Cloud Run + IAM
│       └── variables.tf     # Deployment variables
├── server.py               # FastAPI WebSocket server
├── voice_client.html       # Browser UI with visualizer
├── Dockerfile              # Cloud Run container
├── requirements.txt
├── .env.template
└── README.md

---

How It Works

1. Browser captures microphone audio at 48kHz, resamples to 16kHz PCM
2. WebSocket streams audio chunks to the FastAPI server
3. Google ADK Runner feeds audio into Gemini Live API (bidirectional streaming)
4. Gemini processes speech, decides to use tools, and generates audio responses
5. MCP tools execute via stdio subprocess (financial queries, web search, or Revit commands)
6. Tool results feed back to Gemini, which generates a spoken response
7. Audio response streams back through WebSocket to the browser at 24kHz
8. UI shows real-time conversation transcript and tool activity with timing

---

Hackathon

Built for the Gemini Live Agent Challenge (Google AI + Devpost, March 2026).

• Innovation/UX: Voice-first BIM automation — a genuinely new capability
• Technical: ADK + MCP + Named Pipes bridging WSL2 → Windows → Revit
• Real-world: Architects spend hours clicking through Revit menus. Voice commands collapse that to seconds.

---

Author

Weber Gouin — BIM Ops Studio

Principal / BIM Specialist. Building the bridge between AI and architecture.

---

License

MIT