Nomad is a voice-first collaborative travel assistant built with the Google Agent Development Kit (ADK). It demonstrates bi-directional voice streaming (low latency, interruption handling) and multi-agent orchestration using Gemini 2.5 Flash.

• Backend: Python (FastAPI) + Google ADK
  • Orchestrator: NomadConcierge (gemini-live-2.5-flash-native-audio) routes tasks.
  • Subagents: Flight Specialist and Lifestyle Specialist (both use gemini-2.5-flash).
• Frontend: React + Vite + WebSockets
  • Real-time audio visualization.
  • Live transcript and system activity dashboard.

cd backend
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python app.py

Server runs on http://0.0.0.0:8000

cd frontend
npm install
npm run dev

Client runs on http://localhost:5173

Developers can customize the agent's behavior and models in backend/config.py:

• Model Selection: Change ORCHESTRATOR_MODEL or SUBAGENT_MODEL to test different Gemini versions.
• System Instructions: Modify NOMAD_INSTRUCTION to change the orchestrator's persona or FLIGHT_SPECIALIST_INSTRUCTION / LIFESTYLE_SPECIALIST_INSTRUCTION to tweak subagent behavior.
• App Name: Update APP_NAME for session tracking.

The application provides real-time visibility into system performance and agent reasoning:

• Turn TTFB (Time to First Byte): The time from when the user stops speaking (VAD detected) to when the first byte of the agent's audio response is received.
  • Note: This accounts for VAD silence duration to give a true "end-of-speech" to "start-of-response" measurement.
• Subagent Execution Time: The specific duration taken by a subagent tool (e.g., consult_flight_specialist) to process and return a result.

The right-hand panel in the frontend visualizes the agent's "brain":

• Active Subagent: Shows which specialist is currently working (e.g., "Flight Specialist").
• Subagent Args: Displays the parameters passed to the tool (e.g., {"destination": "Tokyo", "date": "May 2025"}).
• Subagent Response: Shows the raw JSON/Text output from the subagent.

Low latency, handled directly by the Orchestrator.

User: "Who are you?" Nomad: "I'm Nomad, your collaborative travel assistant! I work with a team of specialists to help you plan trips."

Higher latency, involves tool execution and multi-turn reasoning.

User: "Find me flights to Tokyo for under $1000 in May." Nomad: [Internal Thought: User needs flights -> Call Flight Specialist] > System: [Visual: Active Subagent "Flight Specialist" appears] > Flight Specialist: [Tool executes: Checks mock DB] -> Returns "ANA, $850, May 20th" Nomad: "I found a great option on ANA for $850 departing May 20th. Does that work for you?"