MeetMind

A real-time AI meeting co-pilot powered by Gemini Live API

"The smartest voice in the room is always yours."

---

What is this?

You are in a meeting. Someone asks a question you should know the answer to. You do not know it. You cannot Google it without looking distracted.

MeetMind sits in a browser tab while you attend meetings. Whisper a question, get an instant answer through your earphones. Nobody knows.

---

Features

• Real-time voice AI via Gemini Live API
• Private audio responses through your speakers only
• Screen awareness - share your screen and ask what you see
• Text input for when you cannot speak out loud
• Voice Activity Detection - AI stops the moment you speak
• Live conversation transcript

---

Live demo:

https://meetmind-671715875630.us-central1.run.app

Screenshots

1. Landing Page - the homepage before starting a session.

2. Live Session Active - Connected status and Listening... with animated wave bars.

3. Voice Transcripts Working - user speech and agent response in real time.

4. Screen Share in Action - MeetMind describing your screen when asked.

5. Typing a Question - silent text input with instant AI response.

6. Screen Awareness - asking what is visible and getting a grounded answer.

---

Architecture

+-------------------------------------------------------------+
|                        Browser                              |
|                                                             |
|   Microphone  -->  Web Audio API  -->  PCM 16kHz chunks     |
|   Screen      -->  ImageCapture   -->  JPEG frames          |
|   Text Input  -->  WebSocket Client                         |
|                          |                                  |
|   Audio Output  <--  AudioContext (24kHz scheduled queue)   |
|   Transcript    <--  WebSocket Messages                     |
+--------------------------|----------------------------------+
                           |  WebSocket
                           |  audio / text / screen
                           |
+--------------------------|----------------------------------+
|                     FastAPI Server                         |
|                                                             |
|   upstream_task    browser messages --> LiveRequestQueue    |
|   downstream_task  Gemini events    --> WebSocket           |
|                                                             |
|   asyncio.gather(upstream, downstream)                      |
+--------------------------|----------------------------------+
                           |  Google ADK / StreamingMode.BIDI
                           |
+--------------------------|----------------------------------+
|                   Gemini Live API                          |
|                                                             |
|   Input:  audio/pcm 16kHz + image/jpeg + text              |
|   Output: audio/pcm 24kHz + transcriptions                 |
+-------------------------------------------------------------+

Tech Stack

Layer	Technology
Frontend	Vanilla JS, Web Audio API, Canvas
Backend	Python, FastAPI, asyncio WebSockets
AI	Google Gemini 2.0 Flash Live
Agent SDK	Google ADK, LiveRequestQueue
Audio In	PCM 16kHz, ScriptProcessorNode
Audio Out	PCM 24kHz, AudioBufferSourceNode

---

Project Structure

meetmind/
├── README.md
├── screenshots/
│
└── app/
    ├── main.py                  # FastAPI entry point, route definitions
    ├── requirements.txt
    ├── Dockerfile
    ├── pytest.ini
    │
    ├── core/
    │   ├── config.py            # env vars, audio and screen constants
    │   ├── session.py           # ADK runner, session creation, RunConfig
    │   └── pipeline.py          # upstream_task, downstream_task
    │
    ├── meetmind_agent/
    │   ├── agent.py             # Gemini Live agent, system prompt
    │   └── __init__.py
    │
    ├── static/
    │   └── index.html           # full frontend -- UI, audio, WebSocket
    │
    └── tests/
        └── test_meetmind.py     # pytest suite

---

Getting Started

Prerequisites

• Python 3.11 or higher
• Google API key with Gemini Live access
• Chrome browser

Setup

git clone https://github.com/areychana/meetmind.git
cd meetmind/app

python -m venv .venv
.venv\Scripts\activate        # Windows
# source .venv/bin/activate   # macOS / Linux

pip install -r requirements.txt

cp .env.example .env
# add your GOOGLE_API_KEY to .env

uvicorn main:app --reload

Open http://127.0.0.1:8000 in Chrome.

Running Tests

pip install pytest pytest-asyncio
pytest tests/ -v

---

How It Works

Voice Activity Detection

A ScriptProcessorNode monitors audio energy (RMS) in real time. When energy spikes across 3 consecutive frames, the AI is interrupted immediately, ensuring your voice always takes priority.

Audio Scheduling

Incoming 24kHz PCM chunks from Gemini are decoded and queued sequentially using AudioBufferSourceNode.start(when) with a rolling nextStartTime cursor. This prevents gaps and overlaps, delivering seamless playback.

Bidirectional Streaming

The FastAPI backend runs two concurrent asyncio tasks per WebSocket connection:

• upstream_task - Consumes browser messages (audio blobs, text input, screen frames) and pushes to LiveRequestQueue
• downstream_task - Iterates through runner.run_live() and forwards audio chunks and transcription events back to the browser

Screen Context

Screen frames are captured every 5 seconds at 1280x720 resolution (JPEG quality 0.4) via ImageCapture.grabFrame() and sent as realtime blobs. The agent references screen content only when you explicitly ask.

---

License

MIT

---

Built for the Gemini Live Agent Hackathon by areychana, 2026.