A real-time AI meeting co-pilot powered by Gemini Live API
"The smartest voice in the room is always yours."
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.
- 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
Currently offline. The original Cloud Run deployment is gone along with the free trial that paid for it, so that link was returning a 500 and has been removed rather than left to rot.
The screenshots below show it running.
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.
+-------------------------------------------------------------+
| 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 |
+-------------------------------------------------------------+
| Layer | Technology |
|---|---|
| Frontend | React, Vite, Tailwind, Web Audio API |
| Backend | Python, FastAPI, asyncio WebSockets |
| AI | Google Gemini 2.5 Flash Native Audio |
| Agent SDK | Google ADK, LiveRequestQueue |
| Audio In | PCM 16kHz, ScriptProcessorNode |
| Audio Out | PCM 24kHz, AudioBufferSourceNode |
meetmind/
├── README.md
├── deploy.sh # Cloud Run deployment
├── architecture.png
├── screenshots/
│
├── bench/
│ ├── latency.py # end to end latency benchmark
│ ├── mock_server.py # fake server for checking the benchmark
│ └── test_latency.py
│
├── web/ # frontend source, builds into app/static
│ ├── package.json
│ ├── vite.config.js
│ └── src/
│ ├── App.jsx
│ ├── components/ # Board, Transcript, Composer, StatusPill
│ └── lib/ # audio.js, session.js and their tests
│
└── app/
├── main.py # FastAPI entry point, route definitions
├── requirements.txt
├── Dockerfile
├── .dockerignore
├── .env.example
├── pytest.ini
│
├── core/
│ ├── config.py # env vars, model and audio settings
│ ├── 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/ # built frontend, generated by web/
│
└── tests/
└── test_meetmind.py # pytest suite
- Python 3.11 or higher
- Google API key with Gemini Live access
- Chrome browser
git clone https://github.com/pyarchana/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 --reloadOpen http://127.0.0.1:8000 in Chrome. The built frontend is committed, so
this works without installing node.
cd web
npm install
npm run dev # vite on :5173, proxies /ws to :8000
npm test
npm run build # writes into app/staticapp/static is generated. Commit the rebuild whenever you change anything in
web/src, or CI will flag the two as out of step.
pip install -r requirements-dev.txt
pytest tests/ -vA 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.
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.
The FastAPI backend runs two concurrent asyncio tasks per WebSocket connection:
upstream_task- Consumes browser messages (audio blobs, text input, screen frames) and pushes toLiveRequestQueuedownstream_task- Iterates throughrunner.run_live()and forwards audio chunks and transcription events back to the browser
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.
Measured rather than claimed. The harness lives in bench/, and bench/README.md covers how to run it and how to read it.
Client side, fixed by the constants in web/src/lib/audio.js and pinned by the
web tests:
| What | Cost |
|---|---|
| Mic frame size | 64 ms, so no client timing is finer than that |
| Speech to playback cut | up to 192 ms, three frames of confirmation |
| Silence to end of turn | 1600 ms, twenty five frames |
Server round trip numbers need a run against a real key:
cd app && MEETMIND_TIMING=1 uvicorn main:app --port 8000python bench/latency.py --runs 20 --label "local, wifi" --out bench/results/local.jsonThose numbers are what the mic buffer was moved to 1024 samples for. It used to be 4096, which put barge in at 768 ms, longer than the whole server round trip on most networks, so the thing making interruptions feel slow was the client rather than the model.
Measured offline, no key needed, in bench/results/audio_tradeoffs.json:
| Frame size | Barge in floor | Wire | CPU per second |
|---|---|---|---|
| 1024 samples | 192 ms | 42.1 KB/s | 330 us |
| 4096 samples (today) | 768 ms | 41.8 KB/s | 271 us |
Going to 1024 sample frames cut the barge in floor four fold for 0.3 KB/s and sixty microseconds of CPU, with no server change. The cost is sensitivity: a transient now only needs 192 ms to trigger, so if false barge ins appear, SPEECH_FRAMES is the dial.
Base64 over text frames costs 33.7 percent against binary, which is 10 KB/s on a 42 KB/s stream and 68 microseconds per frame. Real, but not worth a protocol change on its own.
Built for the Gemini Live Agent Hackathon by pyarchana, 2026.