Your second pair of eyes, powered by local video LLMs. Because, you know... it does take a village.
The AI Baby Monitor watches a video stream (webcam, RTSP camera, …) and a simple list of safety rules. If a rule is broken it issues a single gentle beep so you can quickly glance over and check on your baby.
Obviously, I'm not going to put my child in danger just for the demo, so here're videos of:
- People using smartphones, when rules say you shouldn't ❌
- Baby being safe and playful with a parent ✅
 📵 "No smartphones" rule – alert fired |
 👶 Baby walking – no alert |
| 🛡️ Private‑first | Everything runs locally; nothing ever leaves your network. |
| ⚡ Realtime-ish | Works on consumer GPUs at ~1 req/s. |
| 🦾 Video LLM | Uses the Qwen2.5 VL by default, served through vLLM. |
| 🔔 One beep alert | Deliberately minimal & quiet — just look when it beeps. |
| 🖥 Live dashboard | Streamlit viewer shows the live stream + LLM reasoning logs in real time. |
| 📝 Easy rules | "The baby shouldn’t climb out of the crib", "Baby should always be accompanied by adult" … just edit YAML. |
| 🏘️ Multi-rooms | Supports multiple rooms. Just add another YAML with instructions. |
Prerequisites • Docker + docker‑compose • One GPU • Python 3.12 with uv
# 1 — clone
$ git clone https://github.com/zeenolife/ai-baby-monitor.git && cd ai-baby-monitor
# 2 — copy .env.template into .env
$ cp .env.template .env
# 3 — build & start all services (Redis, vLLM, video streamer, Streamlit viewer)
$ docker compose up --build -d
# 4 — start the watcher on the **host**. unfortunately playing sound in Docker container and propagating to host is really painful. so no docker compose here
$ uv run scripts/run_watcher.py --config-file configs/living_room.yaml
# 5 — open the dashboard 👉 http://localhost:8501. You can also open the dashboard on your phone http://{host_network_ip}:8501Heads‑up The first run downloads the model (~6 GB), builds docker image and may take a few minutes.
Add or tweak rooms in configs/*.yaml:
name: "living_room"
camera:
uri: "0" # webcam index or RTSP URI
instructions: # natural‑language rules for the nanny model
- "The baby shouldn't do anything dangerous."
- "An adult should be in the room if the baby is awake."- Multiple rooms? Edit
docker-compose.ymland createstream_to_redisper room. Pass in new room config to streamlit viewer. Spawn newrun_watcher.pyprocess on host for new room config. - Swap the model? Set LLM_MODEL_NAME in .env. Check vLLM supported models
stream_to_redis.pycaptures frames and pushes them to Redis (short realtime & long subsampled frames queues).run_watcher.pypulls latest N frames, encodes instructions and frames into prompt and sends them to local vLLM server, receives structured JSON, writes logs & plays a beep if receivesshould_alert = True.- Streamlit live‑updates the latest frame + llm logs.
This project is NOT a replacement for adult supervision. You should NEVER leave your baby alone.
It's meant as an additional guard for situations when you inevitably get distracted for a tiny moment, and your child is doing something dangerous. Thus just a beep sound as a notification.
It’s an experimental hobby tool — use responsibly and at your own risk.
MIT © 2025 @zeenolife
Notification sound from Mixkit, used under the Mixkit Free Sound Effects License.
Videos used for demo from Pexels, used under their license