This project provides a local, private-first AI Baby Monitor that uses video Large Language Models (LLMs) to detect safety rule violations in a video stream. It's designed for parents and caregivers seeking an additional layer of awareness, offering real-time alerts via a gentle beep when predefined rules are broken, all processed on consumer hardware.

How It Works

The system captures video frames from sources like webcams or RTSP cameras and pushes them to Redis. A separate watcher script retrieves recent frames, constructs prompts with natural language safety rules, and sends them to a local vLLM server (defaulting to Qwen2.5 VL via vLLM). The LLM analyzes the video and rules, returning structured JSON. If an alert condition is met, a beep is triggered, and a Streamlit dashboard displays the live feed and LLM reasoning.

Quick Start & Requirements

• Install/Run:
  1. git clone https://github.com/zeenolife/ai-baby-monitor.git && cd ai-baby-monitor
  2. cp .env.template .env
  3. docker compose up --build -d
  4. uv run scripts/run_watcher.py --config-file configs/living_room.yaml
  5. Access dashboard at http://localhost:8501
• Prerequisites: Docker, Docker Compose, one GPU, Python 3.12 with uv.
• Setup: Initial run downloads a ~6GB model and builds Docker images, taking several minutes.
• Docs: README

Highlighted Details

• Runs entirely locally for privacy.
• Achieves approximately 1 request per second on consumer GPUs.
• Utilizes Qwen2.5 VL model served via vLLM.
• Configurable safety rules via YAML files.
• Supports multi-room monitoring.

Maintenance & Community

• Primarily a solo hobby project by @zeenolife.
• No explicit community channels (Discord/Slack) or roadmap mentioned.

Licensing & Compatibility

• License: MIT
• Compatibility: Permissive MIT license allows for commercial use and integration with closed-source applications.

Limitations & Caveats

The project is explicitly stated as an experimental hobby tool and not a replacement for adult supervision. Playing sound from within Docker containers to the host system is noted as difficult, requiring a separate host process for the watcher.