An edge AI system that watches unattended places: homes, warehouses, substations, construction yards. It detects what's happening, tracks it, records sealed evidence, and sends an alert. Runs free, locally, no paid API keys, no special hardware.
- Full pipeline: webcam/RTSP source, detect, track, debounce, record, sign, send to backend
- A real trained detector (YOLO12 on Pascal VOC) live in the pipeline
- Evidence chain: signing, Merkle anchor, OpenTimestamps, chain of custody
- Edge resilience: store-and-forward outbox, heartbeat/silent-site detection
- Reasoning: rule-based context suppression, template and local VLM descriptions
- Platform: live dashboard, event stream, notifications, multi-tenant orgs
- Federated learning, proven in simulation
- Perception fleet: fall detection, re-ID, anomaly, audio, fire/smoke
- 90+ tests passing (
python -m pytest tests/ -v)
- Frontier VLM descriptions and SMS notifications: real, tested code paths, just waiting on a paid API key/provider account
- ONNX to TensorRT quantization: needs real edge hardware to quantize against
packagedetection class: needs better training data- C2PA-certified signing: current signing is real, just not certified
- Fleet-scale learning, OTA updates, real deployment: all need an actual deployed site
python3 -m venv .venv && source .venv/bin/activate
pip install -r requirements.txt
cp .env.example .env
# terminal 1: backend
uvicorn cloud.backend.app:app --port 8000
# terminal 2: pipeline (uses your webcam)
python -m edge.mainOpen http://localhost:8000 for the live dashboard: tracking feed, events,
severity, logs.
Run the tests (no camera needed):
python -m pytest tests/ -vTraining the detector: training/train_yolo12.py, training/train_rfdetr.py,
or training/colab_finetune.ipynb for a free-GPU bake-off.