Northbridge Watch

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Northbridge Watch

Northbridge Watch: quiet Agentic Vision

Northbridge Watch is a home-perception and policy agent for Avery & Morgan Quinn. It detects package, spill, and empty-fridge scenes, then chooses noop, auto_log, notify, or human_required. Calm porch frames stay quiet; spills always escalate; busy multi-package scenes ask a human.

This is a judgeable Agentic Vision loop: perception → decision → action. Visual evidence changes the branch rather than merely producing a caption.

What it does

• calm_porch → noop
• package_at_door → notify or auto_log
• busy_porch and floor_spill → human_required
• empty_fridge → notify or quiet auto_log

Each AgentTrace records OpenCV labels, confidence, metrics, bounding boxes, chosen action, reason, and side effects.

How it is built

OpenCV 5 is substantive and pinned (opencv-python-headless==5.0.0.93), not a wrapper around a vision-language model. The AnomalyDetector uses HSV masks, morphology, contours, and rectangularity for packages; HSV/Lab color features and blob solidity for spills; and bright-interior plus Canny edge-density features for an empty fridge. Runs produce annotated JPEG evidence and stamp cv2.__version__.

The PolicyEngine applies severity, quiet hours, and multi-object rules. The AWS SAM path provisions API Gateway POST /ingest, arm64 Lambda, S3 event storage, and SNS-style notifications/escalations. Local JSONL uses the same schema as S3, so the offline demo mirrors cloud behavior without AWS keys.

Fastest demo:

git clone https://github.com/dl88jing/northbridge-watch.git
cd northbridge-watch && ./demo.sh

Live demo: https://northbridge-watch.vercel.app Repository: https://github.com/dl88jing/northbridge-watch

Demo scenes are synthetic; real deployments should obtain consent, minimize retention, default to on-device perception, and never use biometric identification.

Built With

• agentic-vision
• amazon-web-services
• aws-sam
• opencv
• python
• vercel