Your road hazard agent, right in the dash.
RoadGuard streams your back camera, detects cracks and potholes on-device, dedups the same hazard across a few seconds of frames, and asks a tool-using agent to recommend one of four actions: urgent, standard queue, needs more evidence, or human review. No install, no upload, no video leaves the car.
How it works
Three moving parts, all working together in the car.
Camera + YOLO
Your back camera feeds a YOLOv8 ONNX model running on-device via onnxruntime-web. Four hazard classes (D00, D10, D20, D40) at a few frames per second on any modern phone.
Tracker + dedup
An IoU tracker in the browser collapses the same hazard across a few frames into one submission. A short-TTL cache on the server dedups the resurface so you never pay the LLM twice for the same pothole.
Agent verdict
Perception then Research then Decision agents cite the exact maintenance procedure from the local corpus and recommend one of four actions: urgent, standard, needs more evidence, or human review.