I work in edge AI at Advantech, and I've already worked with Metis cards through that role — so I know what the silicon can actually do, and that's exactly why I want to push it further with this project.
The problem: traffic cameras record everything and understand nothing. A camera watching an intersection can technically see a truck losing its tarp or a bike cutting into a lane it shouldn't, but nobody's watching live, and by the time footage gets reviewed the moment's gone.
Haytham is the vision layer of a bigger physical-AI platform I'm building, and for this challenge I want to build and prove its core loop live on Metis: detection + tracking on a live feed, a small VLM reasoning over flagged events, and an agentic layer that turns what it sees into an actual citation — location, timestamp, plate read, violation description — instead of just a bounding box nobody looks at.
Hardware: x86 platform + Axelera Metis M.2 module.
What I want Wingman to build with me:
- Detection/tracking pipeline for vehicles, pedestrians, cyclists on a live/RTSP feed
- A VLM layer that looks at flagged frames and describes what's actually happening, not just what class it is
- An agentic step that drafts a citation-style report from that description — plate, location, timestamp, violation
- All iterated live on the Metis M.2, not simulated
Why edge: video shouldn't leave the site — bandwidth, latency, privacy. The reasoning and the citation drafting both have to happen where the camera is.
Demo: a live feed with real-time overlays and a running log of auto-drafted citation reports, not raw detections.
