30 hours, public safety, working prototype.

Public-access venues — concerts, malls, transit hubs — generate enormous CCTV footage that nobody has time to watch. Radar feeds from local infrastructure go to one screen, CCTV to another, and the dispatcher correlates them by phone. By the time the situation is understood, it is already half over.

We had 30 hours to build something better. The team went to work on a Friday afternoon. We stopped the clock at 2pm Sunday with a working demo and a $10K check.

A WebSocket-backed dispatcher console. CCTV streams ingest into a YOLOv8 inference worker that emits object events. Radar feed is ingested over UDP and rendered as a heatmap on Mapbox. A correlation layer joins them on geographic proximity and surfaces incidents to a single timeline.

• CCTV ingest: per-stream worker, batched inference, frame-decimation under load.
• Radar ingest: UDP receiver, geo-projection, decay model.
• Correlation: spatial join with a configurable confidence threshold.
• Console: timeline + map + raw stream view, switchable by incident.

The hardest part of a hackathon win is not the technical one. It is making sure the demo runs at 11:47am Sunday in a noisy hotel ballroom with a flaky projector and a judge who has already seen 30 of these. We pre-recorded a fallback clip but rehearsed the live demo eight times. The live demo worked.

• ASU News article on the Malindo team and the LERS project.
• Devpost project page with the demo video and the team list.
• Devils Invent archive page on ASU's site.

Edge inference. The 2023 build ran inference centrally; a 2026 version would push YOLO to the cameras themselves and reserve the central console for correlation and dispatch. That cuts the network bill, the latency, and the privacy footprint at once. I would also re-do the correlation layer as a learned model rather than a spatial rule — the rule was the right call for 30 hours and the wrong call for production.