BlackoutGuard

BlackoutGuard is an on-device, offline road-safety agent built for the Google DeepMind Gemma Edge track. It targets the failure case behind most nighttime pedestrian deaths: a dark road where headlights and a normal RGB camera see nothing. An event camera — which reports brightness changes rather than full frames — keeps seeing through that blackout, a detector boxes the road users, and the result becomes a structured incident. On the vehicle's own compute, Gemma 4 (served locally through Ollama, loopback only, no cloud or API keys) reads each incident and writes one short spoken warning, answers natural-language questions over the night's incident log, and downgrades advisories the operator overrides. A cached on-device voice line speaks the warning with no synthesis and no network. The whole loop — see, reason, act — runs locally on real recorded DSEC night data; the signature demo is unplugging the network cable on camera and watching it keep working. If Gemma is unreachable and there's no cached line, it shows a placeholder and never fabricates an advisory.