IREX turns a sentence into a camera detector

IREX unveiled StreamVLM, a beta vision-language system that turns plain-language prompts into video detectors, while its platform says it manages more than 250,000 cameras across eight countries. The product announcement says operators can add detectors without collecting datasets or training models; the guide says inference runs on a customer-network GPU node and records prompts, thresholds, and events. That governance boundary echoes Microsoft’s small-model security routing case: public-safety teams can cover long-tail conditions such as flooding or graffiti, but should treat prompt-defined surveillance as a governed pilot, not a shortcut around review.