Most facilities have already invested in CCTV. The cameras are mounted, the cabling is run, the storage is paid for. What they don't have is anyone actually watching the feeds in real time. The footage gets reviewed after an incident, sometimes, if anyone remembers to pull it. The cameras are doing maybe 20% of what they could be doing for workplace safety.
HUB closes that gap by adding the watching part. The cameras stay where they are. The model runs alongside them, watches the feeds in real time, and surfaces the things that matter to the safety team. No rip-and-replace, no new hardware in most cases. The 20% becomes closer to 100%.
What HUB is
HUB is a video analytics platform built around generative AI. The interesting part isn't that it watches cameras; lots of products do that. The interesting part is that the detectors aren't pre-baked into a fixed set. You can describe a new hazard in plain language or show the model a handful of examples, and it'll learn to spot that thing.
That's the practical difference from the previous generation of rule-based video analytics. Rule-based systems do exactly what they were programmed to do and nothing else. HUB models adapt over time, learn the quirks of your specific site, and add new detectors as your needs change. The longer they run, the better they get.
For industries where the risks are varied and don't stay still, the adaptability is what makes the difference between a system that works for six months and one that's still useful three years in.
The no-code part
Most safety teams don't have ML engineers on call. If deploying a new detector required writing Python, hiring a vendor, or commissioning a research project, almost nobody would do it. HUB lets a safety officer or supervisor train a new detector from a text prompt or a handful of example images, in minutes, no code involved.
The mechanic underneath: instead of labelling thousands of bounding boxes per detector, you describe what you want or show the model 5-10 examples. The generative AI handles the rest. A new "missing extinguisher" detector for a particular bracket type? Five photos and you're running. A "ladder left unattended" detector for a specific area? Same thing.
The practical effect is that the cost of trying a new safety detection is small enough to actually try it. Most safety AI never gets used because the deployment cost is too high. Lowering that cost to "an afternoon" changes what's possible.
Plugging into the cameras you already have
HUB connects directly to standard CCTV feeds. RTSP, ONVIF, the protocols every IP camera speaks. You don't need new cameras; you don't need to rewire anything. The model pulls the feed, runs inference, and emits events to your existing alerting and reporting tools.
The interesting capability inside the model layer is natural language reasoning plus zero-shot recognition. The model can flag situations it hasn't been explicitly trained on if it can reason about them from the language description. That extends the detection surface beyond just the things you've labelled.
From a deployment standpoint, this is the easiest argument to make to operations. You're not asking them to rip out a working camera system. You're asking them to plug a model alongside it. The operations risk is low, the value capture is high.
What "real-time" actually means here
Real-time in this context means tens of milliseconds, not seconds. The model on an edge box next to the cameras processes each frame and flags events with low enough latency that the alert can drive an actual intervention rather than a post-event review.
For high-risk work, this matters. A worker stepping into an exclusion zone gets flagged and warned within a second or two, not after the next supervisor walk-through. That's a different kind of safety system than what was possible before; the response loop closes inside the window where intervention actually prevents incidents.
Over time, the model gets better at the specific quirks of each deployment, and the safety team gets better at tuning the alert thresholds. The system isn't static; it's a tool that the safety function uses, improves, and extends.
AI safety video analytics isn't a science-fiction idea anymore. It's working at real sites today, doing real work. The teams that started early have a head start; the model is well-tuned, the workforce has adapted, the data flywheel is producing useful insights. Teams getting started now will reach the same place faster than the early adopters did, because the technology is more mature.
The longer view is human-AI collaboration in safety. The human safety team isn't going anywhere; what changes is the leverage they have. Better visibility, faster signals, richer data. The cameras you already have, doing the work they should have been doing all along.