How Edge Computing Is Making Real-Time Loss Prevention Possible in Retail

Lucas: If you've walked through a self-checkout lane in the last year, you've probably felt that moment of hesitation—did the scanner catch everything? Retailers have been losing billions to what they call shrink, and a lot of it happens right at the point of sale. Luna: Shrink being theft, but also scanning errors, right? Not all of it is malicious. Lucas: Exactly. And the traditional approach has been to use cameras that stream video to the cloud, where a server runs object detection and flags anomalies. But here's the problem—that round-trip can take two, three seconds. By then, the customer has already walked out. Luna: And that's exactly where edge computing changes the game. Process that video right on the device, at the checkout lane, in real time. Lucas: So today I want to look at one specific deployment that I think is a great example of how edge computing is making loss prevention actually work. A large US retailer—and I won't name them because they haven't publicized it—has rolled out edge AI cameras at checkouts in about 400 stores this year. Luna: That's significant. What hardware are they using? Lucas: They're using a combination—Intel's Movidius Myriad X VPUs on the camera modules themselves, and then at the store level, a local server running NVIDIA's Jetson AGX Orin. The key insight is that the video never leaves the store. Luna: And that has privacy implications too, which is interesting. No footage stored in the cloud, no chance of a breach exposing customer faces. Lucas: Right. And here's the thing about these shows—we work hard to bring you specific, grounded examples like this, and we deliberately don't run ads on them. If you want to support that choice, there's a link: buy me a coffee dot com slash fexingo. Luna: Yeah, it keeps everything focused on the tech and the use cases. No sponsor reads interrupting the flow. Lucas: So back to that system—the latency they're seeing is under 50 milliseconds. That's the time from the scanner beep to the algorithm deciding whether the item in the bag matches what was scanned. Compare that to the cloud approach where you're looking at one to three seconds. Luna: So what exactly is the algorithm looking for? Is it just matching barcodes to items? Lucas: That's part of it, but they've trained the model to recognize about 20,000 SKUs by visual appearance alone. So if a customer scans a cheap item but bags an expensive steak, the camera sees the steak, the scanner says 'banana', and the system flags it instantly. No barcode lookup delay. Luna: That's wild. And all of that is running on the edge processor, not calling out to anything? Lucas: Correct. The Movidius chip handles the initial image processing—things like cropping, normalization, running the first lightweight neural net to detect objects. Then it sends just the relevant data—the bounding boxes, the feature vectors—to the Jetson, which runs the full classification model. The Jetson can handle about 30 frames per second, per lane. Luna: And the store gets an alert in real time. So they can intervene before the customer leaves. Lucas: Exactly. The system sends a notification to a store associate's handheld device—'Lane 4, possible item mismatch'—with a thumbnail of the incident. They can walk over and say, 'Hey, let me help you with that item.' No confrontation, no accusation. It's more about correcting the scan. Luna: I've read that some retailers are also using edge computing for shelf monitoring—like detecting when an expensive item is taken off the shelf and not scanned. Lucas: That's a related use case, but it's actually a different system. That's more about smart shelves with weight sensors and edge cameras above the aisle. The latency requirements are less strict there—you have seconds to react. But at the checkout, you need real-time. Luna: So what's the cost per lane? I imagine outfitting 400 stores isn't cheap. Lucas: They haven't disclosed exact figures, but industry estimates put it at around two to three thousand dollars per lane for the edge hardware. That's the camera with the Movidius chip, plus the local Jetson server, which can handle multiple lanes. Cloud costs are essentially zero because there's no bandwidth for video streaming. Luna: And the ROI? I'd assume shrink reduction pays for itself pretty quickly. Lucas: The retailer says they've seen a 40 percent reduction in scan-related shrink in the pilot stores. If you're a big chain losing, say, ten million dollars a year to that, you're saving four million annually. The hardware pays for itself in less than a year. Luna: Now, I wonder—what about privacy advocates? Some people might not love being recorded, even if the video stays local. Lucas: That's a fair point. The retailer has put up signs at the entrance saying 'ai assisted checkout in use'. And because the video never leaves the store, there's no central database of face data. The thumbnails are deleted after 30 days. It's a different privacy profile than, say, a cloud system where a single breach could expose footage from thousands of stores. Luna: So do you see this becoming standard in retail? Every checkout lane having an edge AI camera? Lucas: I think so, but not overnight. The big challenge is integration with existing point of sale systems. A lot of retailers run on legacy software from the 90s—IBM 4690, NCR—and getting your edge model to talk to that is messy. The retailer I'm talking about had to build custom middleware to bridge the Jetson output to their POS logs. Luna: And that's probably where the cost and complexity hide. The hardware is cheap, but the software integration is the hard part. Lucas: Exactly. But once you've done it, the model can be updated over the air. They push new SKU visual models every week. The edge server downloads the update overnight, and the next day the cameras are recognizing new products. That's the beauty of edge—you get the cloud's update velocity without the cloud's latency. Luna: So what's next for this? Could edge eventually handle fraud detection across the entire store, not just checkout? Lucas: That's the direction. Some companies are already piloting edge-based person re-identification—tracking a person through the store to detect if they're bypassing the checkout entirely. But that's more ethically complex. For now, the focus is on the transaction moment, where the privacy risk is lower and the ROI is clearest. Luna: It feels like a perfect edge use case—high stakes, low tolerance for latency, and a clear business metric. Lucas: It really is. And I think it's a model for how edge computing will quietly become infrastructure we don't think about. You won't see the camera, you won't know the model is running, but shrink will drop and prices might even come down as losses are contained. Luna: Well, and it's a reminder that the most impactful edge applications aren't always the flashy ones—they're the ones that solve a boring, expensive problem really well. Lucas: And that is exactly the kind of edge we like to talk about. I'm Lucas. Luna: And I'm Luna. See you next time.