LPRC - CrimeScience – The Weekly Review – Episode 209 Ft. Dr. Caleb Bowyer (Loss Prevention Research Council)

Episode Date: July 17, 2025

In this episode of the CrimeScience Podcast, LPRC's Founder and Executive Director, Dr. Read Hayes, sits down with Research Scientist, Dr. Caleb Bowyer. The conversation explores Dr. Bowyer's background and his role with the Affect, Detect, Connect framework in relation to loss prevention and life safety protection. They focus on the "Detect" part of the framework while leveraging the 5 Zones of Influence to highlight the possible detection signals throughout the journey to crime.

Hi everyone, and welcome to Crime Science. In this podcast, we explore the science of crime and the practical application of this science for loss prevention and asset protection practitioners, as well as other professionals. Welcome everybody to another episode of Crime Science, the podcast, and this is the latest in our weekly update series. And we're going to talk a little bit about This is the latest in our weekly update series. And we're going to talk a little bit about crime science, but also data science and how we're leveraging technology from sensors to decision platforms, all augmented by AI. And to fit all this conversation

into really the operational framework that we've got here at the LPRC. So I'm joined today by Caleb Boyer, a research scientist at the LPRC. And Caleb has many roles as a scientist here, but what we do is we strive to leverage his training and expertise and experience leverage his training and expertise and experience in data, data science, AI, and all the above to solve problems. And Caleb heads up our DTEKT initiative, and we'll talk a little bit about DTEKT Effect Connect again, like we have in prior podcasts. So welcome today, Caleb.

Thank you, Reid, for the kind introduction. It's a pleasure to be here. Excellent. So what we'll do is let's just talk a little bit about Detect and then we'll go into you. But what are we trying to get done here at LPRC? What's the framework look like? And what role does Detect and your role within Detect?

OK, so Detect is part of Detect Connect Effect. It's one of the three pillars and in detect I'm trying to detect earlier in time and for further away in space from a retail store. The further ahead in time and further away in space we can detect an offender, their movements, their coordination, the better we can disrupt their nefarious activities.

And I think you make a really, really good point, Caleb, when you say not just earlier in time, but farther out in space, right? That's what we're aiming for, to deter them, to disrupt them. If we're detecting somebody though that's coming our way, it's probably more about disruption than detection. And piggybacking off of that statement, when we talk about detect, there's many ways to detect. You have aural signatures, visual and textual, but also digital. So when we're thinking about further away in space, it could even be in their house when they're online online running their mouth on Reddit, so different subreddits, Facebook posts, X, anywhere like that online is an opportunity

to detect further away from the store. Okay, excellent. So given that, let's do this. Let's go back and now start. Well, why are you uniquely qualified, Caleb, to think about to operationalize DTEKT in all its forms? What's your background? What was your interest? And how did you get to where you are right now? Sure, so I've been at the LPRC almost two years now. I'm finishing up my PhD at the University of Florida

in the Department of Electrical and Computer Engineering. My specialty at UF is reinforcement learning, which is how do you train autonomous agents to make decisions and make predictions in real time from noisy data and disparate data sets? And so combining all that information together is partly kind of what you're trying to do here with DTEK. So I like approaching these research questions with a variety of tools in my tool bag. And can you real quickly for the listener,

you mentioned noisy data, what's noisy data? And then kind of talk about, and why is that important? And then the same thing with disparate datasets. Okay, sure. There's many sources of noise. Different noise could creep in because you could be missing data within a dataset. Noise could be corrupted in transit. There's a lot of ways noise can creep in and signals can be corrupted. Also, just text in general could be muddled. Maybe they're not saying exactly what they're intending even, so

there's like there could be noise in what their intention is versus what they're actually posting as well. For an example. OK, good and disparate data sets what's that look like? And disparate data sets what I mean by that is like kind of just distributed data. So data coming from a variety of sources so for DTEK and Zone 5 that could be many different platforms combining all that information once we've scraped it or through API or some other means. And then how do we take that data and then kind of operationalize that to feed it to a model

to then predict on that and then make like thread identifications and such. So let's kind of also stick with some basics right now. Sensors, models, there's decision platforms. Let's start. What's the sensor? What are you trying to sense? And you mentioned some of the things. Sure, yeah. I'll give a quick overview of that. But also back to my background real quick. I'll just close the loop on that. So I'm finishing up the PhD. I'm defending in summer of 2025 and then graduating. So I'll be walking across the stage

and then once I get my PhD, that'll just be a big win for the LPRC and for me personally, so I'm excited about that. Now back to the overview of DTEK. So we talked a little bit about warning indicators. We also have other considerations such as sensors or tools. How do we make those measurements? The sensors or tools allow you to make those measurements. A tool could be an OSINT platform, a sensor could be a camera in the ceiling, it could be an LPR out in the field. All of these sensors are making measurements,

and then once we have those measurements, we can build data sets from those measurements. And this data doesn't always come in, you know, periodically at set times. It could be at off times, even after a business is closed, it could come in very unusual times. So we need to be thinking all the time about how do we make sense of this data?

When is it coming in? What what do those patterns look like? And then thinking to other sensors or tools all the time, building out sensors or tools across the bowtie. Okay, excellent. So sensors are picking up what we want. You mentioned signals, characteristics, or what we're interested in signatures even where we might know who it is once we figured out. So what's an example of a digital sensor, an example of an aural of a visual textual? OK, let's start with aural first.

So think about computer vision. So it could even be a public view monitor when you walk into a store. So it picks up a face. And there's a distinction there between face detection versus facial recognition. Face detection, it just recognizes that there's a face there. No, it's not saying, oh, that's John Smith's face and he's age 24. It's just recognize there's a face.

Facial recognition would keep track of PII and other information across each visit of a person to the store. So that's an example of an RL sensor. You mentioned digital sensor, a simple example of that could be made data, cellular tracking cell phone signals. And then you have that digital information

and you're tracking it from place to place. So on the digital front if somebody's showing up it could be the apps they're using, it could be emissions from different devices and things like that. Is that some of what we're sensing out there it sounds like? Yes, another digital sensor could be geofencing. So if you set up a geofence, you can pick up every cell phone that comes in the area within a predefined area. Excellent. So we wanna know early as we can,

as far left of bang or as far away in time and space as we can get early detection, further distance to give our would be or could be victims a chance, right? A better chance. So what's an example of a text type of a signal or signature that we might want to pick up earlier in time or even during the event or even post event so we can put these people in timeout? Sure, so a text event we may want to be able to detect could be a malicious post online.

It could be someone dropping like a review on a company's page or, and then they make a threat against the company like X company, I'm gonna come shoot up the store tomorrow at this time. That you would want to know about that and be aware that that happened. And so if you trained a model to look

for these kinds of threat statements, you could detect these earlier in time. Okay, excellent. Now, aural, acoustic, things we can hear. What are the examples of some of the signals we're interested in and working with and the sensors we use to do that? This is a really interesting area. There's classical examples like gunshot detection sensors, glass breaking,

but in the last year or so we've been looking at more novel sound classes such as detecting EAS beeps going off. Also different sliders and asset protection device noises, like different tags, being able to go from these analog sounds to digital alert capabilities, that's gonna be a big win for in-store detection and so forth. Okay, so we've got digital, we've got aural or an acoustic,

you know, in acoustics we know we talk about, it could be the sounds of the people, the vehicles, the crime tools like weapons. How can AI, how can AI help us maybe better detect and classify, and I asked you what classification means. How do we do that? Sure, so just quick about classification. You have these predefined classes.

That's a dog bark. This is a person talking. So then AI, once you train the models on these kinds of sound datasets, it knows what the spectrum for that looks like. Depending on how you're doing the feature engineering, the AI model would go from that feature space to predict one of those output classes.

And that's what we're doing. We're building these individual sound detection models that are connected with anomalous or threatening scenarios like shoplifting or someone coming in the store just harassing someone. Okay, excellent. So we are, we come up with, we hypothesize, you know what, these are some sounds that we believe are associated with bad outcomes or potentially could, or there are a series of sounds and so we're going to classify or identify those. We're going to get a data set

of those sounds and presumably other sounds, but we're going to classify, we're going to label, we're going to annotate the ones that we're interested in and not the ones we're not, and then presumably the model begins to get trained. Tell us a little bit more about training an AI model. Again, it could be classifications of images, of movement, objects, sounds, and so on. So what you're talking about, there is supervised learning. So you're labeling these different chunks of data that are associated with the class of interest. With the classical example, it could be,

this is an image of a cat, label it a cat. this is an image of a cat label it a cat this is an image of a dog label of the dog but you can do the same thing with sound files so if you have a five second sound file of this is a dog bark you can then classify that's a dog bark you have another five second that's a cat meow that's a cat meowing so then if you do repeat that for hundreds of different iterations of sound recordings over time, you can train an AI model to recognize that Arley visually and other means. OK, excellent. Now you mentioned supervised learning.

Are there alternatives that are here or seem to be emerging to supervised learning? Yes, there's there's unsupervised learning techniques. There's also semi supervised learning techniques. I won't get too much into the weeds to some of these other methods, but for the most part, we can build really accurate and reliable models with supervised learning approaches.

And thinking more to the future now, with transformer models, you're able to fine tune and train on a lot fewer representations. They're able to learn from even hundreds to build like a semi-decent model. Excellent. So we're trying to make people and places safer. And we're trying to provide tools to do that.

In this case, detection tools, alert mechanisms for the management or whoever at that space or place, those people that are vulnerable in that space. We're trying to provide a version of that for somebody maybe at the corporate office or in the police law enforcement agencies headquarters. They might have a real-time crime center, for example. So talk about how are we developing this, Caleb? How are we leveraging our bow tie model, our zones, our before, during, and after events

to operationalize what we're doing, prioritize, and execute what we're trying to get done? So far today, we've been talking about Detect that which that's the left half of the bow tie all the way up to during incidents so the situation's gone kinetic. In that it's much more serious because the bad guy's in the store he's doing the deed but after the fact after if he escapes the store and he's not apprehended or the law enforcement officer or whoever doesn't prevent this crime, and he gets out, he's back in the community. We have an opportunity before he come back to the same store, or a

different store of the same chain, or maybe he's going to attack a different chain next, but with the same tactic, same strategy he did with the first bow tie, we have an opportunity to investigate those crimes. So that's the right half of the first bow tie. How do we digitally investigate that? How do we, what do we pick up? What do we learn from the first bow tie that we can then use to better

thwart later victimizations. Excellent. So, yes, you know, I've spent time, we've spent it with Dr. Stickle with Ben and with others going through and just meticulously picking apart the entire journey to harm that a bad guy, a red actor, a criminal offender might leverage as they contemplate,

as they plan, potentially plan, as they scout, acquire, travel, attack. So we're trying to very meticulously come at this. So why are we working so meticulously along the bow tie and leveraging that framework? What's that get us? What does that maybe protect us from by doing it this way? So the bow tie allows us to approach solving this problem

in an organized way and concentrate our efforts where we can have the biggest bang for our efforts, so to speak. And it allows us to be organized in our research. Excellent. I mean, that's it, right? We want to be meticulous because life safety, brand and reputation,

obviously financial performance or ruination even are at stake because of crime, especially chronic persistent crime at specific places and the fear of crime victimization that individuals have, people have, organizations have, corporations. So that results in avoidance behavior. People don't want to go there, don't go there at certain times, don't want to work there. The business doesn't want to operate there anymore. So we're trying to get ahead of these things and not miss anything by using this bow tie model. Tell me a little bit about how does DTEK fit in with

affect and with connect and are they always black and white or can it be a little blended? So that's a very good question. I'll say a little bit about the interplay between DTEK connect effect. So in terms of connect, you wouldn't have anything to connect if detect wasn't making detections. So that's the connection there is if I'm making good detections, how do I share that information into whom? Who do I pass that those threat event

information to? Now, affect, if we can kind of, if we're detecting them, maybe if we detect a violent offenders coming to our store, maybe we can steer them away at the parking lot, or maybe as they're approaching the parking lot, or maybe in zone three, before they enter the store, maybe we can intercept them there. So through a different means, that could be blinking lights, maybe some message played on a PVM.

There's a variety of means, or maybe we greet them with better customer service, kind of let them know that we know who they are and we're not playing games. Excellent. So I like that. And we want our listeners to know that. You know we try and have fun and be friendly at work, but at the end of the

day we try at the beginning of each week and each day to remember. What we're trying to do is critical vital. It is about life safety and fear, and so we want to be methodical. It doesn't mean we won't miss something, make mistakes. Everybody, every human can do that, but just to let the people know that we're really trying to methodically approach everything that we do here and have frameworks to help us stay focused. We're also trying to use these frameworks so that our partners in research, our partner practitioners, the technology solution providers, that all of us are sort of looking at the same map,

that we're looking at the same place on the map, that we're using the same terminology, creating a profession out of a career field and things like that. So what else do you think that listeners might want to know about DTEKT, about leveraging sensors and AI models? So what else do you think that listeners might want to know about Detect about leveraging sensors and AI models? So I think on the horizon for Detect is a better use of the

emerging large language models, as well as vision language models, real time. And then like we said earlier, the interplay between Detect and Connect. And so over the next year, I would like to make more progress on merging those two, Detect, Connect and Effect as well. So better merging of all three of these into strategy. I think that's critical what you're saying and what I want to do is

let everybody know LPRC members or otherwise. Please look into LPRC, what we're up to, the University of Florida Safer Places Lab. lpresearch.org is our website and it's a place you can go and learn more. Look for us on LinkedIn, both LPRC and myself, Reed Hayes, on there and that should, in addition really to the rest of our team, give you a lot more insight into what we're up to and why we're doing it. We always talk a lot about what are we trying to do, why are we trying to do it, and how are we proposing or actually doing it. What why how is important.

So lpresearch.org. So everybody please stay in touch and stay safe. Thanks for listening to the Crime Science Podcast presented by the Loss Prevention Research Council. If you enjoyed today's episode, you can find more Crime Science episodes and valuable information at lpresearch.org. The content provided in the Crime Science Podcast is for informational purposes only and is not a substitute for legal, financial, or other advice. Views expressed by guests

of the Crime Science Podcast are those of the authors and do not reflect the opinions or positions of the Loss Prevention Research Council.