Can AI Prevent Crime? Inside Next-Gen Physical Security

Speaker A: Hi Adey, how are you doing? Thank you for joining AI for Business.

Speaker B: Yeah, my pleasure to be here. Thank you Sarah for the invitation.

Speaker A: Sure. Um, today we're talking about one of the most under innovated industries that's now being fundamentally reshaped by AI and that's physical security. Um, my guest today, Uday, is the founder CEO of Smart Sentry AI S E N T T R Y. And it's a company that's using AI agents, computer vision and real time intervention to create what they call a smarter, more affordable guard service. UD comes from IIT Delhi, an MBA from Duke, multiple startups, worked with Fortune 1 hundreds. And he's now bringing AI to um, the worlds that's still probably working with analog cameras and expensive labor codes. Guard labor. Um, the issue of security, a lot of people, the way we think about it, it's, we just imagine, oh, AIs and surveillance and things like that. Um, but it also has that kind of security and protection layer that we are using and we kind of want to kind of rely on on a day to day basis. I would love to know, um, more about first your journey. What triggered, what was the moment that you, uh, a little over six years ago, the moment that you said, hey, I want to go and build this, I want to go build this company and have my own product.

Speaker B: Yeah. Uh, I live in Mountain View. Uh, I can actually walk to one of the Google's offices. Pretty high tech place. And uh, I was shocked when my garage was actually broken into and someone stole my bikes. I just could not fathom like how with all the security cameras generally a safe environment, um, someone could commit the crime. And uh, frustrated, I actually went to Mountain View police department complaining that no, this should not ideally happen. The police officer uh, said something very interesting. He said, um, look, everyone has cameras, but who's watching them? If no one's watching them, the criminals know it and they will break in and they will steal stuff. Your goal is to make sure someone's watching. And then I'm thinking, you know, hey, at 2:00am, 3:00am I'm sleeping. How, um, am I going to be able to watch the cameras, find that someone is there and then uh, you know, call the police. All right, Then I was brainstorming with him and I said, you know what? While I'm sleeping, I can have my AI pretend to be me and yell at anyone who is outside my garage for more than five minutes.

Speaker A: That's hilarious. It really wasn't.

Speaker B: So I proposed this idea and uh, I was pretty Happy that, you know, hey, I have this background in AI. I've got four patents, and I, uh, thought I can go rig up a system like that. You know, Sarah, this probably happens only in the Bay Area, right? As I was leaving, the cop tells me like, hey, I really like your idea. Everyone comes to me with problems. No one comes to me with a solution. You know, the first guy was come up and said, I know something that can fix this problem that affects millions of people worldwide, right? And he was very intrigued. And, uh, he said, look, uh, are you taking any investments in your company? I said, oh, my God. Like, you know, Silicon Valley, right? This is why it is Silicon Valley, right?

Speaker A: Yeah.

Speaker B: Even a police officer thinks of, uh, contributing to innovation, investing in companies. And, uh, I'm not kidding. The same day I went and incorporated my company, and I brought him in as my first investor and advisor.

Speaker A: No way.

Speaker B: Uh, that's how we got started. Uh, amazing origin story, right?

Speaker A: Yeah, it's amazing. Tell me about that piece. Does it really kind of. It would be fun actually to have it, uh, impersonate you and then kind of, hey, what are you doing at my door? Go, that.

Speaker B: No, that's exactly what we've done. Um, we have some patents also around this. Um, and we started off trying to understand what is this space, right? Started very like, uh, at a home level where you have like maybe couple of cameras, couple of places, or entrances to God. And then understood very quickly that it's a problem that affects commercial, uh, buildings, you know, restaurants or office buildings. A lot more, A lot, lot more than residential, right? Because at least at homes, people are there, uh, at night, the chances of breaking in are a little bit lower. But think about unattended properties, right? Car dealerships, uh, office buildings with expensive equipment inside, even, like dental offices, right? Where they have stuff inside construction sites. So these are broken into a lot more frequently. From our learnings of what, um, we achieved at the home level, we decided to start working with these businesses. And as we were working, we actually got approached by two large companies, Nvidia on one side, saying that, hey, we have got this office campus. There are too many signals coming from the camera saying something happened, something happened. There are a lot of signals coming from the doors saying that the door is open, someone has forced the door open. And my agents in my security operations center, they're getting overwhelmed by the number of alarms coming and they're not able to respond fast enough. If there's a real security incident that's happening, uh, you'd be shocked. To know in many, uh, of these large enterprises, uh, because of this kind of, uh, fatigue due to the false alarms that keep coming in from camera systems and other systems, they can get to real events only in about 10 minutes. I mean, when you look at like all these mass shootings and other problems, uh, and again, people keep thinking, how could someone break into the museum, right, the Louvre museum, and steal all these things and get away? This is exactly the problem. Sometimes the cameras are not working. If the cameras are working, it's such a tedious job to keep looking at the cameras. 24 by 7, right? If you have security guards, I mean, how many guards can you deploy and where can you deploy? Right? You cannot guard every square inch of your property 24 by 7 with either security guards or with cameras. That is where AI becomes our savior to protect our community. Right?

Speaker A: Half of the Hollywood movies that are based on security guards falling asleep at the cameras or just looking somewhere else. I'm going to be, you know, I feel bad for them now.

Speaker B: Exactly right? And, um, again, the way I describe this, right, is, um, the life of a security guard is boring 99% of the time because nothing bad is happening, right? And now to kill time, they're either their mobile phone, TikTok, talking to a friend, you know, the 1% of the cases, or maybe even less. It gets too interesting. Now you have to deal with a real threat, you know, and the person can be armed. Unfortunately, in many of these cases, including, uh, the Hollywood movies, the first person to be shot is a security guard. We have to think about it from their perspective also, right? How can you protect security guards from these incidents? I mean, think about the solution to all of this is having a layer in between, just like you described. A copy, a, uh, digital copy of me, or a security guard who is monitoring these cameras, the doors, the property, 24 by 7, looking for anomalies. Hey, this is suspicious. Why is this person here at 2am when nobody should be there? Why is this car coming, you know, the reverse, not in the regular way that I've always seen. Right. In fact, this car doesn't have any license plates or it's like a business area. Why is there a group, uh, of four people walking in? Right? Especially with backpacks, with masks and potentially burglary tools, right? You see anything suspicious, you want this AI layer, the AI agent to come on and almost pretend to be a security guard and warn them, hey, uh, what you're doing is violating our security policies and protocols. I'm warning you to leave before you commit any crime, and we have to come and prosecute, arrest you and prosecute, right? So that layer is very, very important. Now, what else can we do? Right? So think about it like, we have a customer in Jamaica, uh, where their biggest problem is exactly this. Even if their security guard is armed, on the other side, there might be a gang of four people, all armed, and they can shoot at this security guard. And they lost a security guard to one of these, uh, incidents. So for them, all they're asking is, can you please equip my guard with information so that before they respond, they know what happened. And in fact, better still, if your AI can take care of the problem on its own by pretending to be a security guard, say, yelling in a very loud voice, hey, you, get out of here. And activating loud sirens and loud sounds, right? Lights, maybe flashing lights, making it clear that, hey, if you are caught in the act here, you will be prosecuted by police and you will be behind bars. And just remind them this is not something they should do, so that that itself should take care of most of the cases. And in the exception, like now, that this person is still not listening to you, still trying to commit a crime, you have bought very, very valuable time for either security guards or for law enforcement to respond very intelligently. Right. They know who is committing the crime, what crime is being committed, are they armed or unarmed, and how can we tackle them, and how can we. If the crime happens, uh, we know how to identify them, catch them, and recover the losses. Right. So you can see how this can be the best solution to protect properties from crime and hopefully deter crime from happening in the first place.

Speaker A: One example just came to my mind when I was growing up. Um, my family, we were living in a community. You see these people that are at these, um, kiosks, and they. Yeah, you know, that sort of. Yeah. And they actually, um, um, he lost his life to it, and it was very tragic here.

Speaker B: And this is what we want to prevent, you know?

Speaker A: Yeah, yeah, that would be. That is something that, um, I saw for your son, and I think that can bring value. Ah, a lot. A question that I have is, where do you. How do you. So there is false, uh, negatives and, um, all of that. So how. How do you deal with the false thing, especially when you're bringing AI in the loop? Um, how would you make sure that the data that AI, your AI model, is being trained on to detect that there's a problem? They are actually, you know, the good data, obviously. And what is the accuracy rate of it that you're getting.

Speaker B: Yeah. So now coming into the world of AI, right, from security, the problem is that of uh, exactly that, like you know what is true, what is false, right. And what is our model, uh, truly predicting? And because AI is not a perfect science, right? So it's probabilistic, so you always have some level of error. We have to now make conscious choices on whether you're going to stack, um, heavily towards reducing false negatives or reducing false positives. Right? Let's understand this a little better. If you reduce a lot of false positives, um, that, that makes a lot of people happy. Why? Because, uh, you every alert that you process, right? With a remote guard, it's at least a dollar per event. So think about having 10 false positives a day. Now you're looking at $300 a month of just processing cost, right? Lot of incentive is to reduce the false positives. Now look at the other side. If you lean your algorithm towards reducing the false positives, which is like a guaranteed savings of let's say $300 a month, uh, you know, 3600 a uh, year. Right. On the other side, the false negative can be catastrophic, right? It can be a loss of a life, it can be loss of uh, property, $20,000, $30,000, $50,000. Right. But the probability of an event happening is only once in probably 10, uh, years for a residence and once in three years for a home. Right. So what we have seen is typically a lot of companies in the space, they try to reduce false positives which they can measure on a day to day basis and hope for the best that nothing bad happens. Right? So ah, with our AI, we very consciously chose to reduce the false negatives a lot. So we want to keep that number less than one in thousand, as in if thousand events happen. Right. We won't allow more than one where we cannot detect a human. Right. In that area, I'm not talking about the crime, but just a presence of a human should not be missed. So that is our threshold.

Speaker A: Could we kind of, in this space, um, kind of give me an example of what is a false positive and what is a false negative. How does it look like in um, in this space?

Speaker B: Sure. A false positive is basically where, um, the algorithm thinks there is a human, but there is actually no human. Uh, and extending it further, there is a human, but that human is not committing any crime, but is just there by mistake. Right. So unfortunately there are too many errors. Even in the first category where there are um, statues, uh, there could Be posters, there can be a television or shadows, um, from trees that look like humans. If you look closely, they all can say, oh yeah, that does look like a human. Right. So especially at night, uh, you can imagine the lighting is not very clear. So any human being. Right. Uh, is detected through shadows and the outlines. And there are many shapes that look like a person. A person's head, person's hand, person's leg, right? Think about branches of, uh, trees. They look like hands or legs, right? So the AI algorithms trying to be very, um, sensitive to avoiding false negatives, which is missing people. They try to show everything as a person. And that's when you get overwhelming number of alerts saying that there's a person, there's a person. When you go and see it's a branch, it's a shadow, it's um, a, uh, you know, it's a twig or it's a poster. And it annoys the hell out of all these security, uh, guards. And they start losing trust in the system. They think, you know, oh, this algorithm doesn't work. This product does not work. And it gets worse. Now they stop paying attention to even the real ones, right? So that is what a, uh, false, uh, positive does again, extending it. You can have uh, a common citizen, you know, who may just wander into the property, right? And you know, may realize that this is a protected area and may leave. But if you think it's going to be, if it's a criminal now, you may just overreact, right? So that's also a false positive. You need to understand the behavior. You need to know this is not a crime. And then once you persuade them, they may leave. So all those incidents where a, uh, crime has not occurred, right? But your system says there's a crime happening here, is a false positive. And it can be very expensive for companies to process these false positives with a human. Anytime a human is involved and they need to apply this judgment, it can be between a dollar per event to if there is a visit by a person, right? Like think about calling a police officer. It can be $500, it can be thousand dollars. Looking at even the opportunity cost, that officer could be responding to a real emergency instead of showing up at a place where nothing is happening, right? That is the cost of a false positive. A false negative is where, um, someone, and we have seen crazy examples of a person dressed in complete black, okay, and walking in and merging into the shadows and um, sneaking in and committing a crime. It's hard for even a person, a human, who is looking at the camera to be able to detect them. Uh, you know, and makes it much harder for an AI algorithm to detect that. Right. Because they don't have the data set to show that this can happen or has happened in the past. Uh, AI learns from patterns. And unfortunately, if, uh, the criminals keep coming up with clever ways of camouflaging, uh, and then getting in, it makes it very hard to train the system. This is where the false negatives have a higher potential to create, uh, damage. And there are a lot of strategies to reducing false negatives. But again, the cost of a false negative is much, much higher. Right. Uh, it could be loss of property, vandalism, it could be loss of life. So, um, it's tough to put a number on what the value of or the damages due to false negative is. You need to at least, uh, estimate. And then you have to create a system that can reduce the false negatives while also reducing number of false positives so that the overall cost of managing the system is lower.

Speaker A: So kind of what I'm hearing is that you want to create a balance between, okay, the false positives. We don't like them, but they're going to cost us damage. Not just even the monetary value less than the false negatives. So you've trained your models in a way that they'll detect the false negatives more. And I think then you're torn between, okay, we want the false negatives, but we don't want to put a lot of alarm to security. You know, numb them about the alarms and to get them annoyed so that they won't, they will stop responding. How do you solve this problem? Predominantly? It seems like a challenging problem to solve in this space.

Speaker B: Yeah. So this is the secret sauce. And we actually have some patents around that on the process and some of it is actually proprietary built into. Our algorithms are not even shared. But, uh, just to give you an idea, right? Um, everything that you are doing today, you know, can be defeated because there are people who are trying to break in, right? And they keep trying and testing the defenses. Um, our best good analogy I'll give you is, um, that of, uh, protecting ourselves from like, you know, the viruses or bacteria, right? So they mutate and they keep testing our defenses and knowing very quickly this kind of attack has happened. Learning from it and then, uh, deploying a solution or like, uh, a preventative mechanism across your whole portfolio, across all the humanity, is how we protect our results. And unfortunately, security, uh, uh, where you're protecting properties and people is very similar. What uh, works today, will not work tomorrow, unless you keep the whole thing fresh. Right? So one of our approaches here, Sarah, is exactly that. We keep very close attention, you know, on, um, what is happening in the world. For example, this thing that they attempted at Lourdes is pretty clever, right? Uh, I mean, dressing as a construction crew and breaking in is pretty old, but bringing in a ladder like that and then getting inside, bringing these angle grinders, it's all new, right? And a lot of museums have now understood, oh, this is a way people can break in, but then you have to break it down into elements. Um, at this particular location, uh, did you expect construction crew? If you did not, why are these people there? Right. Uh, a crane. What is it doing during operating hours? Right. So there are ways to detect anomalies from the prior patterns, which can be a very good clue to what is unusual. And then you don't need to just rely on AI and we are very proud of our smart God offering where AI is doing 99% of the job, but the 1% of the job still involves human in the loop. Right? We actually have a remote guard. Look at it and seeing. Okay, wait a minute. This does not look right. Or even dispatching a person to go there and check it out, right? To see why is this crane here? Why are these people here at this time? That is how we have a very comprehensive solution, uh, that not only learns from the prior patterns, but also involves humans to actually intervene, live, and then fix problems from happening in the first place, and then learn from that also so that that pattern is not repeated in any other place, anywhere else.

Speaker A: Got it. So what I'm hearing is that the way that you want to kind of, um, solve this dilemma is that there is a human in the loop. So I'm assuming that you have a large or growing, at least call center sort of organization back there that, okay, these alarms are coming in, but before we're going to trigger the police and, you know, make the alarm go live, there is a human sitting over here. And as opposed to a regular guard sitting at the monitors, just not knowing where to look, is it something happening or not? They will look only when something is happening, and they will probably. You'll have a good estimate of how m. What is the capacity of an average person and average guard that can take care of, like, how many alarms, you know, per a range of time? I'm just going, my, my brain is going like, and too analytical right now. But I'm assuming this is the way the bringing the human in the Loop, but at the right place at the right time, not just over there sitting idle. That's probably, I'm assuming, the way that you're solving the false positive, false negative challenge. Uh, this here.

Speaker B: Absolutely. So there are 2, 3 sources of information that make it more powerful, right? One is the whole body of knowledge of humanity on what kinds of crimes happen. Right? Uh, and how do people typically break in, how do they dress up, what do they wear, uh, what do they use? So that is all part of our algorithm already. The second thing is, what is normal for this particular place and time? Uh, if it's a school, and if it is 2am you don't expect anyone, any human to be there or much activity to be there. If there is change in that activity, that is worth examining, right? So those are the sources of, uh, data. Then you come to a human to examine those anomalies. Right? And definitely before any law enforcement or a human is dispatched, we want that to be examined by a person, by a human. So that is basically how we have designed a solution that is very effective, both from a prevention, um, point of view, crime prevention point of view, and also from a cost point of view. And this system only gets better over time, right? Uh, because the two sources learning from the past crimes and second, learning what is normal or abnormal for the place, those two help reduce the incidents tremendously. The third thing we do, right, again, before we get the, um, human in the loo, the AI agent also tries to pretend to be a human and probes the situation, asks questions, provocative questions, right? And kind of throws the person off balance, you know, so that, um, it can scare people away. It can just take care of the situation, right? And then, only then it goes to a human, right? Even, ah, in the remote, the call center kind uh, of thing, what we call as a security operation center. So even for that person, we want to reduce the load by using the AI agent to the best we can. And when that fails, that's when we want this person to come in. And when that person is not able to take care of it, then you bring in a person on site, like a law enforcement or an armed guard dispatch.

Speaker A: You reminded me of a funny, um, memory that I have. I'm like, okay, maybe that was one of your systems. I don't know. It was a couple of months back, and I was traveling in the east coast and I was. Me and my friend were walking. It was late in the. Late in the night, I think it was about. It was after. It was dark and everywhere was closed and we were walking. I was like, oh, okay, this. Is. There is this, um, Mercedes, uh, dealership. But all the cars at the gate was open and the cars were out. Um, you know, they were not, like in a covered area. It was like, okay, let's just walk in here. Here. And, uh, we just went inside and we were just walking for ourselves. Like, oh, okay, this model is like this and that. It was just hanging out. And then all of a sudden, from the. What is it called? The speakers, somebody started saying. It's like, as if it was a person. I don't know. I mean, maybe it was AI it was telling us, hey, you know something, as if it was reading it. You need to, uh, say, we're saying it politely. They need to get out of here or we're going to call the police on you. I was like, is it talking to us? How is it seeing us? So we was trying to, uh. At first we was like, let's see where the camera. Is it really us? Is it that we shouldn't be in here because the gate was open? It seems. Okay. So I was like, we were still walking and it kept getting a little bit more aggressive. I was like, okay, we need to get out of here. So, finally got out of there. Tetsuo, speaking about the, um, your. The patent that you were talking about. So is the patent lies. Is it the model because it's six years that you're working on it, um, or is it what is kind of the. The. Where does the value out of the kind of the. The patent lies in?

Speaker B: Sure. Uh, I just want to point out one more thing about, um, your experience. So what does not make it natural is that there is nothing specific about you when you are walking in. Right. The AI, our AI agent is sufficiently advanced now to be able to point you out very clearly, Right. As though it's a human who is talking to you. For example, in your case, if it's you and your friend, right, we'll say, uh, the two, uh, ladies, you know, who are walking in our area, we are talking to you, or I'm talking

Speaker A: to you wearing red with a black over the red hat.

Speaker B: Literally, if you are wearing like a red, uh, dress, right? Saying that, hey, ma', am, you in the red dress, I'm literally talking to you. Right? So this makes it so much more convincing that, uh, it's actually human and you are the one who is being referred to and this is what you are violating, um, as a security policy or in some cases, I mean, flip it to the other Side, Right. Sarah, think about making this more of a welcoming thing. Maybe, like, you know, you are allowed to be there, right? So, um, uh, you can have a message saying that, uh, look, you know, you're allowed to walk around, but please don't, uh, open anything, right? So that could be a way the same system can be used, right? So it's not just, um, a security guard who is preventing crime, but it's also think about, like, concierge or, uh, a attendant, right, who is, uh, letting you walk around, but in a safe manner, right? Uh, so that you are following their protocols. And then when it's violated, then it gets into the security side, right? So there is an advance that is happening here in terms of experience so that people know when they're violating, when they're not violating, and being told very specifically so that that behavior can be corrected. Right? We don't want people to assume that, you know, am I doing something wrong or am I not? Right? And then, because that ambiguity creates an opportunity for a criminal to pretend to be someone who is there by mistake. And they're testing the defenses to see that if nothing's happening, they're going to go and commit the crime. So we need to be very clear to them. Like, I see you. This is what you're doing. This is not allowed or appropriate. And I'm giving you an opportunity to leave before you, uh, commit a real crime or before I, uh, send someone over to escort you out. So that makes it extremely clear. And this was not possible about a year ago, Sarah. Uh, but with the latest advances in all the AI agents, the LLMs, and the, uh, multimodal communication, it is now possible. And we are proud to say that we are one of the first to deploy it in the physical security industry.

Speaker A: This is fascinating. I was going to ask you because your company is over six years old. And, uh, I was like, we know. At least most of us know of AI just in the past, like maybe two, three years, um, that you mentioned, not right now, that is being effective for your business in the past. Is the IP the model that you're building, or are you using a foundational. One of the foundational models that are out there and kind of training it with specific data or where the IP lies in.

Speaker B: Just out of curiosity, it's a combination of all. Um, there's, like, data sets. We have processed billions of images from security cameras to, uh, understand, uh, where there is a person, what kind of behaviors are happening to get that at the AI level. Right. Then how do you apply that to the security uh, industry. How does crime happen? How do you intervene to prevent crime from happening? So we have patents around that on the process patterns. Right then on top of it we have a lot of proprietary uh, processes which are nothing to do with technology. But how we combine all these things together, involving human in the loop and retraining our systems regularly, uh, and you're right, we have seen tremendous growth in this space. So the way we started solving these problems many uh, years ago is not how we solve them today. Again as a Silicon Valley company we keep a very close eye on what's changing and we are not afraid to change, we are not afraid to change our uh, approach and then incorporate the latest to make our solution overall work for the customers. So that's where the secret sauce is. Uh actually it's a team, our team is so good, I'm very proud of them um, that they're always willing to see what is happening, what is possible and how can we bring that into our solution so that it's the best it can be.

Speaker A: So it's basically what I'm hearing is that big part of it is in the analysis of whether what activity or what is what, what AI is seeing, the computer vision part of it and also analyzing that, whether or not that is criminal. I'm assuming that is the differentiator because um, I don't know if that is readily available already out there. So that is something that you're doing right? Did I get that right?

Speaker B: Yeah, yeah. And also as I mentioned we are backed by people uh in the law enforcement. Uh we have those with military uh backgrounds, uh, who are our advisors? Uh one of our investors is uh, Brian Tuscan, He's a well known authority uh on uh, basically how do you manage security for large enterprises. He used to be a chief security officer of ServiceNow and prior to that for nearly 20 years at Microsoft. He uh, was a law enforcement official, he was uh, a detective, he uh, uh, uh trained under FBI. So basically we have people who know how to think, um, basically from a crime prevention perspective as a discipline, uh, not just as a hobby. And then you have also people um, who are good at breaking uh, things like red teaming, uh that's a very popular term, uh primarily used in the cybersecurity space. So we learn a lot from cybersecurity to see what is happening in that world and how are um, enterprises protecting themselves from these uh, uh, always changing and always innovative threats that they are experiencing and then looking for the analogies in the physical world. Saying, okay, how can they potentially break in, uh, to a physical facility, right, like office, uh, building, like doors and all. And what can we do to anticipate and harden our defenses? Right. So we learn a lot from those two, from the cybersecurity industry and from the best practices from law enforcement and defense. Uh, there's this whole concept of cpt, uh, crime prevention through environmental design, uh, which is the first layer of defense before any of other technologies kick in. Right? Um, so we think about security in a very comprehensive way, not just, uh, dump some AI, whatever state it is in, and hope for the best. Instead, we look at the problem. What is the problem? What are the things, uh, that are frustrating enterprises, businesses, homes, our communities? And what is the right solution? And AI will be an ingredient, probably the primary ingredient and the core ingredient. But that's not enough. You need to have lot of other layers of, um, defenses. And this is where we get advice from experts in the area to make sure our solution is meeting the needs of the customers.

Speaker A: So is it a model that you've created from scratch, or is it that you're using one of the foundational models as a base and kind of building and training on that or like inferencing on that and stuff?

Speaker B: So it's a very interesting problem. Uh, I'll explain slightly differently. A, uh, lot of data sets that you find out there, right? They're all designed for, um, different purposes. They're not designed for security industry. Right? So the videos that you see on YouTube, they're all like, you know, like the ones we are doing here where people are looking into the cameras. The lighting is perfect, right? You can see people very clearly. Uh, uh, unfortunately, the feeds coming from security cameras are completely different. They're like poor lighting conditions. The angle is completely different. Uh, all you see of a person is probably top of their heads. So it's a completely different data set that you need to train your algorithm on. So in some cases, we had some success with foundational models. In many cases, we had to rebuild our, uh, algorithm based on our own data sets. And also think about the LLMs, right? Um, there's a big effort to make the LLMs, uh, very polite, very respectful, uh, accommodating to the point where they're becoming obsequious. Right. Too flattering, uh, agreeing with whatever you say. Now, how do you make this LLM yell at people saying that, hey, get out of my place, right? So we had to redo lot of things because our use case is very different from that of, uh, the use cases for most other industries. So, um, it's a combination of what is out there because we don't want to reinvent the wheel. But wherever it's not working for our needs, uh, we had to go on, rebuild and build it our way to make it work for us and for our customers.

Speaker A: Got it. Do you use psychologists maybe, um, give to help you with the way that you kind of want to convince people to not commit a crime or if they're doing it, just leave and go, or just be polite because you don't know if they're in the beginning for 100%, if they're like, doing something that they shouldn't be.

Speaker B: So we use a proxy in a way. Right. Uh, instead of going, um, to a psychologist directly, uh, we talk to people in the field what they are doing today. For us, the easiest and the biggest one is a real security guard. I mean a good security guard at that. Right. We ask them, hey, in this situation, what would you do? It comes from their training. Both, um, they've received when they got into the profession, but also the years of experience they've gained, and that becomes the basis for us. And I'm pretty sure, uh, the inputs from the psychologist were part of their training. Right. And also the other source is law enforcement. We asked them, hey, where do you see, by the way, um, the first investor that we had, uh, that we got on board, he was undercover for years.

Speaker A: No way.

Speaker B: Yeah. So he used to be like, mixing with the criminal gangs, understanding their psychology, uh, why do they commit crimes? How do they commit crimes? What stops them from picking, uh, like a particular target? Um, and he distilled all that wisdom and gave it to us. What should you do?

Speaker A: He was your inference model.

Speaker B: Yeah, exactly. He's a combination of law, uh, enforcement, he's psychologist, uh, a criminal mind. You know, all that. Right. And telling us what should you do to stop this from happening? And we are very, very lucky to have people like that on our team to help us design this solution, which is outstanding in every possible way.

Speaker A: Right. This is fascinating. I want to ask now, the physical hardware part of it, is it that, um, you're kind of the, the, the, the hardware of it, is it the cameras or is it just you integrate with, say, the ring cameras that, um, it's available, people are using, uh, what is that part of it?

Speaker B: So a huge design principle. Like when we started, I mean, I did market research. Right. Um, I was a management consultant before. So before you take on anything, you have to understand what, uh, the Industry is doing today. Very interesting observation I had was there are already a billion security cameras. Billion already installed worldwide.

Speaker A: Global.

Speaker B: Okay, globally, right? Think, uh, about all the more cameras. Because it's like we're 8 billion people, maybe more. So I mean, think about all these cameras, right? Like everywhere. On the street corners, on the bridges, on homes and you know, on all the office buildings, campuses, everywhere, right?

Speaker A: Yeah.

Speaker B: Why do I want to go and introduce another camera, right? So why don't I leverage what's already there? And also the biggest cost for most of the security installations, uh, and security monitoring is installing a camera. So think about it, right? You need to bring a truck, you need to have a ladder. The person is, uh, ensure the guy

Speaker A: who went into the Louvre where you can have their ladder.

Speaker B: So, so you can imagine, right, there are a lot of costs to doing anything with hardware. And if it's already kind of working, right, like you mentioned about like, you know, having a camera already in place, if I can leverage that, I cut down costs by 80, 90% of onboarding, right? Uh, so think again. Um, you are a large company, like a Google, right? How many cameras do you think they have? 50,000 or more, right?

Speaker A: In one or is it all through

Speaker B: mountain view across all campuses? Same, right? Um, think, uh, about like a metro transit system. They may have 100,000 security cameras, right? Uh, if you have to go and tell them, like, hey, your cameras are bad, like now let me go and put my cameras. You can imagine the amount of time cost, environmental damage that you're causing unnecessarily, right? So instead what we said is, look, if all these cameras are already there, can I connect to them? And uh, before anything bad happens, before any person is dispatched, can I have my AI work just like a security guard, right? So that was the design principle that we operated with. Um, and so we are a software as a service company, we are hosted in the cloud. So we can run our smart guard solution on any cloud. Azure, uh, aws, gcp, uh, Oracle, I mean you just need any kind of compute, right? And our containerized solution can be deployed there, uh, in their own private instance or we have our hosted like Sentry, AI hosted, multi tenant, uh, uh, SaaS offering as well, right? So with that we connect to the existing infrastructure. Uh, it's like, you know, um, it's like a heavenly song to most, uh, security professionals, right? Oh my God. I don't need to remove what I already invested in, right? Uh, so, and we work from there. We are asset light. We work with what customers Already have. And if things are really bad like you and I cannot even see what's there, right. Then of course we ask them to go refresh it. But then you don't need to everything lock, stock and barrel. Um, try to leverage what you have and replace only what you need to. You will see very soon that our model is going to be a winner from so many perspectives. Cost perspective, time perspective, environmental perspective. Right. You know, why are you, what are you doing with all these cameras that you're removing? Most of them are getting trashed. It's very tragic. Perfectly functional cameras are being replaced by our competitors and thrown into garbage. I would say it's like, it's almost criminal to underutilize current infrastructure.

Speaker A: I'm with you on that 100%. Um, so basically your product is predominantly the software that you can get into get, get that integrated with the softwares of all of the other camera, camera providers that are out there. Question. Is there going to be at some point that we can maybe like, it's crazy, but, um, it might not be super useful actually. But let's say we have cameras on our phones and we have cameras on our laptops and say, hey, I want this, I want to leave this on. I'm just leaving. I want that to be integrated with this software just in case, like I don't have a ring camera in the house or whatever. I'll just leave my laptop open. For example. Um, do you see that? Something that is, you know, have you, you thought about it or something that is.

Speaker B: Absolutely. Our software is agnostic of the source of the video stream or image stream or photos. Right. Uh, you can take it physically with your phone. You can uh, put it on yourself as like a body cam. Uh, it can be on a drone that is flying. It can be on a robotic dog, on a humanoid robo. Um, it can be two cameras, it can be six cameras. Agnostic. We just need to get the stream, we need the context and then uh, the security protocol that needs to be applied. And then it's working. That's what makes the smart guard smart. Right. Um, it's not a particular camera. It's not uh, a camera positioned in a particular way. It's a view of that particular space. If you can give that. And you tell in advance, hey, how do you want this to be processed? How do you want this to be analyzed? That is enough for us. What, what am I protecting and what should I do? What is anomalous and what should I do when I see something anomalous? Just like you are training a security guard. You tell them, you stand here. If you see someone wearing a red shirt, tell them to go away, right? So that's, you know, you have identified a place, uh, a view. You told them what's anomalous, and what should they do when they see that. Those are the three things you provide to our AI agent, which is the smart guard. And that's exactly the behavior that will be, uh, exhibited by our smart card.

Speaker A: So I'm assuming that right now you have the human is in the loop. But eventually, when the model is, like, accurate, at least, um, I would love it to be 100% accurate. Never have any alarms, right? That is false. Um, but when it gets to a standard point of accuracy that is, uh, kind of, um, reliable, then is it in the plans to kind of get rid of that human in the loop layer and just kind of have the model run because now it's there and can do it autonomously again?

Speaker B: My analogies come from, like, biology and from, like, cybersecurity, from defense. Uh, I personally don't see that happening at any point of time. As long as we humans exist, uh, our behaviors are expected. There are people who will try to cause harm, try to steal, and they keep trying innovative ways of doing it. And so just like we protect ourselves against bacteria, viruses, which constantly mutate our cybersecurity, uh, solutions constantly evolve to stay a step ahead of the bad actors. I think physical security will be the same. Uh, there will be new ways people will try to break in, and you need better ways to protect ourselves. Um, of course, the cost dimension will be different. Right. In cybersecurity, you no longer have a person sitting and watching every email that you get to see whether it's spam or not. It's now pretty much delegated to the, um, uh, algorithms. Right. But you do have a person who examines once in a while to see what's working, what's not working, and keep tweaking the algorithm a little bit. So I think physical security will evolve to that point where humans are involved mostly to oversee. Right. Uh, but then you will see a very wide coverage of, uh, the physical security with this solution. So think about all the housing communities, all the offices, uh, that today are not protected because you cannot afford security guards. You cannot put enough security cameras there. Right? So tomorrow, uh, this will become very different. In fact, uh, our, um, investor has coined this as a Silicon Valley security model where most of the work is done by AI using cameras and using other sensors and bringing in elite level Security personnel, right? And, um, you may need more of them than we have today, but that will cover much, much better. Right. And 24, 7 as compared to what we have today. A very patchwork of, uh, security solutions, barely covering what they can and still have museums broken into.

Speaker A: So basically what you're saying is that we're still going to need humans because of the fact that we are humans and we keep learning how to do things. And yet now that whether or not we're going to have all of the 360 degrees of, um, how humans are going to be super creative and what kind of creations they're going to come up with. So that's the reason the security guards are not going to go away. They're just going to be equipped is what I'm hearing.

Speaker B: They will be better trained. You need more of them. But then I basically foresee crime reducing tremendously, uh, with this kind of solution, uh, in place. So again, think about it, right? Today, a security guard may cost $15,000 a month. Right. Uh, because of that, there are a lot of gaps because many people, businesses cannot afford them. But if the cost of monitoring through our smart guard drops to like 2000, 3000 now, suddenly a lot of these places can afford these security guards with better presence. Now, the opportunity for those criminals reduces tremendously. Right. It no longer becomes like a career for most of them. Right? So that is what will reduce the crime rate overall.

Speaker A: And I guess my other question would be, what is your, um, basically target buyer? Is it mostly, um, you know, commercial buildings and maybe like government buildings? Anywhere that needs security and surveillance? Or does it include, um, residentials and people who want it for their own, um, you know, condos, houses, apartments.

Speaker B: Yeah, three kinds of ICPs that we are seeing. Right. Um, the first one is rightly so. Uh, it's mostly commercial real estate, where you have office campuses, hospital campuses, school campuses, uh, that have something valuable, whether people or assets, and, uh, need protection. So on where today a lot of security guards are deployed and they have large incidents of crime happening. So they're definitely looking for these kind of solutions. That's our first target customer. The second is of course, in the public sector. Right. These are all government, um, buildings, uh, public spaces. Think, uh, about airports, uh, concert halls,

Speaker A: um, roadways with the traffics and all that.

Speaker B: Yeah, yeah, the traffic cameras. Right. So there's a lot in the public sector as well. You know, there are a lot of them now open to these technologies coming from the private sector, uh, uh, being adopted there and all the way to defense. Right. Borders, for example, the wars. So they need better ways to know what is happening. So there is a lot of opportunity there for the solutions to be deployed. The third one of course is residential. But typically doing it for individual homes is going to be very expensive. But where the homes are aggregated in the form of like multifamily apartment complexes, gated communities, uh, condo complexes, high rises, even single families, uh, in an HOA or um, as uh, some kind of a grouping. Right. And could even happen through government's support. Any of those help distribute uh, the cost across multiple units. That is our third ICP where a um, group of homes can use this solution to kind of help each other out, um, and pay a portion of the m. Uh, total cost.

Speaker A: Got it. Well, this was an amazing chat. I know that there's um, it's. It's one of those things that there's. There's a lot of questions that come out of it because there's a lot of applications of it. It was great. I loved it. I want to um, kind of wrap it up with one question and that is, is there anything that you want to mention that I didn't ask that? You were hoping that I asked.

Speaker B: So the last thing I actually want to leave with is what is the objective here? Right. So our goal is to improve the safety and security of our communities. When somebody comes and steals even like a hundred dollar bike, you have a sense of violation, right. Uh, of an invasion of your privacy, the sense of security. And I've talked to many people who had these burglaries and uh, break ins happen. Um, it leaves an emotional scar and it's uh, almost like it's traumatic. And they're always now worried, is everything safe or not? Right. So that is what we want to address. And my whole goal with Sentry AI and as kind of the name implies, right, Is to be this sentinel, a watchful solution that prevents things from happening and gives you the peace of mind that no, everything is safe. Uh, your family is safe, your assets are safe. Uh, you can rest assured, you know, uh, you're okay. Right. You have other things. Go and take care of it. Rest, um, assured that no, everything is okay. Right. So that is what we are aiming for at the home level, at uh, the place where you work, right. At the uh, business and commercial level and at the public spaces. Right. Uh, a concert that you go to, a game that you are going to go watch the roadways you're traveling in. You should not have that fear something bad may happen. So that is my vision, that we will all be in a safer, secure, uh, community where we don't need to worry about these things and let technology handle all of that. Uh, and we just go about our lives oblivious, uh, to that, you know, any dangers that may happen. Right. So I'm really hoping, you know, for example, you and I don't fear a tiger, you know, attacking us. Why we know we, we have systems in place to keep them all these uh, harms away from us. Right. Uh, I'm expecting with sentry AI, we get to that same place where the criminals and criminal activities are reduced to the small point where you don't need to worry about it anymore. So that's my big vision I'm hoping, you know, and I'm working hard with my team to make that a reality.

Speaker A: I love that vision. Um, and um, you know, you're working hard on it. And one thing that um, when you were speaking about this, that came to my mind was that our um, feeling of privacy is also another aspect. How much we're going to feel good. Privacy is something we're entitled to, right? And all of this, that is with, with all of the, I mean from the security aspect is one thing, but the other aspect of it that people feeling, being, being surveilled, you know what I mean? Um, and I know that that is something that's on a lot of people's minds. Um, it's, it's kind of for another conversation. But it just comes to my mind that it has just these two sides. I want to be secure, but I also want to have my privacy as well. So there has to be somewhere like either a balance or kind of boundaries really created, uh, for both, for both of these sites.

Speaker B: So the short answer to that, and it's a long topic definitely, right, is uh, a short answer is you will have much better privacy with AI based solutions than you will have ever, uh, with people based solutions. So think about it, right? If you have a security guard outside your home, the guard knows every person who is going in and out, right? But if you design a security through AI to say that hey, alert only when you see something unusual, you'll have far better privacy. And it's the same analogy across the whole, uh, home, work and public place you can design. Unfortunately people are not thinking about it that way. Uh, we actually published a paper on it on how you can design. There's always like fire can be used in a wrong way, right? A knife can be used in a wrong way, but used the right way. It can be very, very beneficial. We solidly think AI is like that. You can actually design it with privacy safeguards in mind far better than the privacy you have today. I think it's in our hands to do so and we are champions of that. I'm pretty high on having those privacy safeguards built in. Uh, and again we can uh, discuss in more detail, uh, on the next uh, uh, podcast. But it's something that's dear to me and I do believe we can achieve both privacy and security, uh, with crafting an intelligently designed, uh, AI based solution.

Speaker A: Um, amen to that. I would love to see that. There's a lot of things that is, you know, just needs to be created. The safeguards, the regulations, uh, that is, that has the human's interest, the citizens interest in mind. There's a, there's a lot to it. It's a huge, uh, conversation. Huge topic that I'm very much excited about too. But yeah, it needs its own, uh, you know, separate convers. It was really great talking with you. I learned a lot. Um, I think it's a cool product that you're working on, which is why I wanted to speak with you about it. And the way people can reach out to you is, is your LinkedIn good?

Speaker B: LinkedIn. Come to our website. We are available on multiple socials as well. The best way, find me on LinkedIn, send me a note. More than happy to get on a call, meeting, um, and discuss more about how AI can really help us.

Speaker A: All right, thank you very much.

Speaker B: Yeah, thank you, Sarah. Appreciate your time as well. Thank you.