1 00:00:00,309 --> 00:00:03,853 Back in 2018, in the wake of the Parkland High School mass shooting, 2 00:00:04,373 --> 00:00:07,817 a group of former Navy SEALs banded together and launched a company 3 00:00:07,937 --> 00:00:11,440 called Zero Eyes. Their mission is 4 00:00:11,580 --> 00:00:15,024 to prevent and mitigate these types of violent 5 00:00:15,084 --> 00:00:18,447 events. And the means that they're going about achieving that mission is 6 00:00:18,547 --> 00:00:22,612 through technology. Here we're talking about artificial intelligence, specifically 7 00:00:23,013 --> 00:00:26,477 object detection algorithms. So they went and they trained 8 00:00:26,978 --> 00:00:30,743 algorithms designed to detect the presence of 9 00:00:30,903 --> 00:00:34,047 guns. Then they combine those algorithms, they 10 00:00:34,128 --> 00:00:37,852 run those algorithms on CCTV cameras. to 11 00:00:37,992 --> 00:00:41,675 flag specific instances of person 12 00:00:42,096 --> 00:00:45,458 with one of those weapons. So my guest 13 00:00:45,518 --> 00:00:49,241 today is Tim Solcer, who is one of ZeroEye's co-founders 14 00:00:49,381 --> 00:00:53,665 and the company's CTO. And we talked beyond just 15 00:00:53,825 --> 00:00:57,448 how the models work and how they were made to work and validated to 16 00:00:57,468 --> 00:01:01,310 work. but into the impacts that they're seeing on 17 00:01:01,350 --> 00:01:04,471 the ground, the impact that they're having, the ways in 18 00:01:04,511 --> 00:01:09,194 which this level of automated security is 19 00:01:09,314 --> 00:01:12,575 changing operations on the ground, and where it all could go. 20 00:01:12,975 --> 00:01:16,037 So, if you are new to the pod, I 21 00:01:16,097 --> 00:01:19,539 am your host, Ian Kreitzberg. If you are not new, you 22 00:01:19,579 --> 00:01:22,740 knew that already, and thanks for being here. This is The 23 00:01:22,780 --> 00:01:34,185 Deep View Conversations. Tim, 24 00:01:34,245 --> 00:01:37,647 thanks so much for joining me today. I appreciate you having me on. Yeah. So, 25 00:01:38,807 --> 00:01:41,969 you know, we've connected before. It's been a 26 00:01:42,789 --> 00:01:46,251 couple of years, which is kind of crazy. But ZeroEyes has been around, 27 00:01:47,432 --> 00:01:51,854 you know, even before that. You guys got your start in 2018. Let's 28 00:01:51,874 --> 00:01:55,196 just start there. And then there's plenty of other stuff to 29 00:01:55,236 --> 00:01:59,558 dive into. But with the kind of foundational 30 00:02:03,800 --> 00:02:06,883 So we got started in 2018 shortly after the shooting in 31 00:02:06,923 --> 00:02:10,645 Parkland, Florida. And that shooting was particularly 32 00:02:10,685 --> 00:02:14,368 terrible because the shooter was in a stairwell underneath 33 00:02:14,809 --> 00:02:18,031 a security camera for, I think, three or five minutes before the 34 00:02:18,071 --> 00:02:21,234 first shot was fired. So we looked at that 35 00:02:21,254 --> 00:02:24,456 and we essentially had this idea of, you know, if 36 00:02:24,476 --> 00:02:27,919 somebody was watching that camera, they would have been able to stop that shooting. 37 00:02:28,239 --> 00:02:31,822 And why aren't we watching cameras? It's basically because there's not enough attention 38 00:02:31,862 --> 00:02:35,545 span and people hours to be able to have eyes on every camera everywhere. 39 00:02:36,645 --> 00:02:40,187 And at the same time, AI was 40 00:02:40,267 --> 00:02:43,968 progressing, object detection was becoming a reality. I'd 41 00:02:44,008 --> 00:02:47,329 worked with some computer vision in a previous startup, and 42 00:02:47,349 --> 00:02:50,471 it was a natural progression to be able to say, you know, if we 43 00:02:50,491 --> 00:02:55,072 can detect any number of objects, faces, dogs, 44 00:02:55,152 --> 00:02:58,514 cats, why can't we detect guns on security cameras and use 45 00:02:58,534 --> 00:03:02,392 that as a means to provide proactive 46 00:03:02,432 --> 00:03:05,609 situational awareness during a shooting. Early days of 47 00:03:05,649 --> 00:03:08,911 ZeroEyes, we got started basically just looking for, can 48 00:03:08,951 --> 00:03:13,373 we build a model to detect guns? That was the first MVP. 49 00:03:13,753 --> 00:03:17,515 And we started off by collecting images from Google 50 00:03:17,555 --> 00:03:20,876 images, web scraping images, open image data set, and 51 00:03:20,936 --> 00:03:24,518 basically trying to build a model that would 52 00:03:24,538 --> 00:03:28,420 detect guns. And we tested on videos like clips 53 00:03:28,440 --> 00:03:32,082 from The Matrix and, you know, random images that we scraped 54 00:03:32,102 --> 00:03:35,635 from the web. We found out really quickly that Those 55 00:03:35,675 --> 00:03:39,318 types of models that are trained off web images 56 00:03:39,418 --> 00:03:43,741 are not easily transferable or generalizable to 57 00:03:43,821 --> 00:03:47,384 security cameras. And so we deployed our first model on a security camera 58 00:03:47,404 --> 00:03:50,606 and the performance was terrible. The images didn't 59 00:03:50,646 --> 00:03:54,189 match up to what the AI was trained on. So the 60 00:03:54,229 --> 00:03:57,732 next step from there was we went out and bought some cheap 61 00:03:57,752 --> 00:04:00,914 security cameras on Amazon and hung them up in our 62 00:04:00,954 --> 00:04:04,317 CEO's backyard, which is where our office was at the time. We were working out 63 00:04:04,337 --> 00:04:07,928 of his basement. And we 64 00:04:07,968 --> 00:04:11,148 found out really quickly that generating our own data in 65 00:04:11,168 --> 00:04:14,529 a realistic environment using a real camera, a 66 00:04:14,569 --> 00:04:17,850 real sensor, was the trick that we needed. And 67 00:04:17,890 --> 00:04:21,331 so we invested most of our early time in the company into 68 00:04:21,371 --> 00:04:24,792 just building a quality model and building an organic data set. 69 00:04:25,063 --> 00:04:28,425 That's the perfect jumping off point, right? I mean, we 70 00:04:28,445 --> 00:04:31,867 hear about this all the time, kind of no matter the application, if you're talking about, 71 00:04:32,408 --> 00:04:36,370 you know, studying whales off the Pacific coast, or in 72 00:04:36,590 --> 00:04:40,092 your case, identifying weapons, it all comes down to data, 73 00:04:40,933 --> 00:04:45,942 the quantity of the data and the quality of the data. So I Tell 74 00:04:45,982 --> 00:04:49,285 me more about that. How much 75 00:04:49,405 --> 00:04:52,568 data did you have to collect? And how did 76 00:04:52,608 --> 00:04:56,732 you think about coming up with different ways of varying 77 00:04:56,852 --> 00:05:00,135 the types of data that you're collecting to making sure 78 00:05:00,155 --> 00:05:03,418 you're getting different angles of, I guess, as great 79 00:05:03,438 --> 00:05:07,001 a variety of weapons as you could kind of conceive of? Yeah, 80 00:05:09,063 --> 00:05:12,405 Being a co-founder with four former Navy SEALs, we 81 00:05:12,445 --> 00:05:17,747 never had a shortage of weapons to use as examples. But 82 00:05:17,967 --> 00:05:21,288 exactly that, garbage in equals garbage out. 83 00:05:21,589 --> 00:05:24,690 And we view our data set as probably the most important thing in the 84 00:05:24,750 --> 00:05:28,031 company right now. In the early days, we never had enough data. 85 00:05:28,412 --> 00:05:32,093 That was 100% of our problem. And 86 00:05:32,113 --> 00:05:35,775 we spent a lot of time traveling to different environments, using 87 00:05:35,835 --> 00:05:39,313 security cameras to collect data and with 88 00:05:39,353 --> 00:05:43,215 different backgrounds, different lighting conditions. And that was 89 00:05:43,235 --> 00:05:46,916 really the challenge that we identified early on was security 90 00:05:46,936 --> 00:05:50,877 cameras are mounted in all different types of environments. Security 91 00:05:50,917 --> 00:05:54,518 camera quality, video quality from cameras varies greatly from 92 00:05:54,558 --> 00:05:58,160 camera to camera and manufacturer to manufacturer. So 93 00:05:59,280 --> 00:06:03,001 it was really important for us to have representative examples of 94 00:06:03,182 --> 00:06:06,543 all of the types of environments that we really wanted to detect guns in. 95 00:06:08,257 --> 00:06:11,919 It started in Mike's backyard and progressed to local schools. 96 00:06:12,260 --> 00:06:15,502 We spent every weekend for months in 97 00:06:15,522 --> 00:06:18,764 the early days traveling to different schools and using their 98 00:06:18,804 --> 00:06:23,867 camera systems to record weapon data. And 99 00:06:23,887 --> 00:06:27,269 then the next step from there was we really 100 00:06:27,289 --> 00:06:30,891 wanted schools to allow us to record data when there was actual 101 00:06:31,672 --> 00:06:34,934 people in the camera views. But 102 00:06:34,974 --> 00:06:38,176 schools were not very willing to have us walk around during 103 00:06:38,196 --> 00:06:41,897 the daytime with guns, with active students in the hallways. 104 00:06:41,977 --> 00:06:45,298 So we then came up with the idea, well, if we can't 105 00:06:45,338 --> 00:06:48,899 collect this data in real time, how can we use the customer 106 00:06:48,939 --> 00:06:53,360 backgrounds to build in some generalization 107 00:06:53,520 --> 00:06:57,241 into the model? And our solution was to build a green 108 00:06:57,281 --> 00:07:01,763 screen AI lab, essentially. So we 109 00:07:02,303 --> 00:07:05,824 built out a 5,000 square foot AI lab with 110 00:07:05,904 --> 00:07:09,763 full green screen walls. And we hung about 111 00:07:09,843 --> 00:07:13,305 100 security cameras throughout the warehouse. 112 00:07:14,825 --> 00:07:18,227 So we were able to walk around with ZeroEyes employees with guns 113 00:07:18,527 --> 00:07:21,728 and then overlay our customer backgrounds behind them to 114 00:07:21,768 --> 00:07:25,090 give that context in the scene that we weren't able 115 00:07:25,130 --> 00:07:28,552 to record in real life. And that's where 116 00:07:28,592 --> 00:07:31,713 we've been for the last few years. But there's a lot of 117 00:07:31,753 --> 00:07:35,175 trends around synthetic data, synthetic data 118 00:07:35,215 --> 00:07:38,349 generation that We'll probably obsolete that at 119 00:07:38,389 --> 00:07:42,494 some point. But we've 120 00:07:42,554 --> 00:07:45,798 invested a lot as a company in having a really high 121 00:07:45,838 --> 00:07:49,481 quality data set that represents as many possible scenarios 122 00:07:50,301 --> 00:07:53,942 So beyond that, you talked a little bit about the algorithms, and 123 00:07:54,142 --> 00:07:57,923 I just want to nail a little deeper into 124 00:07:57,963 --> 00:08:01,444 that. You said back in 2018, we were seeing 125 00:08:01,504 --> 00:08:05,505 advancements in object detection, which is basically, that's 126 00:08:05,645 --> 00:08:09,346 the core thing behind what enables your 127 00:08:09,386 --> 00:08:12,807 technology. That's also one 128 00:08:12,827 --> 00:08:17,089 of the major algorithms in self-driving cars. Um, so 129 00:08:17,109 --> 00:08:20,751 it talked to me about beyond the dataset, 130 00:08:20,891 --> 00:08:24,272 building the system itself, the algorithms, um, 131 00:08:25,233 --> 00:08:28,915 how do you validate that? What was the process 132 00:08:28,975 --> 00:08:32,456 like? And, uh, you know, you're building from scratch. You're 133 00:08:32,496 --> 00:08:35,778 not piggybacking off of other systems. It sounds like, 134 00:08:35,858 --> 00:08:39,540 so I imagine if the data collection process was 135 00:08:39,620 --> 00:08:43,042 intensive, the algorithm construction process. 136 00:08:45,322 --> 00:08:48,863 Yeah, there's a few different steps to it. Because when you think about the 137 00:08:48,963 --> 00:08:52,564 entire video pipeline, the processes of the data, there's many different steps 138 00:08:52,684 --> 00:08:56,205 from simply just decoding the video, then passing 139 00:08:56,245 --> 00:08:59,865 it through an inference engine, and then object tracking. So 140 00:09:01,546 --> 00:09:05,146 it's changed quite a bit over the years. But in the early days, we started with 141 00:09:06,567 --> 00:09:10,528 probably the best object detection technology at the time was fast R-CNN or 142 00:09:10,628 --> 00:09:13,831 faster R-CNN models. We were just starting 143 00:09:13,851 --> 00:09:17,173 to see YOLO models be released, which I think we're on 144 00:09:17,834 --> 00:09:21,776 maybe the eighth or tenth iteration of YOLO models at this point. And 145 00:09:21,996 --> 00:09:25,138 it's kind of diverging quite a bit. But in 146 00:09:25,158 --> 00:09:28,580 the early days, it really affected our hardware processing. I 147 00:09:28,620 --> 00:09:31,922 mean, we were trying to solve this problem of 148 00:09:32,102 --> 00:09:36,164 being able to process real-time video at 149 00:09:36,204 --> 00:09:39,499 the customer's site. Because customers are sensitive to 150 00:09:39,539 --> 00:09:43,763 their security camera data being sent 151 00:09:43,803 --> 00:09:47,187 somewhere else. So we identified really 152 00:09:47,227 --> 00:09:50,510 quickly we have to do this with GPUs. We have to do it with GPUs on 153 00:09:50,550 --> 00:09:54,194 premise. And so the model selection came down to what 154 00:09:54,254 --> 00:09:57,758 model gives us the best balance of accuracy and compute 155 00:09:57,798 --> 00:10:01,697 efficiency. For us, faster RCNN 156 00:10:01,737 --> 00:10:05,318 models were the highest performing at that time. They 157 00:10:05,338 --> 00:10:08,779 were definitely more computationally heavy, which 158 00:10:08,839 --> 00:10:12,860 meant that we couldn't load as many cameras per GPU as we wanted to, which 159 00:10:12,920 --> 00:10:16,462 affected our economics. But over time, the 160 00:10:16,502 --> 00:10:19,983 great thing about research and academia is that they're constantly putting 161 00:10:20,023 --> 00:10:23,246 out new stuff. And for the most part, it's open source. So 162 00:10:23,266 --> 00:10:27,230 we're able to use open source off-the-shelf 163 00:10:27,291 --> 00:10:30,514 models, which has progressed to, I think we're 164 00:10:30,594 --> 00:10:35,399 using Ultralytics models today, YOLOv5 165 00:10:35,800 --> 00:10:39,084 or 8. But with 166 00:10:39,244 --> 00:10:42,717 those advancements in model algorithms, that's 167 00:10:42,757 --> 00:10:46,259 brought with it increases in accuracy. As we're increasing the 168 00:10:46,480 --> 00:10:50,062 quality of our data set, the models, the algorithms themselves 169 00:10:50,102 --> 00:10:53,404 are getting better, and the speed is getting better. So we're able 170 00:10:53,444 --> 00:10:56,927 to run models on more cameras, higher 171 00:10:56,967 --> 00:11:00,149 resolutions, higher frame rates, which all of 172 00:11:00,189 --> 00:11:03,511 those things turn into better detection performance for us. And 173 00:11:03,531 --> 00:11:07,416 we expect that trend to continue in the future. Obviously, 174 00:11:07,456 --> 00:11:10,719 we're still talking about object detection models today, which is kind 175 00:11:10,759 --> 00:11:14,162 of like single frame analysis, but with 176 00:11:14,222 --> 00:11:17,705 the potential for large language models and 177 00:11:17,845 --> 00:11:21,028 vision transformers and things like that, we see a lot 178 00:11:21,048 --> 00:11:24,530 of advancement in the future, not just on the detection side, but also on 179 00:11:25,031 --> 00:11:28,314 the context side. What can we communicate about the 180 00:11:28,374 --> 00:11:32,337 image itself to our customers to provide the best situational intelligence? 181 00:11:33,246 --> 00:11:36,850 Hmm. Yeah. I'm glad you brought that up. Cause I was going to ask you about that. Uh, 182 00:11:36,870 --> 00:11:40,413 but the, the, that big area of advancement that 183 00:11:40,453 --> 00:11:43,776 we've seen over the past couple of years, it's all kind of centered around large 184 00:11:43,816 --> 00:11:47,280 language models, uh, transformer architecture, the vision models 185 00:11:47,320 --> 00:11:50,723 that you, that you mentioned. Um, so that, that 186 00:11:50,783 --> 00:11:54,026 is something that you're exploring early 187 00:11:56,027 --> 00:11:59,770 Yeah, it's something that we see as a huge opportunity because today 188 00:11:59,810 --> 00:12:03,132 we're providing this situational awareness of gun 189 00:12:03,212 --> 00:12:06,575 or no gun. There's either a gun in the image or there's not. And 190 00:12:06,595 --> 00:12:11,078 then we have a human in the loop that allows us to maximize 191 00:12:11,118 --> 00:12:14,701 our detection performance while eliminating false positives. So 192 00:12:14,761 --> 00:12:18,864 customers will never receive a false positive. But 193 00:12:18,904 --> 00:12:22,047 the potential with large language models is now you can extract a little 194 00:12:22,067 --> 00:12:25,281 bit more context from the scene. And you can also provide a 195 00:12:25,321 --> 00:12:28,803 little bit more dynamic inputs. So looking for 196 00:12:28,903 --> 00:12:33,005 more dynamic scenes or different objects in different configurations 197 00:12:33,105 --> 00:12:36,606 that wouldn't be realistic with a single model 198 00:12:37,930 --> 00:12:41,232 We'll stick on that side for a second. You mentioned something else that I want to get into in 199 00:12:41,252 --> 00:12:45,836 a minute, but that'll keep. On 200 00:12:45,856 --> 00:12:49,198 the language model side, the idea of greater context on the images that you're talking 201 00:12:49,238 --> 00:12:53,001 about, right? That's really interesting, because like you said, right 202 00:12:53,041 --> 00:12:56,223 now, gun or no gun, that's kind of it. 203 00:12:56,684 --> 00:13:00,386 Other context, I mean, in a security situation 204 00:13:00,407 --> 00:13:03,969 like that, any added piece of information is 205 00:13:04,369 --> 00:13:08,473 probably going to be very nice to have. But 206 00:13:08,553 --> 00:13:11,796 I wonder, as you start experimenting with 207 00:13:11,896 --> 00:13:15,800 deploying or incorporating additional 208 00:13:15,840 --> 00:13:19,703 models of different types, I 209 00:13:19,743 --> 00:13:23,467 wonder how that changes your validation process. Because 210 00:13:23,527 --> 00:13:27,691 the gun or no gun thing is a little more of a straightforward thing, 211 00:13:27,791 --> 00:13:31,334 it's a different type of model. Language models have 212 00:13:31,774 --> 00:13:35,556 reliability issues and 213 00:13:35,697 --> 00:13:39,058 I wonder how much that might be a challenge or if 214 00:13:39,119 --> 00:13:42,360 there's a way maybe through specific data sets, 215 00:13:42,641 --> 00:13:46,623 small language models, other techniques and safeguards to 216 00:13:48,224 --> 00:13:52,555 Well, I think there's two parts of that. One is Just 217 00:13:52,595 --> 00:13:55,796 like all good security comes in layers, I think the same thing is going to 218 00:13:56,036 --> 00:13:59,498 be applicable for IAI. In 219 00:13:59,558 --> 00:14:03,139 our case, I think there'll be multiple layers of AI models 220 00:14:03,179 --> 00:14:06,580 in the future that process all of the detections that we generate, 221 00:14:07,021 --> 00:14:11,952 but also kind of highlights the value of that human in the loop. If 222 00:14:12,793 --> 00:14:16,115 we were to send a detection directly through to 223 00:14:16,135 --> 00:14:19,238 the end customer and provide some layer of analysis from a 224 00:14:19,298 --> 00:14:22,720 large language model, it's very possible that that's incorrect and 225 00:14:23,041 --> 00:14:26,163 that it could confuse or cause some sort of 226 00:14:26,203 --> 00:14:30,840 issue in the response, the critical incident response. And 227 00:14:30,940 --> 00:14:34,121 today we have a human in the loop that essentially does the same thing, but 228 00:14:34,141 --> 00:14:38,343 they're also capable of performing other actions. So during 229 00:14:38,403 --> 00:14:42,244 a critical scenario, when we see a gun and click dispatch, that's 230 00:14:42,324 --> 00:14:45,846 also initiating a call to 911 dispatch 231 00:14:46,286 --> 00:14:49,627 that's closest to the camera location. And someone on 232 00:14:49,667 --> 00:14:52,828 our team is also getting on the phone with points of contact at the 233 00:14:52,868 --> 00:14:58,170 customer to be able to verbally communicate these things. I 234 00:14:58,190 --> 00:15:02,436 don't see the value of having a human in the loop ever being fully 235 00:15:03,617 --> 00:15:07,542 automated or deprecated because 236 00:15:08,683 --> 00:15:11,907 the value during a critical scenario is we're able to 237 00:15:12,007 --> 00:15:15,492 communicate directly to that POC and not rely on 238 00:15:17,830 --> 00:15:21,532 And so I, I guess what you were just talking about there, that that's the point that 239 00:15:21,553 --> 00:15:24,755 I wanted to get to and make sure we drill down on. Um, cause we've 240 00:15:24,775 --> 00:15:28,957 been kind of talking around it for anyone who's not familiar, um, 241 00:15:29,358 --> 00:15:33,460 with what you guys do, right. It's, it's. Object 242 00:15:33,500 --> 00:15:36,923 detection designed for, uh, you know, to, to scan for 243 00:15:37,243 --> 00:15:41,126 weaponry and, uh, CCTV footage. That's 244 00:15:41,346 --> 00:15:44,608 all linked up with warning systems. You have teams of people 245 00:15:44,648 --> 00:15:48,510 that review flags from the system, but I 246 00:15:49,271 --> 00:15:53,012 would love if you could kind of walk me through, if I was a point of contact, if 247 00:15:53,072 --> 00:15:56,854 I was one of your customers and we put in however 248 00:15:56,874 --> 00:16:00,956 many dozens of cameras we had and 249 00:16:02,237 --> 00:16:05,278 something happens, your camera picks up a 250 00:16:05,318 --> 00:16:08,720 flag, can you just walk me through what 251 00:16:08,760 --> 00:16:12,041 that process looks like from kind of inception of the 252 00:16:12,081 --> 00:16:15,403 model says we might have something here and whatever 253 00:16:16,583 --> 00:16:20,246 I'll take a step back and just talk about it from the entire value chain, because 254 00:16:20,266 --> 00:16:24,049 you're absolutely right. We didn't actually cover that in the beginning. We're 255 00:16:24,069 --> 00:16:27,891 connecting to real-time security cameras, all of the customer's existing security 256 00:16:27,911 --> 00:16:31,614 infrastructure, and we're pulling a RTSP 257 00:16:31,634 --> 00:16:35,056 feed, a real-time streaming protocol feed, from that camera 258 00:16:35,237 --> 00:16:38,579 and running an AI on the video feed frame by frame that's 259 00:16:38,639 --> 00:16:41,938 looking for the presence of a gun. at the point at which our 260 00:16:41,978 --> 00:16:45,200 AI says, it's pretty confident that this object is a gun, 261 00:16:45,580 --> 00:16:48,721 it's going to send that detection to a human in the loop. And the 262 00:16:48,761 --> 00:16:52,423 human in the loop is in our ZOC, our Zero-wise Operation 263 00:16:52,483 --> 00:16:56,065 Center. We have one located just outside of Philadelphia here, 264 00:16:56,305 --> 00:16:59,466 and then another in Honolulu, Hawaii. So we're 265 00:16:59,506 --> 00:17:04,108 able to make use of those time zone differences. The 266 00:17:04,408 --> 00:17:08,408 operators in the ZOC are performing an analysis to 267 00:17:08,448 --> 00:17:11,511 verify whether or not the detection has a real gun in it. 268 00:17:11,871 --> 00:17:15,654 So they see a real gun and they click dispatch. That 269 00:17:15,714 --> 00:17:19,057 dispatch button initiates all of our third-party alerting 270 00:17:19,097 --> 00:17:22,980 methods. So we have a dashboard and 271 00:17:23,000 --> 00:17:26,382 a mobile app ourselves, but we also integrate with 272 00:17:26,562 --> 00:17:30,422 local 911. We integrate with other third-party services 273 00:17:30,482 --> 00:17:33,702 in order to get the detection information to 274 00:17:33,742 --> 00:17:36,943 the customer in the best way that they can utilize it. 275 00:17:37,163 --> 00:17:41,144 So from dispatch, our 276 00:17:41,284 --> 00:17:45,205 ZOC operators will get connected to local 911. They'll 277 00:17:45,225 --> 00:17:48,926 communicate and verify that a local 911 has access 278 00:17:48,986 --> 00:17:52,447 to the alert image, and they're able to generate an 279 00:17:52,507 --> 00:17:56,028 incident based on that. But at the same time, we're also calling 280 00:17:56,269 --> 00:17:59,551 points of contact at the customer site. And what we're trying 281 00:17:59,571 --> 00:18:04,795 to communicate is basically what we're visualizing 282 00:18:04,875 --> 00:18:08,178 on our side. So the 283 00:18:08,218 --> 00:18:11,880 ZOC has a specific script that they stick to, but it's essentially something 284 00:18:11,920 --> 00:18:15,223 like we have an alert, a zeroized weapon 285 00:18:15,243 --> 00:18:18,405 detection alert of what appears to be a 286 00:18:18,445 --> 00:18:22,920 person brandishing a rifle in this setting. That 287 00:18:23,040 --> 00:18:26,463 is really the initiation for an incident 288 00:18:26,503 --> 00:18:30,426 on the customer side. Customers have all different standard operating procedures 289 00:18:30,506 --> 00:18:35,170 of how they want to be notified and what they do following a notification. But 290 00:18:35,370 --> 00:18:38,653 I do foresee more automation in the future around that, where 291 00:18:39,974 --> 00:18:43,277 today we're basically handing off situational awareness to 292 00:18:43,317 --> 00:18:46,599 the customer and allowing them to respond. But 293 00:18:46,659 --> 00:18:50,022 I see a lot of benefit in ZeroEyes being involved 294 00:18:50,042 --> 00:18:53,255 in that response in some way. whether it's something as 295 00:18:53,295 --> 00:18:56,456 simple as providing expertise in 296 00:18:56,796 --> 00:19:00,318 how a school or a commercial office building should respond 297 00:19:00,358 --> 00:19:03,880 to that threat, all the way through to initiating other 298 00:19:05,341 --> 00:19:08,822 forms of response, like dispatching alerts 299 00:19:09,062 --> 00:19:12,864 into access control systems, providing 300 00:19:13,024 --> 00:19:17,566 one-click access control capabilities, so 301 00:19:17,586 --> 00:19:21,625 that customers can actually take action from the ZeroEyes Alert. Um, 302 00:19:24,268 --> 00:19:27,571 Kind of in line with what you were just mentioning about being more involved in 303 00:19:27,711 --> 00:19:31,594 the action. I wonder how far something like this goes, 304 00:19:32,815 --> 00:19:35,958 you know, and, and right. And thinking about where you are right 305 00:19:35,998 --> 00:19:39,421 now, where you might want to go as you keep growing, 306 00:19:39,942 --> 00:19:45,223 since you're analyzing real time footage. Right? 307 00:19:45,443 --> 00:19:48,646 Like that's kind of the whole fundamental crux of 308 00:19:48,706 --> 00:19:52,268 what you're dealing with. If there was an 309 00:19:52,308 --> 00:19:56,131 attack, you know, a person brandishing the rifle, or maybe the person's in 310 00:19:56,171 --> 00:20:02,277 a hallway, they're locked in a room, they're moving. Would 311 00:20:02,297 --> 00:20:05,980 you be able to track the movement of these 312 00:20:06,060 --> 00:20:10,002 kinds of assailants? And in an action stage, 313 00:20:10,863 --> 00:20:14,165 beyond just saying, there's a person here, do something about it, 314 00:20:14,365 --> 00:20:18,368 could you physically or do you physically you 315 00:20:18,388 --> 00:20:21,929 know, speak to local police 316 00:20:21,989 --> 00:20:25,169 officers on the phone and say, okay, this is exactly where the 317 00:20:25,209 --> 00:20:28,370 person is. They're in this hallway. They are doing this right now. They are 318 00:20:28,430 --> 00:20:32,290 moving. They just turned left kind of like an overwatch thing. 319 00:20:32,430 --> 00:20:35,571 Is that, is that a function at all of what you 320 00:20:37,651 --> 00:20:40,712 So yes and no. Uh, we aren't, uh, we 321 00:20:40,752 --> 00:20:44,192 don't have visibility into the customer's camera systems to the extent that 322 00:20:44,232 --> 00:20:47,393 we could actually provide that video or watch a 323 00:20:47,413 --> 00:20:51,081 shooter. walk from camera to camera, and we do that for privacy reasons. Basically, 324 00:20:51,141 --> 00:20:55,602 our AI is the only piece that has access to the live video. But 325 00:20:55,622 --> 00:20:59,182 that being said, you're highlighting on a really key point of our value. And 326 00:20:59,222 --> 00:21:02,303 that is, even after the first shot is fired and we know there's a 327 00:21:02,343 --> 00:21:05,564 gunman on site, we don't know where 328 00:21:05,584 --> 00:21:09,264 he is at any specific point, or the first responders 329 00:21:09,304 --> 00:21:12,425 don't know. And so every time that shooter walks in 330 00:21:12,465 --> 00:21:15,526 front of a new camera that's running zero-wise, and we're able to get a 331 00:21:15,566 --> 00:21:18,869 new detection, we are dispatching that and updating law 332 00:21:18,889 --> 00:21:22,352 enforcement. We do have some local police 333 00:21:22,372 --> 00:21:25,874 departments that use our mobile app and they get the alerts directly to 334 00:21:25,914 --> 00:21:29,496 the mobile app for each individual officer. But 335 00:21:29,516 --> 00:21:32,779 the general feedback we've gotten is that they would prefer to get that 336 00:21:32,819 --> 00:21:36,041 information over the radio from dispatch. And so 337 00:21:36,141 --> 00:21:39,953 our biggest point of contact is that 911 dispatch center. And 338 00:21:40,233 --> 00:21:43,915 our integration with a company called Rapid SOS gives 339 00:21:43,975 --> 00:21:47,597 us the ability to automatically be connected to the 340 00:21:47,617 --> 00:21:50,838 closest 911 dispatch center to where our 341 00:21:50,878 --> 00:21:54,079 customer is located based on their camera location. So that 342 00:21:54,119 --> 00:21:58,261 puts us in direct phone contact with 911 PSAPs, 343 00:21:58,561 --> 00:22:01,723 public safety answering points, and those PSAPs are the ones that 344 00:22:06,024 --> 00:22:09,406 And so because of that integration, any officers on 345 00:22:09,426 --> 00:22:12,747 the ground, as well as points of contact at 346 00:22:13,087 --> 00:22:16,249 whatever building or location you're 347 00:22:16,269 --> 00:22:19,551 dealing with, they know the flag was picked 348 00:22:19,651 --> 00:22:23,392 up by, you know, camera 22B 349 00:22:23,473 --> 00:22:26,594 and hallway seven on the second floor or whatever it is. 350 00:22:26,934 --> 00:22:30,696 So they have a deeper level of situational awareness beyond 351 00:22:33,624 --> 00:22:36,885 Definitely. You also highlight a really important point, 352 00:22:36,925 --> 00:22:40,506 which is a unified mapping interface. Because 353 00:22:41,666 --> 00:22:44,947 if our customer is looking at a map, and we're looking at a different map, 354 00:22:45,287 --> 00:22:49,588 and police responding are looking at a third different map, it 355 00:22:49,988 --> 00:22:54,889 becomes really confusing trying to communicate about landmarks and 356 00:22:54,929 --> 00:22:58,557 where things are located. And so having a unified mapping 357 00:22:58,577 --> 00:23:02,200 interface, which we partner with a company called CRG, Critical 358 00:23:02,240 --> 00:23:05,443 Response Group, and they produce really high 359 00:23:05,503 --> 00:23:08,744 quality interior maps. When 360 00:23:09,145 --> 00:23:12,587 our customer has a CRG map, they're able to provide that to us, and 361 00:23:12,667 --> 00:23:16,210 we are able to overlay that on our dispatch map so that both 362 00:23:16,250 --> 00:23:19,973 the customer and our ZOC operation team is 363 00:23:19,993 --> 00:23:23,375 looking at the same map. And wherever possible, we also include 364 00:23:23,415 --> 00:23:27,238 the local 9-1-1 PSAPs in that when we send our alerts, so 365 00:23:31,694 --> 00:23:35,937 Now, you mentioned for the ZOC, for the operators that 366 00:23:36,157 --> 00:23:40,179 review these flags, that no 367 00:23:40,559 --> 00:23:44,482 false positives, I think is what you said. How are these 368 00:23:44,622 --> 00:23:47,984 operators trained? And I guess, what's 369 00:23:48,084 --> 00:23:51,947 their protocol? in terms of validating that 370 00:23:52,027 --> 00:23:55,310 whatever has been flagged is real or 371 00:23:55,550 --> 00:23:58,893 is not a concern. And I mean, I guess taking 372 00:23:58,913 --> 00:24:02,055 that a step further, if you get a flag that turns out not to be a 373 00:24:02,115 --> 00:24:05,498 gun, but maybe it's something a little weird, is that 374 00:24:07,760 --> 00:24:11,343 I would say the operators back in the ZOC have really difficult jobs. 375 00:24:11,903 --> 00:24:15,186 They are on a daily basis, they see 376 00:24:15,206 --> 00:24:18,464 guns on a daily basis. A lot of times it's 377 00:24:18,524 --> 00:24:22,567 toy guns or ROTC rifles, but they're constantly required 378 00:24:22,607 --> 00:24:25,808 to make split-second decisions, which 379 00:24:25,848 --> 00:24:30,211 is why we focused on hiring veterans and former 380 00:24:30,271 --> 00:24:34,193 law enforcement into the ZOC, because those individuals 381 00:24:34,233 --> 00:24:37,915 back there both have the training and understanding what 382 00:24:37,955 --> 00:24:41,437 a gun looks like and how to respond to a critical scenario, but 383 00:24:41,697 --> 00:24:44,839 also as a part of their former careers as 384 00:24:45,819 --> 00:24:49,080 military and law enforcement, A lot of times they were doing very 385 00:24:49,120 --> 00:24:52,921 similar jobs. They either had watch posts or 386 00:24:53,001 --> 00:24:56,362 they were performing some sort of surveillance where they had to 387 00:24:56,823 --> 00:25:00,203 do basically the same thing and 388 00:25:00,243 --> 00:25:03,744 be able to, during a critical scenario, remain 389 00:25:03,784 --> 00:25:07,805 calm and calmly communicate critical 390 00:25:07,846 --> 00:25:11,506 pieces of information that could be happening at the same time people 391 00:25:11,567 --> 00:25:15,408 are actively under threat. So we've 392 00:25:15,728 --> 00:25:19,329 had a lot of success hiring former law enforcement, 393 00:25:19,389 --> 00:25:22,851 former military personnel into the ZOC. And like 394 00:25:22,891 --> 00:25:26,452 I said, they have a really difficult job. On a daily basis, 395 00:25:27,012 --> 00:25:30,674 they have to understand unique standard operating 396 00:25:30,714 --> 00:25:33,815 procedures from different customers and be able to communicate to the 397 00:25:33,855 --> 00:25:39,197 customer in the best way that the customer needs that information. And 398 00:25:39,237 --> 00:25:43,258 how that translates into real life, I mean, We've 399 00:25:43,578 --> 00:25:47,701 made dozens arrests at this point of people 400 00:25:47,721 --> 00:25:51,503 that had guns in areas that they shouldn't have. But 401 00:25:51,523 --> 00:25:56,347 we also communicate with customers on a daily basis about non-lethal 402 00:25:56,527 --> 00:26:04,152 gun threats. And like I said, every customer has their own SOPs. For 403 00:26:04,192 --> 00:26:07,974 instance, some customers want us to disregard any 404 00:26:08,014 --> 00:26:11,336 toy gun detections. Other customers still want to know about them. 405 00:26:13,462 --> 00:26:16,644 And that extends to a wide array of 406 00:26:16,744 --> 00:26:20,267 scenarios that include all different types of guns 407 00:26:20,667 --> 00:26:24,410 being presented from law enforcement or known trainings, 408 00:26:25,090 --> 00:26:28,172 things like that. So as much as possible, we 409 00:26:28,213 --> 00:26:32,040 try to communicate with our customers so that we're aware of any 410 00:26:32,100 --> 00:26:35,481 reasons that there should be a gun on premise, but generally our customers want 411 00:26:35,521 --> 00:26:38,662 to know regardless of what type of gun it is or what the scenario is. 412 00:26:39,303 --> 00:26:42,744 And that's an awesome point of communication for the ZOC. 413 00:26:43,104 --> 00:26:47,265 So even if we see something that we believe is non-threatening, it's 414 00:26:47,646 --> 00:26:51,087 easy enough just to call up the customer and have that conversation and 415 00:26:55,103 --> 00:26:59,506 Yeah, I'd want to know. I'd want to know. You 416 00:26:59,546 --> 00:27:02,709 mentioned that your work, I 417 00:27:02,749 --> 00:27:06,752 guess, has led to a bunch of arrests of 418 00:27:07,092 --> 00:27:10,935 people who had weapons in places that they shouldn't have had them. I'm 419 00:27:10,955 --> 00:27:14,117 wondering what else you can tell me about the 420 00:27:14,157 --> 00:27:18,441 results that you've seen. Like, we've talked about you've launched 421 00:27:18,481 --> 00:27:21,823 in 2018, you've been in operation, you know, and growing 422 00:27:21,863 --> 00:27:26,318 into new places. what 423 00:27:26,398 --> 00:27:29,700 kind of situations have you come across and 424 00:27:33,122 --> 00:27:36,624 Yeah. So since 2018, we've expanded 425 00:27:36,684 --> 00:27:39,986 to, I think we're up to 47 States. We're spread out throughout 426 00:27:40,026 --> 00:27:44,469 the entire country on K-12, um, public 427 00:27:44,509 --> 00:27:48,071 transit, uh, on the commercial side, we're in, um, 428 00:27:48,772 --> 00:27:51,974 big box retail and logistics centers and things like 429 00:27:52,034 --> 00:27:55,630 that. Um, We 430 00:27:55,910 --> 00:27:59,651 haven't had that stereotypical detection 431 00:27:59,711 --> 00:28:03,132 and arrest of somebody that appeared to be, you know, entering a school to 432 00:28:03,232 --> 00:28:07,052 commit a mass shooting. But we've, you 433 00:28:07,072 --> 00:28:10,913 know, in early days, maybe we would see a gun once a month. We're absolutely 434 00:28:10,973 --> 00:28:14,514 seeing guns, dozens of guns on a daily basis in areas that 435 00:28:14,534 --> 00:28:20,135 I never thought we would see guns. And so the 436 00:28:20,175 --> 00:28:23,376 performance of the model itself has proven itself as 437 00:28:23,436 --> 00:28:26,727 we scale. We're generating detections all the time. 438 00:28:29,089 --> 00:28:34,253 And in the scenarios that we've run with local law enforcement and 439 00:28:34,854 --> 00:28:38,117 law enforcement at customer sites, we've been able to identify that the 440 00:28:38,497 --> 00:28:42,080 reduction in response time is considerable. 441 00:28:42,300 --> 00:28:45,723 So comparing response times 442 00:28:46,084 --> 00:28:49,766 without zeroized alerts, First responders 443 00:28:49,846 --> 00:28:53,589 are basically showing up at a school and not knowing where the 444 00:28:53,609 --> 00:28:56,933 shooter is located. So they enter that school, it 445 00:28:56,993 --> 00:29:00,396 could take them five to 15 minutes to clear the school 446 00:29:00,476 --> 00:29:04,100 and locate the actual shooter. And there's been plenty of real life examples, 447 00:29:04,700 --> 00:29:08,304 like Uvalde, where there's been serious 448 00:29:08,344 --> 00:29:11,470 challenges around that. So in 449 00:29:11,650 --> 00:29:15,394 the testing that we've done, we've been able to considerably reduce response 450 00:29:15,434 --> 00:29:18,556 time and direct first responders exactly where in 451 00:29:18,576 --> 00:29:21,719 the building they should be located. It's 452 00:29:21,759 --> 00:29:25,082 something as simple as if first responders show up to the wrong side 453 00:29:25,102 --> 00:29:28,705 of the building and they enter the wrong door, that could mean 15 minutes 454 00:29:28,765 --> 00:29:32,148 in lost response time. Wherever 455 00:29:32,208 --> 00:29:35,512 possible, we're really focused on reducing that response time to get 456 00:29:39,096 --> 00:29:42,700 You mentioned you're seeing dozens of guns every 457 00:29:42,720 --> 00:29:46,904 day in places that you wouldn't expect. Are 458 00:29:46,964 --> 00:29:50,546 these, what are those kinds 459 00:29:50,586 --> 00:29:53,928 of situations? Is that, are people just kind of caring and 460 00:29:53,948 --> 00:29:57,449 they're walking around and they're caring? Is there intent to 461 00:29:57,489 --> 00:30:02,312 violence? Do those always lead to kind of reactions in 462 00:30:02,452 --> 00:30:05,793 terms of law enforcement, but in terms of, you know, the customer that 463 00:30:05,833 --> 00:30:09,115 you're securing, is there responses to that or is it, oh no, 464 00:30:09,275 --> 00:30:12,616 he's okay. I mean, obviously each 465 00:30:14,983 --> 00:30:18,624 Yeah, short answer is it totally depends. But 466 00:30:18,644 --> 00:30:22,004 I would say that for the most part, fake guns are 467 00:30:22,024 --> 00:30:25,665 starting to look more and more like real guns. If you pick up an airsoft rifle at 468 00:30:25,785 --> 00:30:29,806 your local Walmart, that airsoft rifle is almost indistinguishable 469 00:30:30,026 --> 00:30:33,666 from an actual AR. And even 470 00:30:33,706 --> 00:30:37,187 more so when students will 471 00:30:37,247 --> 00:30:40,548 paint the tips, paint the orange tips black or remove the orange 472 00:30:40,608 --> 00:30:43,925 tips. So a lot of times we 473 00:30:44,086 --> 00:30:47,427 don't necessarily know if the gun is a fake 474 00:30:47,447 --> 00:30:50,628 gun or a real gun, and we have to essentially treat it as though 475 00:30:50,668 --> 00:30:53,909 it's a real gun. So I would say that's probably 476 00:30:53,949 --> 00:30:57,170 one of the more common scenarios that we see. Also, a lot of times we 477 00:30:57,210 --> 00:31:00,668 see people using objects 478 00:31:00,988 --> 00:31:04,473 like cell phones and pointing them at each other as though they're real guns. 479 00:31:04,533 --> 00:31:07,897 There's a very popular TikTok challenge that's been going around the last few 480 00:31:07,937 --> 00:31:11,060 years called Senior Assassin, which is about the 481 00:31:11,100 --> 00:31:16,386 most insensitive thing that I think you could do in nowadays climate. But 482 00:31:16,586 --> 00:31:20,088 students are bringing airsoft rifles or fake pistols, 483 00:31:20,148 --> 00:31:23,590 or in some cases, real guns to school in order to perform 484 00:31:23,630 --> 00:31:26,932 these mock senior assassinations and post them 485 00:31:26,992 --> 00:31:30,393 online. I couldn't tell you how many detections we've gotten that 486 00:31:30,413 --> 00:31:33,875 were similar to that, where we see students basically filming 487 00:31:33,935 --> 00:31:37,177 themselves pointing guns at each other in 488 00:31:37,237 --> 00:31:41,335 order to fulfill this TikTok challenge. So 489 00:31:41,355 --> 00:31:44,498 in cases like that, we always respond to the customer. We let 490 00:31:44,518 --> 00:31:47,761 them know what we're seeing. We try 491 00:31:47,801 --> 00:31:52,105 to communicate as much detail about the scene as possible, but ultimately 492 00:31:52,165 --> 00:31:55,848 it's on the customer to respond and execute 493 00:31:58,230 --> 00:32:02,163 Senior assassins, huh? Yeah. Oh, 494 00:32:02,223 --> 00:32:05,344 man. But and so you mentioned, I 495 00:32:05,384 --> 00:32:09,546 just want to, you know, clarify, right, like you haven't been 496 00:32:09,586 --> 00:32:13,287 involved in or encountered directly 497 00:32:13,387 --> 00:32:17,069 any of the kind of mass violence 498 00:32:22,222 --> 00:32:25,343 Luckily, no, uh, I think about this sometimes and you 499 00:32:25,643 --> 00:32:28,884 know, obviously if an event does happen, I want to be there. I 500 00:32:28,904 --> 00:32:32,065 want to be on those cameras to be able to detect it. Uh, 501 00:32:32,085 --> 00:32:35,546 but thankfully we've, we've not been involved in a mass shooting 502 00:32:35,586 --> 00:32:38,787 event. Um, I think it's only a matter of time though. The 503 00:32:38,807 --> 00:32:41,948 more cameras we're on, the more guns we're going to see, the more coverage we're 504 00:32:41,968 --> 00:32:45,329 going to have. And, um, I just pray that when we're in 505 00:32:45,349 --> 00:32:48,750 that situation, uh, that we're able to make the detection before 506 00:32:49,331 --> 00:32:52,668 Mm hmm. Now, 507 00:32:52,989 --> 00:32:56,432 you mentioned that you're in 47 states now, 508 00:32:56,832 --> 00:33:00,655 and you have a ZOC in Pennsylvania 509 00:33:01,716 --> 00:33:05,259 and in Hawaii. So you cover the 510 00:33:05,279 --> 00:33:09,222 different time zones. They're operational all the time. How 511 00:33:09,263 --> 00:33:12,865 does the ZOC scale in kind with 512 00:33:12,945 --> 00:33:16,287 the scale of your emplacements and your 513 00:33:16,327 --> 00:33:19,629 model? Like how many more people do you need for each 514 00:33:24,252 --> 00:33:27,794 Today, we're able to monitor the entire United States from just those two operating 515 00:33:27,834 --> 00:33:31,017 centers. I anticipate in 516 00:33:31,037 --> 00:33:34,519 the future we're going to expand and build operating centers in other locations, but 517 00:33:35,219 --> 00:33:38,521 today we're able to just basically follow a model that says when we add 518 00:33:38,681 --> 00:33:42,063 X number of cameras, we're going to expect to see X number 519 00:33:42,103 --> 00:33:46,545 of additional alerts, and that increases our headcount. So 520 00:33:46,565 --> 00:33:49,766 we follow a pretty linear model in that sense. We also try 521 00:33:49,806 --> 00:33:53,448 to staff people based on alert load throughout the day. As 522 00:33:53,468 --> 00:33:56,570 you can imagine, we get the most false positives during the times of 523 00:33:56,590 --> 00:33:59,751 day that are most active in front of cameras. So if 524 00:33:59,771 --> 00:34:04,153 you think about a school, that's the five minutes every 525 00:34:04,193 --> 00:34:07,955 hour when students are walking in between classrooms. And so during 526 00:34:07,975 --> 00:34:11,657 the day, probably 8 AM to 5 PM, we 527 00:34:11,677 --> 00:34:14,938 staff heavier than we need to during the nights and off times and 528 00:34:14,978 --> 00:34:18,800 weekends. But it's been 529 00:34:18,840 --> 00:34:22,103 pretty standard for us and the more scale that 530 00:34:22,143 --> 00:34:25,485 we have, the more different sites that we're on, the 531 00:34:25,505 --> 00:34:28,967 more data that we include from those different sites 532 00:34:29,027 --> 00:34:33,909 into our model to make it better, the more predictable 533 00:34:34,150 --> 00:34:37,371 our scaling model is. So that part 534 00:34:40,493 --> 00:34:43,754 Now on the privacy side, which 535 00:34:43,774 --> 00:34:46,996 you kind of referenced earlier when I was talking about the whole Overwatch-y type 536 00:34:47,016 --> 00:34:51,191 thing, have 537 00:34:51,311 --> 00:34:54,952 access to these cameras, but you're not watching those camera feeds 538 00:34:55,012 --> 00:34:58,513 necessarily. You get, you'll get like the frame that 539 00:34:58,533 --> 00:35:02,294 was flagged. But you, you train, like 540 00:35:02,314 --> 00:35:05,475 you just said, the data that you gather, you use to 541 00:35:05,875 --> 00:35:09,296 train the models. And so I, and 542 00:35:09,316 --> 00:35:12,957 I'm sure this probably varies customer to customer, but I'm wondering on 543 00:35:13,317 --> 00:35:17,338 the privacy side, you know, if we're, thinking 544 00:35:17,358 --> 00:35:20,979 about schools or public places, these big box retailers, for 545 00:35:21,039 --> 00:35:24,420 example, how you work on the data side 546 00:35:24,480 --> 00:35:28,380 to ensure privacy and what specific components, 547 00:35:28,440 --> 00:35:31,721 I guess, of these images that they'll 548 00:35:31,841 --> 00:35:35,622 use to train your models on? Is it anonymized in 549 00:35:35,682 --> 00:35:39,263 some way, faces cut or blurred out or whatever? 550 00:35:43,823 --> 00:35:47,893 For our training data, we try not to obfuscate 551 00:35:48,273 --> 00:35:51,934 as much as possible because we're concerned 552 00:35:51,974 --> 00:35:56,075 that it will affect the context of the model. But 553 00:35:56,796 --> 00:36:00,557 for any data that's coming from a customer, we do remove 554 00:36:00,597 --> 00:36:03,918 faces so that we're not infringing on any privacy of 555 00:36:03,958 --> 00:36:07,279 the customer. Obviously, it's something that we get the customer's permission ahead 556 00:36:07,319 --> 00:36:10,559 of time. We're very close with our customers in 557 00:36:10,579 --> 00:36:14,118 that respect. But on the data training 558 00:36:14,178 --> 00:36:17,380 side, we've, from the very beginning, tried to 559 00:36:17,441 --> 00:36:21,023 distance ourselves from any sort of biometric detection 560 00:36:21,083 --> 00:36:24,806 or recognition, specifically to address that privacy concern. 561 00:36:25,707 --> 00:36:28,830 So instead of detecting a person and 562 00:36:28,870 --> 00:36:32,453 then detecting a gun, we're strictly looking for a visible 563 00:36:32,513 --> 00:36:35,575 firearm. And all of our data that is 564 00:36:35,715 --> 00:36:41,654 annotated is annotated for that specific firearm. Avoiding 565 00:36:41,674 --> 00:36:45,015 the facial recognition piece, avoiding any sort of biometric analysis has 566 00:36:45,055 --> 00:36:48,257 allowed us to distance ourselves from privacy concerns with 567 00:36:48,317 --> 00:36:51,578 our customers. And because of that, I think our customers trust us. 568 00:36:51,998 --> 00:36:55,419 They trust that we're not looking at their live video. They 569 00:36:55,459 --> 00:36:59,761 trust that we're not detecting 570 00:36:59,901 --> 00:37:03,362 or biased in any way that would affect 571 00:37:06,425 --> 00:37:10,107 And there's also a security side as well. Obviously, there's 572 00:37:10,147 --> 00:37:13,809 security to everything we're talking about your security company, but the security 573 00:37:14,049 --> 00:37:17,751 of the model, the security of the cameras and the warning system. 574 00:37:18,671 --> 00:37:22,293 And you talked about on prem deployments. 575 00:37:23,774 --> 00:37:26,876 But I'm wondering about, and I'm just gonna throw a 576 00:37:26,916 --> 00:37:30,037 kind of stupid, like movie scenario at 577 00:37:30,057 --> 00:37:33,979 you. Would it be possible 578 00:37:34,360 --> 00:37:38,122 or if If possible, how do you mitigate for 579 00:37:38,342 --> 00:37:41,704 someone to somehow gain access to or 580 00:37:41,724 --> 00:37:46,667 hack into your system to trigger a false alarm? I 581 00:37:46,787 --> 00:37:50,189 can't, I don't know exactly how that would work, but if they were to gain access, like, 582 00:37:51,029 --> 00:37:54,391 is that a real security concern on your end? That 583 00:37:54,471 --> 00:37:58,013 your system could become compromised in a way that sends 584 00:38:01,235 --> 00:38:04,658 I would say that's probably lower on my list of concerns. I 585 00:38:04,678 --> 00:38:08,042 mean, getting hacked in general or falling prey 586 00:38:08,102 --> 00:38:11,225 to some sort of social engineering is always in the back of my 587 00:38:11,265 --> 00:38:14,468 mind and probably is for any technology leader or 588 00:38:15,709 --> 00:38:19,213 startup founder. But 589 00:38:19,393 --> 00:38:22,599 that piece of our business, it always runs through the 590 00:38:22,619 --> 00:38:26,565 ZOC. So the only way for an image to be dispatched to our customers is 591 00:38:26,846 --> 00:38:30,352 for somebody in the ZOC to actually click dispatch 592 00:38:30,632 --> 00:38:33,791 and send it out. Like I said, we have 593 00:38:33,811 --> 00:38:37,154 a really tight connection with our customers. So even in the event that 594 00:38:37,174 --> 00:38:40,996 there was an accidental dispatch, we 595 00:38:41,457 --> 00:38:45,580 would have direct line to the customer to be able to deescalate immediately 596 00:38:45,820 --> 00:38:49,142 if that were to happen. So I haven't run any scenarios with 597 00:38:49,162 --> 00:38:53,165 that specific threat in mind. But yeah, 598 00:38:53,405 --> 00:38:56,567 it's always in the back of my mind that we'll get hacked in 599 00:38:56,607 --> 00:39:00,747 some way that will either, you know, expose 600 00:39:00,787 --> 00:39:04,611 some vulnerable information to the world or like 601 00:39:04,651 --> 00:39:08,194 our customer information or, you know, in 602 00:39:08,234 --> 00:39:11,577 this scenario, you know, affect our ability to dispatch or, 603 00:39:11,637 --> 00:39:15,281 you know, send errant information out into the 604 00:39:15,341 --> 00:39:18,784 universe to our customers. So it's always a threat, but 605 00:39:21,926 --> 00:39:25,730 That's good. That's good. I've 606 00:39:25,750 --> 00:39:29,045 got a couple more for you. And the first thing 607 00:39:29,065 --> 00:39:33,608 is something that I've been thinking about, which 608 00:39:33,748 --> 00:39:37,330 is, I wonder how 609 00:39:37,370 --> 00:39:40,832 much you're thinking about or exploring ways that this 610 00:39:41,913 --> 00:39:46,475 work expands beyond gun-specific detection. 611 00:39:47,796 --> 00:39:51,278 You know, Zero Eyes doesn't have anything about guns 612 00:39:51,638 --> 00:39:55,480 right in its name. And the idea of object detection 613 00:39:55,700 --> 00:39:59,401 tied to, uh, warning systems with humans 614 00:39:59,441 --> 00:40:02,703 in the loop connected to first responders seems to 615 00:40:02,743 --> 00:40:05,984 me that it could scale to other situations. I 616 00:40:06,024 --> 00:40:09,125 mean, like you, I even think about a fire, for 617 00:40:09,165 --> 00:40:12,926 instance, in a, in a school building, the same way that, um, 618 00:40:13,807 --> 00:40:17,268 your systems can tell first responders, which camera 619 00:40:17,328 --> 00:40:21,747 flags, you know, an alert. it 620 00:40:21,767 --> 00:40:24,970 would seem to be a very advanced fire alarm for you to be able to tell 621 00:40:25,030 --> 00:40:28,333 someone there's a fire and this is exactly where it is. So 622 00:40:28,533 --> 00:40:32,177 evacuate accordingly, right? Like that kind of thing. Other threats 623 00:40:32,257 --> 00:40:35,340 beyond guns. I wonder if that's something that 624 00:40:36,561 --> 00:40:40,284 From the very beginning, we've been exclusively focused on being the 625 00:40:40,324 --> 00:40:44,348 best at one thing instead of mediocre at a bunch of different things. And 626 00:40:44,388 --> 00:40:47,771 that's served us really well up until this point in the company. That 627 00:40:47,791 --> 00:40:51,352 being said, our expertise in AI and 628 00:40:51,432 --> 00:40:55,393 object detection lends itself really well to expanding into other 629 00:40:55,533 --> 00:40:58,754 use cases like this. I 630 00:40:58,814 --> 00:41:02,395 see it generally as we are a very high-value trigger for 631 00:41:02,555 --> 00:41:05,856 an incident to start. And I would love to expand those 632 00:41:05,916 --> 00:41:09,657 triggers to cover more incidents, things like perimeter 633 00:41:09,677 --> 00:41:12,738 security and intrusion detection, health and safety, to 634 00:41:12,778 --> 00:41:16,503 your point. There's 635 00:41:16,523 --> 00:41:33,426 a lot of work being done on the retail side. with 636 00:41:33,486 --> 00:41:37,287 loss prevention. So I see there's applications 637 00:41:37,507 --> 00:41:40,788 across many verticals that touch the same customers that we do. 638 00:41:41,408 --> 00:41:44,609 I first say that we're going to continue to be focused on guns, but start 639 00:41:44,649 --> 00:41:48,150 to expand into some of these other areas that lend themselves 640 00:41:48,190 --> 00:41:52,271 well. And ultimately, moving 641 00:41:52,311 --> 00:41:55,572 into a future where vision transformers and 642 00:41:55,612 --> 00:41:59,823 large language models are more easily accessible in real time, It 643 00:41:59,843 --> 00:42:03,525 opens up a lot of possibilities. We could take this in a lot of different directions 644 00:42:04,986 --> 00:42:09,309 and ultimately try to solve as many customer value problems as possible. But 645 00:42:10,109 --> 00:42:13,511 yeah, very interested in intrusion detection, other types 646 00:42:13,551 --> 00:42:17,253 of weapons like knives and just aggressive behavior in general, 647 00:42:17,293 --> 00:42:20,915 things like that. And then ultimately we 648 00:42:20,935 --> 00:42:25,724 want that trigger to initiate some sort of valuable response. Today, 649 00:42:25,804 --> 00:42:29,206 that is us sending situational awareness to our customers that 650 00:42:29,226 --> 00:42:32,327 they can respond, but we're still sending a person to deal with a 651 00:42:32,347 --> 00:42:36,329 very dangerous situation. And so I would love to initiate 652 00:42:36,429 --> 00:42:39,950 additional response like, you know, 653 00:42:40,170 --> 00:42:43,452 locking down doors or sending a drone to verify the 654 00:42:43,492 --> 00:42:47,535 incident and just have additional eyes on, I think is a 655 00:42:47,575 --> 00:42:50,997 very likely future for us. If you can imagine a first 656 00:42:51,037 --> 00:42:54,538 responder showing up to the scene after there's been a drone there for five minutes, 657 00:42:54,918 --> 00:42:58,680 showing them what exactly is going on, they have the ultimate situational 658 00:42:58,700 --> 00:43:02,141 awareness that they wouldn't have otherwise. And given 659 00:43:02,161 --> 00:43:05,523 the dangerous nature of first responder jobs and security jobs, 660 00:43:06,103 --> 00:43:09,184 I think it's completely natural that at some point in the 661 00:43:09,244 --> 00:43:12,766 future, there will be some sort of response 662 00:43:16,192 --> 00:43:19,596 Yeah, there's a lot of places you can take it, I guess, when you start thinking 663 00:43:19,636 --> 00:43:23,021 about it like that. Sure. You know, drones, automated 664 00:43:23,081 --> 00:43:26,320 security. I 665 00:43:26,340 --> 00:43:31,223 guess you almost see the beginnings of like a Robocop type thing. 666 00:43:34,005 --> 00:43:37,448 It's definitely hard to differentiate and diverge your thinking between 667 00:43:37,488 --> 00:43:40,970 what is sci-fi and what is real life. And we get questions about 668 00:43:41,010 --> 00:43:45,153 that all the time. Total Recall type analysis 669 00:43:45,193 --> 00:43:50,439 where privacy is a thing of the past. It's 670 00:43:50,459 --> 00:43:53,542 something we're sensitive to, and as much as possible, I want to 671 00:43:53,582 --> 00:43:57,005 avoid privacy issues, because I care about that personally, myself. 672 00:43:58,926 --> 00:44:02,389 But at the same time, there's so much possibility on the response 673 00:44:02,429 --> 00:44:06,373 side to incorporate drones 674 00:44:06,433 --> 00:44:09,776 and robots, autonomous response, into areas where otherwise 675 00:44:12,253 --> 00:44:15,396 Right. I even think specifically for something like 676 00:44:15,756 --> 00:44:20,500 the Hurt Locker, like if that could be done completely with robots, that 677 00:44:20,520 --> 00:44:24,243 would be good. Setting people in the suits to diffuse 678 00:44:24,283 --> 00:44:27,465 these things is crazy. But that also raises an 679 00:44:27,485 --> 00:44:30,748 interesting point that I always think 680 00:44:30,788 --> 00:44:34,691 is interesting because like we've mentioned 681 00:44:34,731 --> 00:44:38,234 several times, you guys have been around for several years before 682 00:44:38,294 --> 00:44:41,756 the kind of boom in AI that we're dealing with now. 683 00:44:43,938 --> 00:44:48,141 And I wonder, I guess, how, 684 00:44:49,422 --> 00:44:52,945 if it got easier to talk 685 00:44:52,985 --> 00:44:57,208 to clients about what your offering is, when 686 00:44:57,888 --> 00:45:01,191 AI became much more in 687 00:45:01,211 --> 00:45:04,613 the public vernacular, And if in a weird way 688 00:45:04,653 --> 00:45:07,734 that also made it a little bit more difficult, because a 689 00:45:07,774 --> 00:45:10,975 lot of it is tied up, as you mentioned, right? With, with 690 00:45:11,515 --> 00:45:14,876 people from outside the industry who are, have a hard time sifting through 691 00:45:15,617 --> 00:45:19,058 fiction and, and, and the hype with what's actually happening. 692 00:45:20,238 --> 00:45:23,540 Yeah. Our biggest challenge in the early days was convincing people that 693 00:45:23,560 --> 00:45:28,301 it wasn't snake oil. And so, I mean, 694 00:45:28,341 --> 00:45:31,713 we ran so many demos. At that point, just trying 695 00:45:31,733 --> 00:45:34,814 to prove to people that we could actually detect guns. And then 696 00:45:34,854 --> 00:45:38,476 they said, well, you know, you're showing us you being 697 00:45:38,516 --> 00:45:42,057 detected with a gun. I want to hold the gun and be detected to 698 00:45:42,117 --> 00:45:45,519 really prove it out. And so a lot of it has been proving out 699 00:45:46,540 --> 00:45:50,044 that. AI security of 700 00:45:50,084 --> 00:45:53,787 the past was maybe embellished 701 00:45:53,867 --> 00:45:57,690 on the sales side to an extent that it caused customers to 702 00:45:58,631 --> 00:46:02,114 think that AI wasn't capable 703 00:46:02,154 --> 00:46:05,376 of performing in real time on security cameras. And we spent the 704 00:46:05,416 --> 00:46:08,479 last seven, eight years trying to change that opinion a 705 00:46:08,519 --> 00:46:13,523 little bit. And then going 706 00:46:13,603 --> 00:46:17,081 forward with the 707 00:46:17,141 --> 00:46:21,004 emergence of large language models and vision transformers 708 00:46:21,124 --> 00:46:24,306 and some of the 709 00:46:24,567 --> 00:46:28,890 common topics around them like copyright infringement, potential 710 00:46:29,570 --> 00:46:33,853 lawsuits, these large language models using 711 00:46:33,913 --> 00:46:38,356 public data in some way. It's 712 00:46:38,396 --> 00:46:41,638 caused a lot of questions about how we collect our data, and that's been one 713 00:46:41,679 --> 00:46:45,084 of our strong selling points to the customer is that All 714 00:46:45,124 --> 00:46:48,827 of this data is organic to ZeroEyes. There's 715 00:46:48,847 --> 00:46:52,130 no risk of us infringing on 716 00:46:52,190 --> 00:46:55,453 any copyrights or using any public data. It's all stuff 717 00:46:55,473 --> 00:46:58,975 that we've meticulously developed in-house 718 00:46:59,396 --> 00:47:02,819 and have scrubbed to maintain the highest quality possible. So 719 00:47:02,839 --> 00:47:06,161 I would say those are probably the two areas that we see overlaps with 720 00:47:08,062 --> 00:47:11,344 Yeah. Yeah, you operate at 721 00:47:11,664 --> 00:47:15,026 the intersection in an interesting way. 722 00:47:15,106 --> 00:47:19,409 It's different from a lot of other AI companies that 723 00:47:19,489 --> 00:47:23,111 I talk to. But my last 724 00:47:24,171 --> 00:47:28,713 point, I guess, to leave off here, is, 725 00:47:29,414 --> 00:47:32,954 you know, I wish you guys weren't needed. And it's 726 00:47:33,115 --> 00:47:36,775 interesting, you start the company, 727 00:47:37,916 --> 00:47:41,617 this team of veterans, in 728 00:47:41,657 --> 00:47:45,198 the wake of a devastating mass shooting and, 729 00:47:45,858 --> 00:47:49,299 you know, that's a problem and a crisis 730 00:47:49,339 --> 00:47:52,500 that hasn't really abated too 731 00:47:52,540 --> 00:47:55,721 much. And so here, you know, 732 00:47:55,761 --> 00:47:59,382 where you sit with a technological solution, and 733 00:47:59,402 --> 00:48:03,743 this kind of goal to, through 734 00:48:05,063 --> 00:48:09,544 detection and better response times to mitigate violence 735 00:48:09,584 --> 00:48:12,765 on the ground, whether that's someone has 736 00:48:12,805 --> 00:48:16,146 an airsoft gun, or whether that's, you know, a mass shooting 737 00:48:16,186 --> 00:48:21,427 might be about to happen, we have to get over there. You're 738 00:48:22,107 --> 00:48:25,549 the idea of the kind of technological solution to a problem that bleeds 739 00:48:25,609 --> 00:48:29,090 beyond technology. I wonder what you think about that and what 740 00:48:29,130 --> 00:48:33,252 other levers should be considered or 741 00:48:33,512 --> 00:48:36,713 if the reality of the world we live in is one where it's 742 00:48:36,753 --> 00:48:39,995 like technology is 743 00:48:40,035 --> 00:48:43,416 kind of the best thing we have left to 744 00:48:47,217 --> 00:48:50,440 I also wish we weren't needed. I think about that almost 745 00:48:50,500 --> 00:48:53,961 on a daily basis. As a company, we follow gun violence throughout 746 00:48:54,001 --> 00:48:57,102 the country really closely. And so I 747 00:48:57,162 --> 00:49:00,502 see on a daily basis the news reports of shootings and 748 00:49:00,522 --> 00:49:05,563 gun violence that happen all over the country. And yeah, 749 00:49:06,103 --> 00:49:09,184 when people come up to me and they say, how's business doing? I 750 00:49:09,224 --> 00:49:13,065 say, it's good. Unfortunately, people keep 751 00:49:13,185 --> 00:49:16,926 committing gun violence. And there's probably 752 00:49:16,966 --> 00:49:20,379 deeper issues at play there. But 753 00:49:20,419 --> 00:49:24,181 when I think about our position with the customer, we're 754 00:49:24,201 --> 00:49:28,123 providing a layer of security. And so when we go 755 00:49:28,163 --> 00:49:33,066 into a new customer, They trust us based 756 00:49:33,146 --> 00:49:36,826 on our track record and expansion to basically 757 00:49:36,846 --> 00:49:40,227 be an expert to them about how their security should look. And 758 00:49:40,247 --> 00:49:43,788 so we're able to really provide this great feedback where, you know, 759 00:49:43,828 --> 00:49:47,088 maybe a customer is struggling with their camera system and 760 00:49:47,108 --> 00:49:50,549 it doesn't even make sense to buy ZeroEyes until their camera system 761 00:49:50,609 --> 00:49:53,830 is upgraded. And so we're happy to make that recommendation to 762 00:49:53,870 --> 00:49:57,228 them because At the end of the day, our performance won't be as 763 00:49:57,308 --> 00:50:01,574 great on older cameras, lower resolution cameras. And 764 00:50:01,774 --> 00:50:06,197 it matters to us that our customers have the highest security posture 765 00:50:06,237 --> 00:50:09,679 that they possibly could. So we're in this really awesome 766 00:50:10,280 --> 00:50:13,782 kind of expert position for our customers. And all 767 00:50:13,802 --> 00:50:16,884 good security comes in layers. We're just one piece of 768 00:50:16,984 --> 00:50:20,486 it. And we're trying to address that gap of being 769 00:50:20,526 --> 00:50:23,828 that first early warning sign to communicate to 770 00:50:23,889 --> 00:50:27,131 customers when there's a visible gun, when there's a weapon that's brandished on 771 00:50:27,171 --> 00:50:33,121 their physical site. Yeah, 772 00:50:33,141 --> 00:50:36,503 going forward into the future, I see a huge possibility for us to 773 00:50:37,283 --> 00:50:40,465 be really ingrained with the customer to be that expert and to 774 00:50:40,505 --> 00:50:44,928 provide that 775 00:50:45,188 --> 00:50:48,703 knowledge that's needed to understand 776 00:50:48,743 --> 00:50:51,844 how to respond to one of these incidents. Hopefully, the vast majority of our 777 00:50:51,904 --> 00:50:55,165 customers will never experience one of these incidents, but with 778 00:50:55,185 --> 00:50:58,986 the prevalence of gun violence and how it's expanding, that's 779 00:50:59,006 --> 00:51:02,346 becoming more and more likely. And it's not as easy for customers to 780 00:51:02,386 --> 00:51:05,827 just say, you know, that's not going to happen to me. They 781 00:51:05,847 --> 00:51:09,008 have to be prepared in some way, and we're in this awesome position to help 782 00:51:09,048 --> 00:51:12,269 them prepare. So I think we're just one 783 00:51:12,289 --> 00:51:15,390 layer in their broader physical security, but we're a 784 00:51:15,430 --> 00:51:19,323 critical layer at this point. I hope I answered your question. That 785 00:51:21,483 --> 00:51:25,264 Thanks. Yeah. But from that perspective, the future 786 00:51:25,444 --> 00:51:28,885 is exciting. But, you know, again, dark 787 00:51:31,906 --> 00:51:35,307 Yeah, we as a company, we have to do well 788 00:51:35,347 --> 00:51:38,703 in order to do good. That's one of our kind of principles. In 789 00:51:38,723 --> 00:51:42,424 other words, we have to be profitable in order to expand to more cameras 790 00:51:42,684 --> 00:51:45,765 so that we can cover enough cameras in order to fulfill our 791 00:51:45,845 --> 00:51:49,025 mission, which is stopping gun violence. And 792 00:51:49,065 --> 00:51:52,146 so we live and breathe it 793 00:51:52,346 --> 00:51:55,967 at a daily basis here at Zero Eyes. And 794 00:51:56,807 --> 00:52:00,328 when people come in to work every day, they're singularly focused 795 00:52:00,368 --> 00:52:03,648 on that mission of ending gun violence. It's a beautiful place to work 796 00:52:03,708 --> 00:52:07,140 because of that, but it does come with Um, 797 00:52:12,222 --> 00:52:15,364 It's a heavy mission. Well, Tim, I appreciate you, uh, 798 00:52:15,684 --> 00:52:19,005 letting me steal you away from it for, for a little bit, um, 799 00:52:19,245 --> 00:52:22,547 and walking me through what you do. Uh, so thank you. Thank