Real-Time Visual AI for Utility Operations.

Hi, pressure to move faster while maintaining accuracy across design inspection as-built documentation and emergency field workflows. Today's session looks at how real-time visual AI embedded directly into field execution is helping utilities close that gap. We'll explore how IQGeo's geospatial work execution approach turns field photos into intelligent actions, combining AI-driven image analysis with live network context and mobile workflows. We'll see how utilities are using these tools to improve data quality at the point of work, verify installations automatically, update network models in real time, and strengthen grid resilience through connected, adaptive workflows. This webinar is designed as a practical preview ahead of DTEK, highlighting how Tier 1 utilities are already applying AI-enhanced field execution in daily operations. Before we get started, a couple of housekeeping tips. This webinar is live and interactive. You can see how submit questions at any time using the Q&A box, which is also the best place to flag any technical issues. We also recommend that you shut down any other applications you have open at this time so you can fully concentrate on the discussion you're about to hear and improve your machine's performance. This session will be recorded and available on demand. A follow-up email with the recording link will be sent to all registrants. Without further ado, here are the speakers presenting today from IQGO. Adrian McNulty is Vice President of Utility Solutions at IQGO. He brings more than 25 years of experience across the energy sector and works closely with utilities to modernize field operations, improve data accuracy, and strengthen grid resilience. Thomas Trulier is Head of AI Solutions for the Americas at IQGO. He leads efforts to bring real-time visual AI into utility and telecom field workflows, helping customers translate field requirements into measurable operational improvements. Full bios are available in the presenter tab. I'll now toss it over to Adrian to begin today's presentation. Adrian? Hey, thanks, Kevin. Good morning, good afternoon, good evening, everyone from wherever you're joining. We're excited to present this topic today. Like Kevin said, I've been in the utility industry for the past 25 or so years. I've been really lucky to work in everything from generation, transmission, distribution, and technology has been really evolving. And I'm sure AI is a hot topic for all of you today. So today we're really going to cover a couple of things. We're going to really talk about how we're using visual AI to enhance field workflows. But even before we get into that, we're going to talk about what AI is and the different flavors. So more of a learning educational steps as well. With me is Thomas. So Thomas, would you like to introduce yourself? Hey, Adrian. Thank you, Kevin. So I'm Thomas. I've been working on computer vision, visual AI for over eight years as part of a company called Deep Homatic that was acquired by IQG over the summer. And very excited to show you what we've got, what we've been working on, and what we will demonstrate at DTEK. Great, thanks, Thomas. And before we jump into it, so I've been with IQG for the past four or five years now. And about 12, 18 months ago, we were looking, we realized we wanted to add AI into our solutions. We felt it had significant value for field workers. So we were looking for a solution set that, first of all, was scalable. So we worked with some of the largest utilities and telcos in North America across the world. But also, we wanted to find a solution that aligned very closely with IQG. And IQG, we focused specifically on telcos and utilities, and in the utility space, specifically in the field workflows, putting applications in the hands of the field users to execute the work. So we're quite a company called Deep Homatic where Thomas came from. And there was a very good alignment in terms of sector, scale of AI, and actually the use cases. We're going to go through these today. Let me just proceed with the slide. So first of all, here's our disclaimer. I have to include that. Looking forward to seeing you all in San Diego at beginning of February. I think you'll all agree, San Diego in February is a good place to be. We'll be at booth 1245. We'll have a number of different demonstrations there, ranging from the base IQG of geospatial solutions, which we'll explain in a second, as well as the visual AI. You can reach out to myself or Thomas. You can use the DTEC app. I'd love to see you there to make an appointment or even pop by and say hello. Looking forward to seeing you all there. So before we jump in, just a quick introduction of IQG, and I'm going to keep this to one slide. So we are focused, as I mentioned a slide or two ago, we're focused specifically on utilities and telco. So we build a geospatial platform that covers the lifecycle of the network. And that ranges from everything from design to as builds to everything you do once your network is operating, asset inspections, planned work and unplanned work. And then visual AI layers across these different workflows. Before we jump into the actual AI content, I was thinking today, like we've been going through constant change and innovation across the 25 years, at least I've been in the utility space. So I wanted to throw a question out there around what do you guys think of innovation? So here's the first poll question. I'm going to read it out. What is the first thing that comes to your mind when you hear the term innovation? Is it do more with the same resources, leverage new technology to solve real business issues, smarter use of data, improve safety and compliance or other? And really, what I'm trying to get us thinking about here is AI isn't necessarily something new, even though it is new technology, we've been going through innovation changes for the last 25 years. And this is just a new flavor, a new technology. But it's important we think about what is what are the outcomes that we gain from applying new technology to our use cases. It looks like some of the results are coming in, we'll give it another minute. So again, what is the first thing that comes to your mind when you see they think they hear the term innovation, do more with the same resources, smarter use of data, improve safety and compliance or other? I think we're ready to push the results to the audience and see where we ended up. All right, so I think you should all see the results now. That's kind of what I thought we'd see. So leverage new technology to solve real business issues we have, and a smarter use of data. And that aligns with the way that we see visual AI, AI in general, is it's how do we apply it to the work we do today to make it easier to do more with what we have today. And that's going to be the topic of today's today's presentation. But kind of emphasize that the journey that I've seen and the dates aren't necessarily important here is more as the change that we've seen over the last 25 years. And this kind of replicates my career in the utility industry. You know, back in the late 90s, we started to digitize our networks with the first enterprise level GIS is and SCADA taking SCADA beyond the substation fence, you know, moving away from mimic boards to full SCADA systems. You know, in the early 2000s, early 2010s, we had the mobility and smart meter rollouts. I remember the 2010 ARRA that really accelerated the smart meter rollouts and the early ADMSs. Then we moved into what I call the operational area. This is more ADMS at scale, and we start to see renewable applications, the early DERNs applications, and then the data age, more advanced integrations across systems, IT, OT integration, and advanced analytics of the data that we're retrieving across systems. And now I feel we're in the age of what I'm calling the AI guided era. So how do we take AI to guide us in the work that we do? And that's going to how the focus of the rest of this presentation is going to be. So at this point, I'm going to let Thomas take over. Thank you, Adrian. So I want to cover quite a bit of ground with you today about a little bit about AI, the different types of AI, and how it's utilized for field work, mostly through the use of computer vision. But I also don't want to just leave you hanging wondering exactly how that can be applied. So we decided to flip the script a little bit. And I will start by actually showing you a first example of our technology being applied. So you can have in your mind a visual representation of everything we're talking about. So let me go ahead, I will go to the next slide, start a video that I will comment, a recording that we made for a specific use case live on inspecting poles and pole assets. So let me press play. So you should now be able to see that video going. This is a screen grab from my iPhone. So you can see we have a mobile application that allows user on the field to capture data. I don't have a poll in front of me. So I upload a photo I took earlier in the field. And you can see that we detect different types of assets. So here is really to look at insulators. I can detect the different types of models. I can count how many there are, and then translate that information into output in terms of does that asset need to be replaced or is it okay. Now upload a closer take of one of the asset to show you what happens when we want to look more closely of the characteristics of a specific asset. I look at the number of skirts, the different elements, visually helping me determining what is that asset. A different example would be looking at joint use. So first actually taking a photo of an entire pole and kind of get an assessment of what's on it in terms of the different equipment. And then looking at joint use. And so we can look at what's attached on the pole, equipment, lines, and translate that back into your system and give that information through the mobile application. So here you see I can have the neutral line, we have different attachments, and we can measure the clearance. So this is a quick glimpse into the interaction of that technology on the field. And once that is done, I have a second video to show you where basically we centralize all that information in the back office. So you can see here I'm on my browser, and you can see the photo I took earlier through that demonstration and all the information that were captured. The number of assets, the type of asset, if something was detected as suspect. And those are all the different jobs that you might be responsible for as a supervisor, as an area manager, and basically control what's going on in the field in a central way, and then translate all that digitalized data into the rest of your system. So if that was a lot, don't worry, feel free to ask your questions in the Q&A, think about what you just saw, and tell me if you want to see anything more, and we'll be happy to delve more into that. And don't worry, I also have a second video to show you with the different type of use case. So keep all that in mind. While I go into giving you a little bit of a broad overview of what we call the different AI flavors. So that might be a little bit basic for some of you, good reminders for others. So I'm not going to necessarily spend too much time. But I just want to make sure we all understand what type of technology we're talking about, and do a quick refresher on some terms and vocabulary. So you might have heard of the term machine learning. This is basically a technique that now is quite established. It's almost all at this point. And it's really the idea that you can have all algorithms that learn with new data. So it's not a set program that has an output input and does not evolve. It's one that we're able to change over time based on experience. So it's a pretty broad umbrella term. In that is deep learning. So deep learning is a more advanced method, where we basically literally took inspiration from the structure of brains, and created what we call artificial neural neural network, which are just layers and layers of a very, very large system that learns just like in machine learning, but can learn much more complex data, such as visual information, to analyze what we see in photos and in videos. And then we have Gen AI. That's something that has been come up in the forefront of our discussion with ChatGPT, Gemini, etc. Gen AI utilize the same systems in terms of a large neural network, we call them the large language models. They ingest an enormous amount of data in order to be able to generalize concepts based on language, and then not just synthesize what is processed, but generate new data based on that assessment. And finally, the last one I'll talk about is the AI agent. That's the newest version of utilizing all this different technology. And it's basically going from processing the input, generating new data to then taking action. So it's actually giving control to those agents of certain digital actions, tools, etc. to automatically evolve in our day to day work environment. And so how does the IQgeo AI platform function? So having said all of that between the different technology, one key aspect that I really want to emphasize is that we won't rely on a single large model such as you would imagine ChatGPT processes. This is not the approach of IQgeo. The IQgeo approach. I'm sorry, I'm getting some feedback noise. So maybe it's one of the presenters. But anyway, I'm going to keep going. We don't rely on a single model, we build very specific deep learning model to handle individual tasks. So in the video you saw earlier, we can detect different asset on the pole, we can detect if the different visual elements of an insulator are giving you information. Those are specific models that are trained to do just that. So not just taking a photo and giving it to one model to give me something, but a workflow of a lot of different models. And the benefits of doing this is that we get much more accurate results. We don't just ask ChatGPT, what do you see on that photo? We actually look for specific things with our specialized model. It's much faster because each of those models are much smaller, they process the data very quickly. It's adaptable, I can bring in more model in my workflow, I can remove some, I can tweak the result. And overall, that makes it cost efficient, easier to deploy, uses less energy. And finally, we use this technology both in cloud deployment, but we also have developed the capacity to put them in phones, in mobile devices directly. What that means is we're able to process the data, give the feedback to the field, even without any connectivity. So now I have a new, oops, sorry, I realized I did not pass my slide. I have a new poll for you. So I should be pushing that poll to the audience. I will read the question, you will see it soon. So where's your organization on its AI journey as it relates to field operations. So option number one, using AI in production today. Or you might be running pilots and proof of concepts. You might be at the research stage. You're interested but cautious. So please feel free to share. We're interesting to see where you stand vis-a-vis this very new technology, especially since now we are seeing there are different levels of deployment, machine learning, deep learning, gen AI agents. So how much commitment is there to this technology? How far along are you on that journey? So I'm getting first responses. I will push it to the audience very soon. Let me do that now. All right. So you might be able to see right now some of the results. Obviously, what's interesting and jump out right away is that nobody answered, not on the current roadmap. So everybody's looking at it. A lot of caution. But it's something that we're talking about. So my goal today is to make sure that you understand the different ways we can deploy this and the level of risk that we're taking in that deployment. And hopefully, it will match your understanding of the technology. Let me see if we're getting a couple more responses. All right. So more emphasis on the research stage. Great. Thank you for your responses. So if we focus a little bit more on our technology and its usage by our customer, I will talk about. So what is computer vision for large distributed asset? A word about where we are and why you can trust us. We have been working. So first at Diplomatic, but now at IQGO on computer vision for over 11 years. And that's been the technology we've only worked for all this time. And in the last couple of years, so about five, six years, we've focused on deploying that technology exclusively for large distributed assets, meaning telecommunication networks, as well as utility networks, gas, electricity and water. And last year, we reached over 30,000 technicians on the field using our technology every day. And that represented the volume of over 20 million work orders processed. So I put a couple logos of our different customers. This is just a small sample. And we emphasize that those deployments are the result of an enterprise ready deployment in terms of data security, data management and access. Our access and deployment of the latest technology. So deep learning and gen AI, as well as our capacity to implement this technology quickly and realize real world financial benefits for our customers. And so, as you saw in my initial video introduction, computer vision is that branch of artificial intelligence that process visual imagery. So we look at what is in photos and videos in order to detect elements, make assessment. Is this a good element? Does it require maintenance? Does it require repair? As well as measurement distances for clearance, for example. Small list of examples for the utility market. We have customers working on meter installs. So that's one of the next demo I will show you. So detecting equipment on full, inspecting quality of asset, expecting damages. And so the rule of thumb that I put at the beginning at the bottom of that slide is something to keep in mind. When you're working with a technology like computer vision, if you can see it in an image, if an expert can make an assessment just based on looking at the image, then so can our algorithm. If you can't see it on the image, then the flip side is obviously the AI will not be able to see it either. So if you have a process currently that involves looking at visual elements, either by being on the field or through photos, that is a process that could be digitized. Not necessarily to fully automate it, but at the very least to extract the physical information from the field and put it in a digital format. And that's the reality of the operation of all our customers. So the reality of the way that we're going to do is by definition, the distributed asset nature of utilities network mean that you have elements that are difficult to reach, that can be far away and not always accessed. And because you have a workforce that work on this asset regularly, it's the opportunity to crowdsource information that will be useful for the operation of your network. So most of you probably already have a way to capture those photos. I know most of our utility customers do and did. And that is an opportunity to centralize this information in order to get a different type of visibility over your asset. But because of the nature of the infrastructure, you get large amount of photos, you get large amount of data. And so what we're seeing is that it usually impedes our capacity to make that data useful. You might be able to centralize. Maybe you don't even centralize it. But the only way to make this large amount of data actually useful is to at least automate the processing of that data. And so that's what the computer vision, the real-time visual AI that IQgeo is providing becomes useful. The moment someone takes a photo in your network, we will be able to process it in order to digitize the information in it and even make sure it's been taken properly. And this is why the real-time analysis on the field is really important. Today, when someone captures information or a photo on the network, it might be that that information is not captured properly or the photo is not framed properly. So now imagine that when a technician takes a photo of a pole, a transformer, a trench, a customer's meter, for example, we'll instantly be able to give them feedback on, well, did you capture that asset properly? If they only took the top of the pole and were interested in seeing if the pole is leaning, for example, we'll be able to flag it and explain, look, we need a full photo from top to bottom. So here is an example. Please retake that photo. What that means is it ensures that the documentation of your asset now becomes accurate. And you can actually trust all the data that comes from the field and know what percentage is useful and detect problem. The next level is when they are performing work, when your field resources are installing a new equipment, doing repairs. By having the real-time analysis, we can also flag problem discrepancies, mistakes, things that were forgotten so that they can fix it while they're still there. A common example might be things as simple as forgetting to close a handhold properly or putting a label on an asset that down the line might be important to have or read an asset number, for example, instead of inputting it manually. So if we detect that something is missing, we can flag it, explain what needs to be changed and give the opportunity to the technician to address it. And so having done that, you increase the rate of first-time write. We talk about first-time write operations and ensure the visibility over your asset because now everything that is captured and done in the field has been already processed in a systematic manner. And what's the result of this approach? It allows you to now have a visual and digital record of your physical asset. And so when I talk about digital, I don't just mean having a photo in a digital format. It's not just about passing from paper to digital information. It's extracting the visual information from the photo that's inherently still very analogous and putting it in a digital format. So not just a photo of a pole, but what's on that pole? How tall is it? In what condition is it? And now once you have all this information, you can input that into your systems to run reports, do back office analysis, actually pull information before next work on the same asset to see what state it was. And those are functionalities that we have on our platform at IQ.GO. But you might also have in different tools. And so the data that we extract from the real world is something that you can utilize in your system. And to wrap up and really drive the point home of the importance of the type of data you digitize, I just want to summarize the three broad way to understand what type of data is analyzed through computer vision and the IQ.GO platform. So number one, when I will talk about context validation is was the photo taken properly? That is what is going to be the basis of your record. So did the photo correspond to the expectation? Is this the asset we wanted to see? Was it framed properly? Is it not too dark, too blurry, or does the sun shine in the background and we can't see anything? So the first layer is making sure you're getting right visual data. Number two is the job conformity. As I explained earlier, is making sure that nothing was forgotten. Any work that was done with photos that might be a before and after the work, we ensure it's been conducted properly. So we capture the job conformity. And finally, we have this broad term of reality capture. It's anything that we actually will not necessarily provide feedback to the technician to because they might not be able to do anything about it, but information that we want to know. So, for example, grid degradation of an asset that a fix cannot be done right away. We extract that information and pass it on to the rest of the system in order to maybe have a review and schedule maintenance work. It could be vegetation encroachment. It could be reading values of assets just to put that in a record and do inventory. So basically, again, digitizing your environment, your physical environment to create that digital record. So now I will show you one more demo. It's going to be a different type of workflow. The first one was on Pulse, if you recall. This one is going to be on the installation of a gas meter. So let me start the video. You should be seeing it. It's going to be popping up on your screen. And you will recognize here, this is the mobile application I was showing to you earlier with the different requirements for that installation. So now I'm taking a photo of that gas meter before. You can see the different tasks. So the tasks are that conformity validation and job conformity. So I take the first photo. Some of the information is already processed. I will jump a couple of the photos not to take too long. And to give you relevant examples. So this is the leak test with a soap test story to detect leaks on the gas meter. That's just to give you an example of things that can be done based on expertise. It's visual. It's visual. So that means we can train in the eye to recognize it. This one fails. We can see that one of the joint has not had the soap being applied to it. And we also detect a potential leak. So as a technician now, I have the opportunity to perform the task better and improve the quality of the job. So I retake the photo. Sorry, I must have run the photo twice. So you can see that now we have detected in the second place. So you can see that now we have performed the test properly. So now we're talking about doing. That needs to be performed by a field resource is done properly before they move on. And that includes potential safety checks like this one. Again, if you want to, if you think about other examples that you would like us to explore, feel free to put them in the Q&A section. I would love to know where you see this technology potentially being useful or if there is any comment in terms of what you would want to see being done differently. That way we can address it towards the end of the presentation. And even if we see you at our booth at DTEK. So now I will go a little bit deeper into our capabilities for doing computer vision on the IQgeo platform. And so I will step back to that initial definition I was providing on deep learning and Gen AI. So those two forms of machine learning are useful for computer vision. So what's the difference? The deep learning is going to be the technology we use to have extremely high accuracy, even on complex tasks. And that's also the one we're able to deploy offline completely. This one requires training. That's really the technology that you utilize to give something extremely specific without any room for interpretation. It's a black and white understanding of what's the standards, what's good or bad. And we can rely on a technology that will be extremely accurate and consistent. So that's deep learning. Then you have generative AI. I call that the vanilla generative AI, basically what's off the shelf. This is if you took a photo of your environment and really gave it to just ChatGPT, just Gemini, something that is what you've got access as a general customer. The strength of generative AI is that you can prompt it with text. So you don't necessarily have trained models, but you can send a photo and ask a question with English. What do you see on that poll? Do you think this is an asset in good condition? This system will probably do okay, but it's not going to be excellent. And the reason for that is it's trained on, you know, the entire Internet and understand language very well. It just doesn't understand your industry that well just because this information is not broadly accessible and there's room for interpretation. But it's pretty good for things that the general public can understand. Is there a label on this asset or how many lines are on this poll? This is the kind of question you can use with vanilla generative AI. And you also just get the average consumer SLAs. And now what we've got is more of specialized generative AI. So those are models using the same technology that are integrated into the IQgeo platform. And it's specialized because in the same fashion as deep learning, we have provided industry specific and knowledge into that generative AI model in order to have a deeper level of understanding. So now when you ask, well, what type of transformer this is, it will be able to understand it. It's still not as good as deep learning, but there are some advantages to it. And the advantages is to combine both technology. With deep learning, you have the capacity to deploy a system that will be extremely consistent with high accuracy to do those tasks for which you have trained the models. With the specialized generative AI, it's a much quicker setup. So if you want to test certain processes, you want to think about new workflow for documentation of your work and digitization of your asset, GenEye might be a way to actually test it out. It's set up in minutes. It's available. And you will be able to tweak the questions you ask it very quickly. And if you realize that, well, this process actually has a lot of value, I just wish we could go beyond a 90 plus percent accuracy and get close to 100. That's when we switch to deep learning and train those models specifically on your standard. And so where does that apply? I've shown you a lot of example based on photo taken by the mobile device by people directly on the field. But that's not the only source of data available. We can also search existing database, so photo you currently have or database of polls, for example, such as Google Street View. But it could also be a batch upload of data taken from a drone providers, for example. So don't think that we're limited to that field approach is the one that we've seen provide a lot of business improvement. But in terms of documentation of your asset, that's other areas that you have to keep in mind. So the rule of thumb still applies. If you can see it in the photo, if you can see it in the video, we can use computer vision as a technology to process that data. So a couple of examples for you to consider. This would be, for example, using existing catapult data or IECS data. So, you know, the system where you would take photos of the poll in order to do precise measurements. Those are type of data we can process if you have those already available in order to provide additional results or using aerial imagery in this case to detect visual encroachment. With a system that's been trying to process that visual insight, it means you can run hundreds of miles and get the results in a matter of minutes instead of potentially having to spend a couple hours a day in order to process that data and get the result that you want. And now we'll give you a couple of additional workflow example. It will look somewhat similar from what we showed earlier. It's just to keep, you know, giving you examples to think about how to apply this to your work. This is a full end-to-end workflow for installation of a water meter this time. And what's good with this view is you can see that it's really a step-by-step process. So, when someone installed an asset, there might be the need for understanding the environment before the installation. What's the state of the previous asset? Potentially extracting values such as the ID of the asset, any readings. And then after the installation. So, you can really compare what was before, what is after, and ensure it's been done properly. And finally, you can have safety check. Is this handhold closed properly, etc.? So, you get the idea. And so, what's important to think about when considering deploying that technology in the field is the workflow. So, when do we capture the photo? What photo do we need? And what is interesting in the photos that we're capturing for us to digitize and for us to provide the feedback to the technician who's doing that work? I have a quick example, a quick, sorry, case study from one of our customers deploying such meters. So, in Belgium, for the past couple of years, that's the one, that's the reason why I chose this one, is they've been using our computer vision technology for over two years. And they're deploying a couple millions of smart meters, both electricity and gas. And the results have been extremely positive. We've reduced the need for manual quality control by 80%, which means that now a supervisor can actually focus on different type of work. But more importantly, we've drastically increased the rate of successful intervention, catching some potential safety problems, and ensuring that no need for revisit, reducing the need for revisit for that customer. Here is an example of the asset they work on. So, those are the gas meter and the soap test that you saw me do earlier for those installations. And you can see here a couple examples of the type of elements in the visual data that we are extracting. Another one is a deployment that we're doing in Ireland. Same type of deployment. What's interesting with this one is it reveals that depending on your circumstances, increasing the documentation and the speed of processing of that documentation is something that can have pretty drastic impact in your operation. In the case of Circe, in Ireland, it's a matter of increasing their output by a significant amount because now the technician were able to move from one job to the next faster without having to wait for the validation that their job was conducted properly, which is something that they were required to do by the regulators. And so, that speaks to the quality of the analysis that we do because the AI results that approves the quality of the job is something that is acceptable in that context. And so, this is a look at the type of equipment that they're installing. This is an electric meter. So, we make sure that the wiring is done properly, the asset is sealed and not tempered with, and properly labeled. So, that's it for the overview of what IQgeo does. I'm really excited to see if we have some questions and happy to answer them. Again, we'll be at DTECH. I think we have one more slide. I'll pass it on to you, Adrian, in a second. But very excited to hear about what you would want us to be able to show. And if you have any additional questions to answer, please, happy to stay on and make sure we answer all those questions. So, thank you again very much, Adrian, passing it back to you. Thomas, could you push to slide 41? I'm not sure I've still got connection. All right, it's done. All right, so, thank you, Thomas. Thank you for going through the different flavors of AI and some examples. Hopefully, everybody found it informative. Like I mentioned earlier, we'd love to see you at booth 1245 at Distributech the first week of Feb, so here's the layout. You'll be able to get your steps in walking across the exhibition hall. We'll have a number of different demos available. We'll have the visual AI demo in a phone format, so you can try it for yourself. So, we'd like to encourage you to come, try it yourself, see the other demonstrations of IQgeo. We'll have the geospatial utility platform, different use cases ranging from general GIS mobility, different types of planned and unplanned work, for example, asset inspections, outage mobility. We'll have all of these different, you can actually use the software yourself. Use it on an iPad, use it on a phone. See what experience your users would actually see in the field. So, I think that's really important to understand the impact that it has to field users. So, I think we've got a poll question to close. So, I think I've got a code. Thomas, could you advance the slides? I'm having some technical challenges. Yeah, absolutely. No problem. So, I put the slide. Do you want me to read it or do you see it? I can see it now. So, like we've mentioned, we'd love to have you drop by booth 1245. So, the poll question here is, would you like to give it a try in person? So, yes, I'll participate during DTEK and you'll like an appointment. You'd like to do it, but you're not at DTEK, so we can follow up with you or no. And I think the value of this is actually seeing how it actually works in real life. You know, looking at different photos, different use cases, how it actually works in the field. Because one of the things that we get asked a lot, Thomas and I, is, you know, how do you start the journey on AI? I think it's really important that you will look at this and it adds value to existing work. It's not something that's going to fund it. It's something that helps us solve business problems rather than just a new technology. Many of you are already taking lots of images today, different workflows. What do we see in the results right now? Thomas, you see the results? I'm having some refresh challenges. And gentlemen, we can also leave that question up for a little while if we want to while we go through some of the Q&A. Up to you, unless you had some more thoughts. But I would say maybe we just dive into the Q&A. We've got several questions that have come in. Please continue to send your questions in as we still have a little bit of time left in the webinar. But I'll just start with this one question here, gentlemen. How can this be used to predict asset deterioration over time and determine failure risk? So, Adrian, Thomas, whoever wants to take that one. Yeah, I can start, Thomas, because this is an interesting topic that we've been asked from two of our existing clients. So within IQG, we do quite a bit of asset inspections on both distribution and transmission and gas assets. So it's different workflows using the geospatial platform. And some of our customers ask us, hey, one of the first step is, can we build visual AI into that workflow as it is? Because they're already taking images. And then the next step is, if we build this repository of images over time, we're going to have a cycle. We're like, let's think of what the cycle of the asset inspection is. We're going to have in the back office all of those different assets. So we can train the models to identify, like the question asked, what is the deterioration? For example, if we're looking at corrosion on an asset, how is that increasing over time? And we can put values against it. To think of a wooden pole, whether it's a woodpecker hole or a crack or the leaning of a pole, for example. We can model that change and put parameters in where it becomes an issue. And then you've got to take actions on it. So I think this is actually a very real, it's a very good question. It's a very real use case. And it's one that we're hoping to solve with some of our existing customers in the next 12 months. And we have a customer who do already a version of this. So for equipment where there is a high touchpoint rate, so where technicians will work on that asset repeatedly over time, we're able to give a health score based on a whole range of visual attributes on that asset. And so depending on the evolution of that health score, we're actually able to trigger maintenance work. And based on the historical analysis of the equipment, then we can extrapolate after a certain amount of time, degradation rate, and make prediction on the degradation of other assets as well. All right, terrific. Moving on to the next question. I see one here. How does this integrate with existing field tools? Thomas, you want to take that one? Because you get that question a lot. Yeah, absolutely. So everything that you saw in the demonstration I was showing, this is based on our mobile application at Diplomatic. All of our customers, there is not a single one who does not already have different tools, different field tools, or tools where they collect the field data. And in most of the time, we will integrate Diplomatic in that workflow. So it's not necessarily for you to think that implementing that workflow requires a new tool to be pushed to the field where they might already have many to juggle with. We will integrate into those tools so that it's part of the flow already existing. So if your technician currently take photo, they don't have to change the way they do it. We integrate the UX and the process directly there so they get the feedback and we don't slow them down. Yeah, let's think about it. Let's say it's coming in through a work order and that work order is in a mobile application. That mobile application, then Diplomatic would launch from within there and then close back into that existing application. If the utility is using IQG in the field, we productize that integration layer, both how you launch the application, but also the data that comes back into IQGS that comes out of the box integration. All right, great. Moving on to the next question here. Where is the data stored and how do you handle security and data residency? Yeah, I'll take this one. So we have servers in the US and Europe. So depending on where you are, the data will be onshore. Primarily, the company I come from, so Diplomatic that was bought by IQG came from France. So we respect all the GDPR regulation, extremely stringent data privacy and protection laws. And that's what we apply across the board. And so onshore and GDPR level of protection. And we also certified, et cetera. So any questions around security, we're happy to share information there. Terrific. This next one that came in, what happens when the AI isn't confident or detects something ambiguous? So that's a very good question. And that question really relates to the way AI works today with Gen AI, right? So that can happen a lot. With deep learning, we actually have a very precise idea of the result of the AI. So when the deep learning model says, I see an insulator there, it's not just saying I see an insulator there like ChatGPT would tell you. It's actually telling us, I'm 82% sure that this is an insulator. That's what we receive. So we have techniques in order to decide what's the threshold in terms of confidence to then give an output that's black or white to the technicians. But we can work on those thresholds. So we are never blind in terms of what's the level of confidence. Now, what happens if the AI is wrong? So let's say the AI says, I'm 85% sure that's an insulator. We say that's good enough. We say that's an insulator. But it's not. That's a pretty extreme example that wouldn't happen. But let's say that's the case. We have ways to finding this out. There's also interaction with the technicians on the field, which means that if they don't agree with the result of the AI, they can actually correct it. That flags it for us. Potentially, it's information, especially at the deployment, at the beginning of the deployment of deep learning models, makes surface photos that have really high level of information in them. If the AI gets it wrong, it's important to extract that photo and retrain it with it. And finally, we can also have situation. And that's more common and very important in operations where the AI is right. It takes a photo of situation and flags something wrong. It could be that the technician has no other choice but that proceed this way because of constraints in the environment. What happens in this situation is we can also have the ability to leave comments and explain why the situation is as it is. And that's a very good way to close the loop and understand the entire environment. But just be assured that, like I said, deep learning, we have very good understanding of what's the accuracy and what's the confidence level. So we are rarely in a situation where we actually give a result and we don't know why and we actually can't address it if it's a little bit off. Someone does ask, kind of related to that, how accurate is this in real world field conditions, especially with poor lighting or bad photos? So that's another excellent question. And I might take a slight detour to answer it because it reveals something important. It's the importance of using your context data to train the model. You talk about poor lighting. That might be because your specific type of work is in these conditions and might not be for somebody else. Often, for example, would be weather. You don't have the same photos of Paul with someone in Alaska than someone in Texas. And so the reason why the AI performs so well, it's because we train it on context specific photos that corresponds to your asset. That's how we address this. Now, how accurate is this in real life? All the system we push will depend on the requirements that the customer give us. So what I mean by that is when we are checking something for safety, like a gas leak, for example, we will, to go back to that idea of a threshold of confidence, we will put a threshold that's extremely high. So what happens is we're sure not to miss a problem. The flip side is we might get a couple of false positives where we say your sub test was not done properly. It was done properly, but we just want to make sure never to miss one. We can be in other circumstances where we're saying we can have one or two problems out of a thousand when the AI says something wrong. It's not going to be a bigger deal. And what's the biggest concern for us is not to flag things more often to technicians so they're not slowed down. That means we'll take a threshold that's slightly lower. So at the end, what that means is the output, the business output, when we say this is what we see, we are overing around 99 plus percent. That is what we get across the board. And if we can get, if we need to be at 100 percent, we push the threshold higher, meaning we get a couple of false positives, but we never miss a problem. So that's how we address this in the real world, because we understand that the situation where we're deployed requires us to be confident in the result that we give, not just precise, but actually be confident in the precision that we provide. And Thomas, just to add to what you said is one of the first steps that you hope you guys saw in the demo was the applications checking the conformity of the image that was taken. So what we see, if we look at a sample set of images that are just taken today in other applications without applying them into AI, there's a high percentage that don't have the quality. They don't have the right focus or the right blurry or there's some flash in it or they don't have the full image of the asset. So that's the test that goes in. That's the first conformity test. But also as the crews there, there'll be a sample of what they should be taking the image to look like. So a trained behavior of how to take the images as well as the accuracy of the model that we built. Well, Adrian Thomas, I thought this was tremendous, a great presentation. And I just wondered if you wanted to leave folks with any concluding thoughts today. I know they'll be able to reach you in a couple of weeks at booth 1245 in DTECH in San Diego. But anything you want to leave the audience with today? I just that this is an emerging technology that we've seen a lot of excitement from existing customers and new IQgeo customers. And like you just said, we'd love to see you in Distributex. Seeing is believing, right, especially with visual AI. So I offer the opportunity to come and put the application in your hands and see how it works firsthand. I think you'll be very pleasantly surprised. All right, tremendous. Well, I want to thank the panel. I want to thank the entire IQgeo team and, of course, all of you for joining us today. This presentation will be archived and can be accessed on the Factor This website if you missed it or missed part of it. A reminder email message will be sent to all registrants, complete with a direct link to the archive as well. You should also see a survey link appearing in the webinar chat. We would really appreciate your feedback there. We look forward to bringing you more webinars in the near future, and we will see you at DTECH in San Diego in a couple of weeks. Have a great day. Thanks, all. Thank you.