AI is reshaping fiber construction.

Welcome to Bite Size Fiber, the podcast where we break down the biggest challenges and best practices shaping the future of fiber and telecom networks, one bite at a time. I'm your host, Ellie Pulse, and today we're diving into a special episode all about AI and fiber construction, specifically how computer vision is changing the way we plan, build, and validate networks in the field. Joining me are three fantastic guests. First up, Clara Soslosky, a computer vision specialist who helps organizations implement AI-driven workflows to improve quality control, speed up builds, and make right first time a reality. Next, Greg Ross, with over 35 years of experience in telecom network design, build, and operations. Greg brings a wealth of hands-on knowledge about how technology can transform construction workflows in the real world. And finally, a regular guest of the podcast, Stefan Schneider, who helps telecom operators to help build a new world. Operators modernize their processes with mobile-first integrated tools that empower field teams and improve operational efficiency. In this episode, we'll explore what computer vision really is, how it's being used for real-time validation in the field, how it can prevent costly rework, and why it's as valuable for subcontractors as it is for operators. Let's get started. Clara, since you work with Diplomatic, can you tell us what exactly is computer vision? Computer vision is essentially a visual AI that can see and interpret images by automatically detecting or measuring or even verifying objects on it. In telecom, it works by analyzing photos taken by field teams to assess the quality of their work in real -time and extract data about fiber network assets. To perform this analysis, computer vision uses deep learning. And this is no science fiction, it's just science. These are AI models which need to be trained on data sets of images to learn how to recognize specific components and patterns at the beginning of a project. How have you guys seen this in your fields being used? Computer vision comes into play all along the fiber rollout process, when field workers are either digging trenches, installing cabinets, or placing distribution points on poles or in manholes. The technology process pictures are taken on site in real-time, and the AI works in two steps. The first one is the photo verification, where we ensure that the image includes the expected elements, is clear, not blurry, not dark, and properly framed. We want to make sure that the photo is of high quality and can be used as a documentation. In the second step, we go into more detail, because we work in a funnel kind of way, because that's where we are going to check the task, the compliance check on the picture. We want to analyze specific things like cable depth, coiling, labeling, and positioning, for instance. The AI validates the photo and associates the tasks. The extracted data is then automatically fed into the operator's system of records. That's some really interesting technology. Greg, can you tell us about how some of these inspections were conducted in the past, and then how this computer vision might change things for the future? Oh yeah, definitely. That's a good question. Currently, right now, the typical way of inspecting what was constructed in the ground, like Claire had said, you could do photos. But they didn't do every installation. Like, they didn't go and take a picture of every hand hole or every conduit that was placed. And I think with the new technology, that it's really going to help the as-built process. So if you could go in a subdivision and you were taking the pictures of all the conduit that was being properly installed, the hand hose that the conduit is leading to, then you could really help with your as-built documentation when you come back through and actually pull the fiber and put in the terminals and the splice cases and stuff like that. You're going to have a much more in-depth picture of all of the areas instead of traditionally what's taking place is construction guys. As they go out and they do the construction, and then they'll take a picture here and there, and then the supervisor or the quality control person will come by and they'll do what they call spot checks, where they don't go to every location to verify what was actually done. And that really hurts the documentation of all the as-built in the future of servicing somebody where you think there might be a terminal, but there wasn't a terminal place and stuff like that. Moving forward, what Claire's talking about is being able to document that and analyze that as you're going. And then there's no need to go back and send somebody out and do a truck roll or something like that. You can do it all during the construction step, which is going to be really, really good for everybody in the industry. Yeah, and you can actually extend this post-construction as well. You can use this technology to help your teams on the field to do things like, say, for example, using optical character recognition to get labels out of assets and populate them directly under a system of record. That reduces manual intervention, mistyping information, or simply forgetting to add that information as part of the work package that you're concluding. Checking for things like inventory checks, like take a picture of a set of ports and see which ones are available. Are those the ones that should have been available? Or are there more available or less available ports? And then influence what you do and capture as a result. And then give them some real-time feedback, like guide the worker across the task to be completed. And give them some visual aids in terms of how to get this done quickly, how to get this done with quality. We definitely know field conditions can be unpredictable. Can computer vision adapt workflows in real time based on what it sees? Yeah, absolutely. And I strongly believe that it should. Field engineers often arrive that site with unknown conditions. So assigning them with very rigid workflows can be impractical. With our AI solution, we want to make sure that the workflow will depend on what the AI sees. It can, for instance, trigger new tasks, action, or adjustments to their workflow. For instance, we work with a large UK network operator, which already uses this. And based on what the AI identifies on the first picture, it automatically unlocks the appropriate workflow, including the required pictures and the AI checks. And IQgeo, our key focus at the moment is to make sure that we get every aspect of the lifecycle fed and enhanced and rich with information from the field. So, right, this have the signs, have plans, have build instructions, be enhanced by what you're seeing on the field. So we're making sure that we're closing that loop and giving construction crews and field technicians information that's actually relevant and doesn't have to go back to HQ for approval. The idea is that we want to keep the project on track. We want to make sure that things get done correctly in the span of time that they were promised to be done correctly without necessarily having so many touch points with folks in the back office having to approve or correct work that's being done on the field. So, Clara, I'm interested. Can you tell us more about how your computer vision software works? Do you feed it a bunch of photos of correct information? Tell us more about how that learns so that it can adjust in real time. Yes. So at the beginning of a project, we need to identify exactly what are the quality standards of each project. And I know that Greg is just stepping off a call where he was giving us some insight as to how poles are installed in the U.S. And with that experience and with example picture, we are able to train AI models to give some feedbacks over the quality check that were scoped. So, for instance, for a pole, we are measuring the distance between the ground and the first cable. That's an example. The AI will also need to see a not OK example. So we really need to have concrete overview of anything that can happen on the field so that we can prepare it to do some real time, real life analysis on the field. Greg, what are some of the things for poles, for example, that would be important for the computer vision to pick out? Or is there so much difference between them where finding those bad examples would be difficult? No. I mean, like Claire said, we're in the middle of developing the AI to detect by either street pictures or a personal picture on site or images. And first of all, we want to identify what's on the poles. So we're going to identify the power and where it's at on the poles and then all the communications that are underneath the power lines. We have a team working on that right now, which would be just an enormous help on getting pole applications set up and preparing for the make ready. If you want to go aerial, making a determination, whether you want to try to continue with the aerial route of getting all the poles measured or you want to just look at the photos and know just by the photos without even sending anybody in the field. Rather, you could, you know, estimate what your cost is going to be and you might make a decision whether you just want to go underground. So the next step, the next evolving issue that we're going to have is once we identify what's on the poles, we're going to actually have measurements of all the attachments that will help us when we apply for the poles to the different power companies to make a new attachment or even overlash to an existing attachment. So the technology that they're speaking about and obviously the help of AI analyzing the pictures that we can take in the field or the images that we can have AI analyze, it's going to be a real game changer once we get to that point to where we can help the provider and the construction guys determine the path before they even get in the field. And it's going to save them a ton of time doing that. And I know that we're talking here about bite-sized fiber. That is great. But we also deal with utilities at IQgeo. We also deal with MSOs who basically have their infrastructure running on coax. This technology is extremely helpful for anyone that has to deploy a service using aerial routes. It's not just for the fighter provider, but it's for everyone. If your service requires you to put something on a pole, an access point, some power amplifiers because you're running coax down the street, whatever you have, this can recognize what's in there and it can give you much more accurate records and help you with future planning and other activities as well. The rule of thumb you have to keep in mind is that as long as a human eye can see it, then the AI will be able to pick on it as well. As I said, we work in two steps. The second step being going into the detail of what's exactly on the picture. As long as this is something that you, Greg, you, Stefan, would be able to see, we can train the AI to see it on pictures. Is there a way to actually train computer vision to recognize stuff that's more than meets the eye? It would have to be visible. For instance, I remember seeing some film or some series when I was younger where they would take from a very blurry picture. They would find a car license plate from very, very far away and be able to zoom in on it until it became completely clear on the image frame. This is not something that's possible with a computer vision. We just see what's to see on the picture, just like a human would. I don't want to repeat what Greg said, but prior to using diplomatic, our customers are doing spot check, like maybe 5% to 25% of the cases of production. And we want to make sure that the back office team can still do 5% to 25% of the check, but we want to orient them to the check that requires some human analysis. So it might be the situation where the element is less visible or the element is different than what the AI is used to. And we want to make sure that the 80, 90, 95% rest of production that's completely what the installation is correct. It doesn't require some human time on it. That's interesting. Yeah, that's interesting. So what does that look like when people are going out into the field? Any disputes that there might be with where things are placed or how could using the computer vision help to make higher quality builds? Yeah, we really want the data that we extract from the picture to be used as a ground truth between diplomatic, between the end client and between the various subcontractors or any given stakeholders that might need that data. Because computer vision generates objective, verifiable records. And that's why we want to make sure that it's shared by everyone and that the issues that might occur are accessible to anyone. You basically see at this point distals being essentially oversight, right? So for the field crew, this is not a replacement. You're not going to attach a camera to a drone and call it a day. This is something that you use to enhance the work that the crew is doing on the field. How about like resolving disputes, right? So what happens when someone saw something on the other side and they went to the side and they take pictures and it shows something different, right? So what is, in your experience, the best policies to apply in terms of resolving those conflicts between human and machine? Yeah. Yeah. We have cases with subcontractors where the approval and payments are processed way faster thanks to computer vision and thanks to the AI verification of the vast majority of production. This means that the quality team can focus on the dispute situation and they will be able to access pictures there which are exploitable. Because from experience, even though we know that some clients were already taking pictures, if they aren't analyzed, then you will find out two months later, maybe two years later, depending on when the dispute happens, that the picture was actually pitch black or completely blurry or is not exploitable. We want to make sure that this situation is avoided. And if there's a dispute in design, if, for instance, a technician was expecting to work on a cabinet and when he comes to the field, he's actually working on the chamber, having visual proof of what is in front of him on the field will be a help to update the design team, the subcontractor, any given person who might need that information. Yeah. And Claire, I just wanted to segue back to what Stefan had said. Like, I understand that when you're taking these pictures and we analyze the pictures, you could actually add geographical representation to that. Right. So I've seen it in my past where some contractors, they would turn in the same picture for five or six different locations. Right. To Stefan's point. So what we need to be able to do, like if a splice case is supposed to be in this hand hole, a terminal is supposed to be in this hand hole and so on and so on. And there's going to be longitude and latitude associated with each one of those photos based on AI analyzing that to let you know that you're geographically in the world to where the contractor saying that you actually are. Correct? Yes, exactly. The geolocation can be attached. And we have also the possibility of refusing a picture that would have been already sent. I also believe that it's very important for each of my projects to make sure that if there's an asset ID to be seen, if there's some kind of identification of the asset that's visible, I want to make sure that we check it to make sure that the technician is working at the expected place. And a combination between the geolocation and the asset ID, I believe, is helpful for the client to make sure that he has the freshest data from the most accurate asset. What would it look like getting computer vision into contractors day to day? It would depend on their current day to day. We know that most of the subcontractors are already taking pictures, but most of the time they are sending them per WhatsApp or they are just putting them on SharePoint at the end of the day. So they do not really get some real-time feedback over them. If there's already a field tool or through IQgeo, we can make sure to integrate seamlessly into the subcontractors' processes and just show the real-time results over there. There's something interesting about this. Like, there's all these ecosystems that we're building, right, to help operators do better work. But ultimately, this all goes back to the technician on the field because technicians on the field have seen things that we wouldn't believe. They've seen splice closures that have been completely left on fire, right? Like, stuff is just connected will and nearly. They've watched cables, like, just beam with the wrong power levels. You're an OLT or an ONT. And if you're not capturing that with computer vision, all those moments will be lost in time. And the knowledge that comes with those moments will be lost as well. Yeah, what level of training and manpower do you think that computer vision can help? Like, when you have these situations, if you were to run into some crazy situation, would there be a lot of time spent? Normally, you have to go back to the office, figure out what to do. We like to show example pictures always directly on the app. And we have had some proof in the past that the implementation of the AI solution was actually used almost as a training tool for new technicians on the field. At some point for a client in Europe, we were surprised by a rise in error for very specific pictures, which was very complicated to take. And we actually understood that it was related to the rise of the arrival of new technicians, because it was in the month of September, I believe, or January. And we actually understood that this was a point that the people weren't enough trained on. And so they were using the AI to correct themselves directly on the field. And we also advised our clients to add some training on that specific installation, which was quite complicated. So I really believe it can be used as a training tool for the field. Greg, how have you seen that in your career running into specific issues? Do you feel like crews without using computer vision usually can tackle any kind of problems? Or is there a lot of back and forth with the back office? Yeah, well, I mean, that's the number one problem today is all the back and forth that you talk about, Ellie, is that really slows down projects. Because to my knowledge, a lot of people in the United States, they haven't used the computer aided construction tool, to my knowledge. Now, some people might be using it, but most people, like we talked about before, they'll go out and they'll do the construction and they'll document any changes in the field. So normally, you know, they'll write up a paper copy map that they have, right? So that's part of the problem. I mean, the main issue that I feel from an experience level is to get the construction guys and the technicians that Stefan were talking about used to looking at a digital copy on an iPad or a laptop in the field instead of a paper map. Because typically what happens right now is the design crew will give the construction guys a paper copy, 11 and a half by 17 copy. They will go out into the field and they'll take a couple pictures here and there that's required from them from the provider in which they're working for. And then they'll mark up the paper copy. And traditionally, that goes back to the designer. And then there's a whole process of redesigning and then going back in the field. What we found, we had a client in North Carolina and Pennsylvania that did away with the paper copies. Okay. We were able to use a network manager in the field. They actually constructed the fiber plant from network manager in the field. So we were able to use that. We didn't have the pictures then. I mean, we could take a picture and we could upload it to the structures periodically. But with the introduction of having diplomatic integrated into network manager or workflow manager for the different customers and getting the construction guys used to using a digital copy in the field instead of a paper copy, that's going to eliminate, Ellie, the back and forth that you're talking about. And boy, is it really going to speed up the time. Claire talked about, you know, the contractors getting paid quicker for their work. Instead of all that back and forth, they have to wait because the provider that's hiring them, they're not going to pay them until they get the work complete. And they're satisfied with the work that's complete, either a spot check or whatever. But if you could do that and get the historical facts that, okay, you're using this AI component to do that for you in the field as you're going and the provider gets comfortable with the contractor, all of that's going to eliminate the back and forth that we talked about. And the design changes could be made on the fly and any discrepancies could be handled in a matter of hours at the most days instead of weeks or months going in the future. That's what I see happening. Yeah. I deal with a lot of customers that are unwilling to part with their paper maps and their old artifacts because that's the way that they've been working for a long time. And even those customers, right, they're very reticent to adopt new technologies unless it saves them time or money. They see the value of using computer vision and reality capture as kind of like the interface between them and the work that's been done because it saves them some significant amount of time in the field. And from even an operational perspective, if you go into the maintenance or break fix and you have to update your system of record with the new state of fast build, because as builds, not forever. Having computer vision and just capturing the details on the changes that you did on the fly can save you a lot of time, money, and grief. So it's only a matter of time before people start seeing how much of this actually augment the workflows that they currently do with a paper map or with a PDF printout or something else and go, oh, well, I don't really need to like struggle with my life as much as I've done before. I can get this done in less time with less friction. We have a situation that I like to talk about in Europe where at the implementation of a project, we knew that the field workers were doing two to three installations per day. And after the AI came in, there was some guy who did 17 in the day. He actually had to stop because there were not more than 17 planned for him because the design team had never planned for more than one week in advance of jobs. And it was a bit easier for him because it was in a multi-dwelling unit. So all the elements were closed to each other. But I have to say that this guy was paid by the job. So I always like to think about how it must have felt for him to be able to do 17 jobs instead of two, maximum three per day. Yeah, it sounds like computer vision is going to make a huge impact and bridge the gap from everyone using paper, too. What are some of the other high value use cases for computer vision outside of construction? We do a lot of home connection and maintenance, for instance. Anything that might occur on the customer's premises. This is a situation where the first time right is very important. We know that from talking to subcontractors, for instance. You have to block a customer for half a day or an entire day at home to do an end-of-line installation. And if you have to come back, if something went wrong, usually this is something that they would hate. So that's why giving some real-time feedback really has a lot of value there. I also see a lot of value in situations where the accesses are kind of complicated. For instance, if you need to go in someone's basement or if you want to analyze what's within a trench. Usually the trench is closed on the same day or maybe the next day, but it's not going to stay open very long. If you do not get the results in real-time, then you've lost the benefit of the AI. Yeah, I'd like to add something to that. Claire brings up a good point. I mean, there's a segue, Ellie, between construction and operations, right? So you do the construction and you go out and do an installation. But having the pictures at the time of the construction will help the person that's in the field. You know, they can go into the system of record. They can look at that picture. They can bring that up and say, okay, there's a terminal that belongs to 123 Main Street, right? So that he or she are going to know, okay, all of that stuff's there. By doing the pictures of all the assets on the construction phase, it just carries over and makes the operations so much smoother. Claire talked about the contractor that was doing the MDUs. They sped up because everything was done during the construction phase. And I'm sure Stephan could speak to this more. But if you can document that all in the construction and your as-builds are up to date, like Stephan said, the as-builds are only as good as the next time you go there. But then Claire's right. You know, you take a picture of putting in the conduit for the drop, the little flower pot that you might put on for the drop going to the home. You document that just like you did with the construction. And then it's a seamless procedure all the way to hooking up the final customer, which that is the end goal for everybody. It's going to develop a lot more trust with the customers, too, because they're going to see more instances of doing it right the first time. I'm down with OSS and I came from that kind of background. So anything that has to do with operations, it's near and dear to my heart and my career. You can only have good operations in the sense of having the data to operate on. Being able to have that handoff point between taking a picture and saying, yes, work's completed, and then automate the process of, say, for example, doing tests and turnout for a new service to create a birth certificate, or doing a retest for an existing service to get proof of work for customer SLAs or internal SLAs. That is massive. For a long, long, long time, people used to do like physical work and go like, I am done and go away. And then someone, sometimes days later, has to go in and actually do proof of work on the work being done. If you can actually start creating these things into a much more seamless workflows where the work being done is not contingent on you manually handing off or someone reviewing work somewhere else, but automated via computer vision and process automation, you start getting into real serious, not just cost savings, but efforts to improve customer service perception from a service provider point of view. It can also be very useful in situations where multiple people have access to the same assets because it's going to give some kind of a sense of responsibility over the state of the asset. If you look at the fiber network in France, you can see some terrible pictures with a cabinet with way too many fibers in them. It's very complicated now that the network is in such a bad state. Whereas when you start from the beginning to document it and to feed the digital twin, then it's easier, I'd say, to make sure that the asset is going to stay in a good, healthy condition for a longer, longer time. What do you think the biggest impact is going to be over the next few years with computer vision and AI? A lot of people started their journey into AIOps with the, we're going to use this basically as a frontline interface for people to not hit customer service reps as hard, right? I'm going to go in and chat with someone that's going to solve my problems. And that has had an effect. It has helped kind of like create the first line of defense for service providers to get information up to customers with a much more natural interface that they can react to. But I think that's selling the benefits of AI machine learning kind of short, right? Things that we have dreamt of and we envision happening in the not so distant future. It's where we can use this information to actually detect change of our time. Do your yearly inspections for polls, for example, and start noticing, hey, this ball's now starting to tilt a degree per year and be predictive about it, right? This tilt, if you do not correct it in five years, it's going to basically rip out all the connections that's on that ball because it's going to be past the slack. Using that same information to create large groups of data sets where you can go in and say, well, is this particular asset performing the way that it is? Because is it the asset the way that it was installed? Time and weather affected it. Events that transpired across the network that may have impacted these things like storms or people driving into a pole repeatedly or people just attaching stuff to a pole forever. It's moving the needle from just simple recognition to recognition with recognizing the pattern. Yeah, I would just like to segue right into what Stefan's talking about. I think, Ellie, in the next two to three years and beyond that, even starting the whole life cycle of planning and getting the customer actually hooked up. You're talking about using computer vision for doing the survey, for analyzing the make ready and the poles that we were talking about. Rather, a provider would want to, you know, analyze whether they want to go aerial, whether they want to go underground. And then all of the documentation that we talked about during the construction of taking a picture of all the assets and the conduits and the hand hose and document all that. All the way up to the actual installation on the premise. I think that everybody is just going to, they're ultimately going to buy off on this, the construction guys, and they're going to see the value of it. Because of the quality of work, because of getting paid quicker, and because of getting more jobs done and a more suitable time and having better documentation. I just think as a whole industry that we had this to look forward to. And doing this over 30 years and dealing with the as-built that we've had to do with the paper copies, I'm going to welcome this more than anybody, I think. I agree with the both of you and technology is actually moving forward in the directions you're depicting. For instance, new multimodal AI models are arriving. We are also more and more talking about Gen AI. What you need to remember here is that we will soon be able to mix different sources of data. So we'll be able to use the field imagery that we are already collecting, using the street and satellite imagery as well, using maybe some data coming directly from the end customer, using the data coming from IQgeo, using all different sources of data to support the design, the deployment, and the maintenance of the fiber. Yeah, I am so excited to see where we will go. So, well, this conversation has been so interesting. So thank you so much for joining. This has been great. Thank you so much. Thank you.