Latest / AI in Manufacturing / Can AI Agents Make Continuous Improvement Truly Continuous?

Transcript

Can AI Agents Make Continuous Improvement Truly Continuous? · AI in Manufacturing

• speaker-0: Welcome back to the AI in manufacturing podcast here on industry forward TV. I'm your host, Good Zai, Mandy Teresa. Our topic of discussion today is going to be centered around the question, can AI agents make continuous improvement truly continuous? So this is a question basically that I was posed by Devin, our guest today in one of his writings, and he's going to help us unpack that for us here today. Welcome to the show and thank you so much for joining us today. speaker-1: Yeah, thank you. Thank you for having me. I'm really excited. This is a it's a great topic and a very unique opportunity, I think, for manufacturing. right now with AI and continuous improvement, I feel like it's it's a very ripe for efficiency. So I'm excited. speaker-0: Yeah, absolutely. So before we jump into the topic of discussion here, could you perhaps start by introducing yourself to our audience? Give us a bit of your background and tell us a bit more about Scrin, what you're all about. speaker-1: Yeah, so my name is Devin Bouchon. I am a software engineer by trade. grew up in the San Francisco Bay Area and started Squint the after basically I had a career in engineering, primarily at the intersection of AI and computer vision, and worked on a variety of things, but in my role before starting Squint, I ended up spending a lot of time with industrial companies. And as I worked with them, the product I was building before Squint was actually just something that brought them data from OT hardware. And that got me closer to understanding the problems they were trying to solve with data, right? And it was things like when something breaks, what do we do? and data helped them identify where the problem was, but didn't help them figure out what to do next, right? And so the feedback I kept getting was. This data stuff is good, but I really need to know how to respond based on what the data says. And so I kept asking people, where is that coming from? And where like what would what do you need to see? And people would say, I need the knowledge in that person's head, you know, the tribal knowledge that has existed for decades, and I'm just not familiar with how to fix this machine when it when it does this behavior. And so I realized that. Someone had to go actually capture that tribal knowledge because the people with most of that industry experience and the the thinking around how to improve operations, what to do when something goes wrong, those folks were like essentially they were going to age out and retire. and especially with manufacturing growing now so much, that was a major issue. And so I started Squint five years ago. and we are an industrial AI company. And what we do is we help you essentially start by capturing tribal knowledge, understanding how process works. We do that by leveraging video. So you can actually record a video of a process, Squint will watch it, and the agent will it's embedded with a lot of understanding of lean methodologies and Six Sigma. And generally we we've built an industrial ontology for it to leverage. So it understands when it sees things, what's happening and what those things are and what they're related to. And it will actually then be able to map the process that it sees. And you can use that to either create standard work or do process improvement. or like, you know, yesterday I was talking to someone and they used that as a starting point for DOE. So they ran like a defect, you know, root cause analysis and figured out what was wrong. and then ended up solving the issue much, much faster than a human that would have designed an experiment and watched, sixteen hours of video in slow motion to figure it out. So just it does so much. and it's very capable. And I think the the progress we've made in the last like six months has just been exponential. So it's a really exciting time in manufacturing. So I hope that's helpful. Quick intro. yeah. speaker-0: Absolutely. mean, you know, tribal knowledge is a big issue. It's a big problem really that I see a lot of companies struggling with. And I think what's particularly interesting about your approach is how you choose to solve it with video agents, I would say. So you've seen a lot of companies solving that, using conflict capturing the workflows through an input. know, dashboard or form or something like that. But it's very interesting the approach that you chose to use video as your primary method of capturing that would you say is what information would you say is not available with, traditional ways of trying to capture that, you know, travel knowledge that you thought video would be more suitable for? is there a specific reason why you decided to go to speaker-1: Yeah. I think manufacturing is building things is artisanal, right? you can have standard work and a recipe, but at the end of the day, you're like you're cooking, you know, you're it's art in some ways. Like the way you actually do the procedure, it may follow the same steps, but there are is nuance. And so If you're actually trying to watch someone and create standard work, or you're trying to watch someone and understand why does this shift run faster? How do we actually take the process improvements that this person has made? it's easier if you can actually visually see and hear and understand what's happening, right? If you just try to look at the timings from the NES, it's very hard to infer what process they followed that led to the improvement, right? You're blind. and so when agents are blind, they can't actually observe the key changes that someone has made to a process. So I think for that reason, we decided video was a very critical part of our platform. So we invested really, really early. It's one of the first things we did was video intelligence. And we've been upgrading this over the last five years. and now we're at a point where we can take, you know, security camera footage, CCTV footage, and we can understand. a process across, maybe a a U-shaped line with 18 operators all moving around doing their own things, and we'll blur all the faces out, we'll understand what happened, we'll create the process, and it will essentially then create a simulation or model of the line, but in in in like a value stream map. And now that you have a value stream map that actually was based on reality, you can start operating on it and you can work with the AI and it actually gives you recommendations on, here is here's the waste that it saw, ranked by how much it's gonna cost or what the effort will take, and what the payoff is, right? And you can say, This is my goal, this is my constraint, these are my affordances, and it will recompute all of that because it sees the whole process. So that's that's kind of what you win by by getting video. It's a very rich format. You can see a lot. speaker-0: Absolutely. I mean, you recently shared about a manufacturing project in Singapore. Would I think it will help sort of you know, explain this more concretely? I thought that was a really fascinating example of what, you AI agents are starting to make possible in manufacturing. So maybe can you unpack that real world implementation for us? Like, what were the challenges? How did you go about implementing the solution and what were the findings or results of that? speaker-1: Yeah, totally. So it was really interesting. This was one of the first times that we were able to test our system. the you know, the the lean agent that we've built is pretty new. It's something we released in the last couple of months. and so this was the first time we were testing it at such a large real world scale. So we went to a facility in Singapore. they do it's like blade repair for turbines. So large things like you know building sized turbines come in and they take them apart and they analyze for you know where repair is needed and there's a lot of different process involved in this. So one of their lines, they actually this customer invited us in and they said, hey, we actually have video. We want to test this agent. So we have a we've created a test. And it was fun. I like I like competition. So it was it was a fun time. But they said we have a test. We've recorded video from two years ago on this line. and it's 144 hours of video. So we're just gonna give you the files and you have this conference room and we'll talk to you tomorrow. Let us know. Let us know what you find, you know? Can you make the process better? Can you represent the line? Can you balance it? Our goal is to try to free up operators because we don't think it needs 18 operators. We think it could do with fewer. And then we would like to move the operators off to another line. And so they said, basically we've we've already done this exercise with humans. so we want to compare the results of the AI to how our continuous improvement team did. so we went and we fed the video into Squint. and we basically let Squint use the lean agent and it applied, you know, all of the lean and Six Sigma methodologies over the last, you know, hundred years as they've gotten better. The AI is modeled after it. And so it actually looked at the process. It understood the, you know, it built out the U-shaped line that they had. it understood each station's responsibility, mapped out the cycle times, created charts of value add, non-value add. time spent, what the activities were, and created an analysis basically at the end. for an analysis for each station and then also an analysis for the line that takes into account things happening across. and so we took that and we said, okay, well the goal this is you can talk to the agent, right? So you can say, look, I want to balance this line. my goal is to keep the correct cycle time. But reduce the operator count required. Can I is that possible? Right. And so it actually runs the computations and simulations and it shows you its work. And it came back and it said, actually, I can free up six of the engineers, six of the operators. So you only need 12 and you can still hit your cycle time. And it rebuilds the Amazumi tarts and a flow simulation to show you how to basically set up the stations. which stations need to be together, how you can batch the work. and it also basically it found some of the waste in the process and it took it out. the bottleneck station it identified and it said, hey, there's some stuff that is happening at the bottleneck that you can prefabricate. So actually you can even lower your cycle time a little bit if you want to. and so it did all of that, right? And we came out and the next day we presented. So we went to the team and I was very nervous because I was like, yeah. I don't know anything about gas turbines. You know, like I I couldn't watch all the hundred and forty-four hours of video. It's impossible to watch all of it. So I watched some of it and our team was like we looked at it and we're like, okay, you know, looks right. I don't know. let's let's talk to the the lean team. So we presented it and there were some really interesting findings. There were things that they didn't see when they did it two years ago, because humans cannot actually Hold in their brain all 144 hours at once, right? So there were associations that the AI could make that they had not realized. So I'll give you an example. it saw there was a process where someone is basically they're sanding, they're polishing a part and it's creating a lot of dust. Now that dust is settling on the part itself, and so they have to use a blower. And they have to blow the dust off. Right. And what it found was it said, hey, based on what I'm seeing, there's a downdraft table already. They're working on a downdraft table, but it's not evacuating the dust at the right rate. So there's like X percent of the time is spent manually blowing dust off. So actually you should go tune your drown down draft table or look into if it's effective. Because there's an opportunity here, because you're you're doing a lot of unnecessary work potentially. This is waste. Right. So there were things like that that the AI can find that it's hard for humans to recognize. And then they were generally they looked at it and they were, I think, pretty blown away. And it's just been it's been an incredible, you know, launching point for this agent since. So yeah, it's it was it was fun. Hope that answers the the question. Yeah. speaker-0: Yeah. you mentioned that this is something that they had, benchmarked using like manual human, you know, you know, analysis. it, would you give us a sense of how much faster was the agent analysis, you know, compared to, the manual? speaker-1: I left this out because it sounds so crazy. But basically we came back and we were like, man, we d we did all this work, like here's our results, what do you think? And they looked at it and they were like, Yeah, this is really good, this is good, this is good. and at the end they told us that it took a team of eight people seventy-eight days to do the same thing two years ago. and the agent did that independently in a single day. Right. That is just like it's I've been ever since then, I've been basically, as I share with customers and people that we get introduced to, I think it's just a it's a crazy moment in manufacturing. This is what when when we started talking about doing this podcast, the thinking was continuous improvement can really happen continuously now. Right? Mm-hmm. This is like it's much faster. You don't have to do a kaizen every six months. You can do a kaizen every day. The agent can go do all of the free work, right? It creates spaghetti diagrams. It looks at the waste, it analyzes the process, it builds you all the charts, it does the interpretation, and it tells you where to go spend your time. So you can take the lean experts and and the process experts and the engineers and you can say, Spend one afternoon. And we can complete the whole Kaizen event, right? That's that's the the impact that this is having right now. speaker-0: Yeah, absolutely. that's a remarkable difference in engineering effort. But I guess it's also sort of like raises an interesting question on why is this work so time consuming in the first place? I think you mentioned about eight people working on a project on multiple number of days, right? mean, manufacturers have been practicing lean and continuous improvement for decades, right? So maybe from your experience. working in this space, like mentioned for the past five years. Why do you this activity still requires so much manual engineering effort and what's of like. What's holding companies from really scaling that out? speaker-1: Well, I think the it's so interesting. I've talked to a lot of different industries, customers in different parts of the world, different types of businesses. one observation I have about the manufacturing process improvement in general is that, you know, a time in motion study was invented over a hundred years ago, like I think a hundred and fifteen years ago. that's how long the time in motion study has been around and it's been running on a stopwatch and a clipboard. Right. and you fast forward since then. After that, we had Toyota production system got introduced. We had, you know, Six Sigma, Lean, all of these upgrades in methodology happened, but they are all based off of clipboards and stopwatches and paper and humans watching and doing manual work. And I think it's so funny because continuous improvement is meant to find the bottleneck and remove it. And in the process of compu continuous improvement, the bottleneck is actually the stopwatch and the clipboard and the human, but no one has removed it in 115 years. Right. And so when I when I started learning more about continuous improvement process and how manufacturing has has done this, it was just ironic to me. I think it's it's one of those things that when you're an outsider At first and you start learning about something, you look at it differently. And the first thing that came to mind for me was we have to automate the stopwatch and the clipboard process. We we must fix this. So that's that's why we started because we had the video understanding and the AI. And now you can take hundreds of years of experience and you can package it into the AI's intelligence and it can go watch the video. And you don't need a stopwatch and clipboard. And that means you can do it nonstop. Right. And so it's just been a massive unlock. I think one of the things that I hear as well, in addition to just like it's hard to do, I think the other thing I hear is things are moving at the speed of light. People don't have time to spend on their continuous improvement practices. Right. I hear there's businesses that are growing because of AI. So energy data centers. all of the suppliers for that. So cooling solutions, rail solutions, you know, transformers, valves, all all of this stuff. they're growing very quickly. And they're growing so quickly, they actually need continuous improvement the most because if they could make more throughput, they could sell more and and raise the the ceiling of the company. But they're actually going so fast they don't have time. And so it's a funny thing, you know, like Continuous improvement takes so long, we'll do it later. and then there's businesses that, are right now impacted by tariffs and volatility in the market, aluminum costs, petroleum costs, right? And they're struggling because they need to become more efficient. But as they become more efficient, one of the answers is actually headcount reduction. And I I see like continuous improvement teams actually get Are the ones to suffer from that because their resources get reduced. And I think this is the answer for both of those types of businesses. We can say if you want to go fast and you need to you don't have time, Squint is one hundred times faster. Right. And if you have cost pressure and you need to become more efficient with less resourcing, Squid can give your team superpowers and can make them one hundred times more productive. So that's the that's the theme I'm seeing. speaker-0: Yeah, absolutely. I mean, it's really, it's really fascinating. of like what you mentioned, this idea of using, you know, store pushes and things like that. was remember I was reading the the book Factory Physics, was it about a few years back? It was mind blowing to me how. What was being done in the 1910s, 1920s is exactly what a whole lot of companies are still approaching how to actually, you know, do this continuous improvement. So it's really. It's really mind blowing. again, for me, how I imagine this needs to change, because we have got this idea of a lean manufacturing, where you are continuously improving a process. But do you think that with AI agents or video AI agents, it also opens an opportunity to move to a more adaptive form of manufacturing? What are your thoughts on that? speaker-1: Yeah, yeah. I think it's a great question. I get really excited thinking about this. So, we're in a state right now where the world needs this technology just to catch up or keep up with the change. But I actually think you know, studying manufacturing and understanding how things work, there's a you know, there's so many concepts like batch sizes and calculations that go into the economics, right? How many What what batch size can we run at? And if we reduce the batch size, it gets more expensive. Well, why does it get more expensive? It gets more expensive because it's hard to change, right? It's hard to actually change from one product to another product. And if you're producing another product, it's not just materials. Even if the materials are the same, you have to change a lot of process that's going into it. And the people who are running the process need to understand the change that's happening. And There's a a bunch of steps, right, in in the process of changeover. And I think Squint has a very interesting role to play here. I think we take it as our responsibility to help the world get to an instantaneous changeover. Right. What can we do to make it so that no matter who you are, what you're doing, Squint can instantly help you understand what the setting should be. how the changeover should run. How can we perform SMED with AI at scale and basically get to the minimal changeover that allows for like, you know, almost every type of good can be custom ordered and be f economically feasible. You know? AI can actually help that. There's a lot of manufacturing basically comes down to a lot of computation and logistics. And AI is very good at computing. But it has to understand what it's looking at. And so that's the gap we're filling. So I I I think there's a massive opportunity there. And if if you actually if you can reduce the change over time, then I think more factories can create a a larger catalog of SKUs and can be more flexible to the demand. And so if supply could match demand as it goes up and down, then I think we would have a much more streamlined, physical world and and the supply chain would be much more flexible. It would just it would just be good for the world and logistics and it would reduce a lot of the risks we have today with, the classic example is when COVID happened and we couldn't get toilet paper. Right. So that kind of thing. Hopefully that never happens again now that we're doing this. speaker-0: yeah, absolutely. Okay, now maybe let's kind of like look at how the actual AI agentic or agentic AI workflow helps sort of like, you know, change or rather help with the, you know, continuous improvement process. Specifically, I would like you to give us a practical walkthrough of your Queensland agent and perhaps the technology behind it that enables your AI to do the interpretation of activities. Identifying West and also know developing some recommendations and yeah more importantly what I also want to understand I mean you mentioned you've got a background in sort of like computer vision If you could paint a picture for us on how that is different from that traditional computer vision because some would argue that we already Have some video analysis on the shop floor, right? This is nothing new. Maybe you can also make that distinction for us speaker-1: Yeah, I think yeah, maybe I'll start with that and then I'll explain how it works. So I I think a lot of the video on the shop floor, I think the shop floor was actually quick to adopt video, right? quality use use cases typically. So it's video that can look at things coming off the line and perform a quick quick like QA, basically. Can we can we look at a potato chip? And see if it's the right color, the right shape, does it pass our past our quality bar so that a human doesn't have to look at everyone? Or so we can go from sampling to actually just looking at everything. Right. I think that has been adopted. There's ROI, but that video analysis is very discrete. It's very much you train a model and it looks for good potato chips. That's great. I think what we're talking about though, is a level of complexity that is not so discrete. You have to be able to watch a a set of humans moving around a space. you have to understand what they're doing with their hands, what they're doing with their feet, what is happening to the the actual like good being produced. if you're building tires, you have to trace the tire through the facility. And you have to trace the people through the facility. And you have to understand the tire and the people interaction. and if there's machines involved and robots involved, now you have to understand the people, the machine, and the the skew, the product. Right? And so that sort of video analysis is actually quite complex. It's even hard for us as humans. It's easy for us as humans to look at a potato chip and say good or bad. so AI had gotten there, and I think now we're at the point where. We're hitting this next level where it can watch all of this. And this is how Squint works. Basically, we've built into the agent. we've built in a hundred plus years of this methodology, right? All of the TPS methods, lean methods, six sigma methods, all all of the different things and standards that have been created are encapsulated and it knows them. And it knows how to apply them. So when it watches the video, it can instantly create an understanding of the process, the value stream, what the people are doing, what the machines are doing, what the interactions are, what the GAMP chart would look like if you were to map out the standard work, you know, combination, what that would be. and then it allows you to give it a goal. So you can say, I would like to, I don't know, I would like to increase my throughput. By 15%. And I have $100,000 of CapEx that I'm able to spend. but I can't hire any more heads. So no more headcount, but I have this much money and this is my goal. Right. And so I give it my constraints, I give it my affordances, and I say go. and then Squint watches the process, it builds out all the time studies behind the scenes, it understands if you give it a procedure. if you give it work instruction or standard work document that you have, it'll use that to understand what it's watching. if you give it schematics or a CAD file or something that's like that helps it, it'll it will study that and understand what you're looking at. And then it'll go watch, right? And so it comes back to you and it says, okay, I can't get to 15% throughput. I can get to 12 or whatever. Here's my here's my analysis. And now here's my work. Let's look at it together. And it'll ask you questions of things it couldn't understand and you can answer them. You can give it feedback. You can say, This doesn't look right. I actually can't, you're asking me to build another welding station. I actually don't have another welder. Here's my LMS data, right? And you can export your learning management system just so it knows the skills that people have. And it can look at that and it will say, okay, you only have one welder, but you have six of this other station, I understand now that we can get this throughput achievement through a different configuration. Right. So it's it's it's very flexible and it can understand all of this and it applies industrial knowledge so that it feels like an AI like an AI engineer, basically. And you're giving it work and it does the work and it comes back with results. You review it, and it hopefully gets you to a result you wanted while you go do something else. speaker-0: Yeah, I mean, does it. Does it work specifically with a goal in mind? I'm curious to find out if could also approach, you know, you know, an open ended investigation. So, for example, you mentioned that you could say, OK, I want to do so much percent throughput. This is what how many resources I've got it. But is there also a possibility that you could give it an open ended investigation? speaker-1: Yeah. And the advice I give our customers is give it very ambitious, open ended goals. It does better with that. Right. So you can actually say, I'll give you an example. We have a customer and they make fiber optic cable. And what they do is this fiber optic cable is made on a line that's like 200 yards long, and at the end they have a a spool and it winds into the spool. They have two of these lines, and one has a defect rate that's higher. So it's the same people that work on both lines. They wanted to run an investigation. So they told Squint, I would like to find out why the defect rate is higher on this line. and I I know it's not my operators, but I don't know what's wrong. And Squint said, okay, is there a way that you can place a video camera inside of the the line so I can watch the cable. And they looked and they were like, yeah, okay, I can put a GoPro inside the strander. And so they put a GoPro there and they recorded 16 hours of this cable going and they gave it back to the agent. And the agent said, Okay, I understand. we should do a design of experiments. Let's do this. Here are the settings that there's three types of settings that can change. low or high. and depending on what they're running, it's low or high. And sometimes there's a defect. And so it said, okay, here's your run sheet. Here's all the settings to try. Go do them all for 30 minutes each and mark the timestamp when you do it and upload the video. So they uploaded the video, marked the timestamps, and it looked and it found micro vibrations in the cable that mapped to the defect In the MES. And it was like, okay, so it's happening at these this two second period where the cable moved a little bit. So what else is happening at that time? It's on this setting. And over here, this operator is following this procedure. So something's wrong, right? And it it made that association. And in this case, it's not removing, it's not like doing waste analysis or lean or it's not exactly that, right? But it is running a DOE. And it came back and it was like, here's your result. You should probably go look at what's going on at that station at this timestamp on this procedure. And then they went and they studied that and they figured out the root cause. Right. So this is just an example, but it can it's highly capable. Every day people use Squint in ways that I didn't know was possible. It is doing things that I we have not even built. It it builds stuff, you know? speaker-0: Absolutely. that's super interesting. Super interesting. now, as mean, manufacturing is like processes, manufacturing processes are really isolated, right? Manufacturing is really interconnected network of processes and things like that. So optimizing one line, you know, may not necessarily mean that you have to like, you know, optimize the, the entire. So whether it's within a site or the entire enterprise in a situation where what happens in a certain line might be might affect the next line or might be because of the previous line. So I want to get a sense of once your agent understands what's happening in an individual workstation or individual line, how does it move towards optimizing on a much bigger scale or from a much bigger across the enterprise or across the site? speaker-1: Yeah, it's a fantastic question. This is exactly I think industrial AI like automation is gonna come in a few stages, right? So the first stage was automating tasks. This we've been doing for a long time, which is you can give it a video, it can make a procedure, right? Basic tasks. Second stage is giving it that goal and letting it run that investigation, right? Which is kind of a mission. If you think about it that way, right? It's like improve my throughput on this line on this bottleneck, go. Right. And we'll go do that. the next stage is actually here's my plant view. Here is all of my system connections. So I have an MES, I have an ERP, and maybe I don't know, I have a CMMS or something, right? And also here's the video footage of my line. security cameras, something like that. can Squint actually look at all of that and help with higher order problems like sequencing? Right? If you have sales data and demand is changing on the weekly basis, how can we use that week's demand and that week's material flow to determine what gets built to reduce WIP and actually optimize for like the the end product that you're producing and selling, right? This is a different type of mission. It's a little bit more holistic, but the AI can still do this if it has access to these things, right? And so that's we're moving into that territory now, that's stage three or level three, as I see it. And then level four is really once you've done that to sites around the world, across an enterprise, can you actually connect them? Can you actually start detecting, you know, There's two changeovers, one in Germany, one in the US. they're on slightly different types of equipment, slightly different product, but in the same company, and they have different practices and identify the best practice and build a new best of class procedure, right? And then send that out to all the sites. So if I go to sleep in LA and I wake up the next morning, I wake up with a two minute process improvement in my inbox. Right. And Squint has already customized it for my process, my operators, and it's built custom training. Right. this is the direction we're going. A lot of this is possible today. the next, the last step is really this is where we're focused over the next couple of months is connecting the gaps between different plants. And so how do we take this to the level that it's always on? It's always acting, and it's always improving and helping you operate more safely. At higher quality, hit your delivery targets at a low cost, right? speaker-0: awesome stuff. Right. think the other thing that I would like to find out from you is, I mean, I think you've already alluded to this idea that to be able to make recommendations, AI needs to understand a lot more than what's actually on video. Right. So this stuff that's not on video on AMS systems and things like that, because the video can show an AI system what is happening. But in some cases, it doesn't explain why processes work. in a certain way, right? So I would like to get a sense on what additional operational context does your AI agent need to make or typically need to make those reliable recommendations. speaker-1: Yeah, I think you know, the way I like to think about it is just ask it. Like this is my recommendation to customers is go ask it what it can do. So say that this is what you're trying to do and it will tell you what data it has, what data it needs, if any, and it will give you options and to the point where you know, it can do things that we've not thought of. And our job as a software engineering company is to build tools for it so it its capabilities grow. And then its job, the agent's job, is to work with you, the manufacturer, to understand what you're trying to do and help you figure out how to do it. And and it's very good at that, right? So it's gonna, it's gonna tell you what systems it needs access to. Typically, I think it's, you know, ERP data is very rich. Most people have it. And so that's a great export or great integration. And then from there, if you have video, it's very rich in terms of what's happening on the floor. So then you see the company performance and you see the floor performance. That's like usually very good. If you have an MES, that becomes really helpful. because you get good time series. You know, if if you have historian data, we can take that. LMS. So if you have training, we can then compare performance against how people are trained. And we can help you identify opportunities for improving training. Right. So there's there's a lot of different combinations, but Squint will actually guide you. speaker-0: Yeah. And maybe at the risk of getting too technical about the integration aspect, if you want to get a sense of, you mentioned that you, you started really collecting data and looking data from the shop floor. sort of like integration mechanisms do you have for, for, bringing data onto, onto your platform? Like do you have like standard OPC UA MQTT maybe just kind of give, give us a sense of what sort of integration mechanism we have available. Yeah. speaker-1: We so we have a number of standard integrations that work with Squint, that you can go and you can click the button and turn it on. but actually the direction I'm going here, that Squint has been building towards, this is what I was trying to say earlier, is kind of like anything is really possible, right? We're getting to the point where AI can build the last twenty percent gap. So we build the eighty percent of the base of what is required and the AI figures out how to close the gap. For your custom need. Right. So actually a big part of that is integrations. So what what we we are developing is a way for you to say, this is my homegrown, you know, blah, blah, blah. This is my MES that I we've created over the last 15 years. and it will actually guide you on how to integrate it. So it will actually tell you, here's what I need to know. Tell me this, tell me this, tell me this, maybe take a screenshot or do something else. And It will try to write the code to integrate on its own. That way we can support a lot more. Because otherwise, what happens is there's such a spread of systems on manufacturing floors, even if, we have an SAP integration, but there's a million versions of that, like, you know, custom deployments that we have to integrate with. And so if you came to me and you said, I need XYZ integration, I would have to tell you in the past, I would have to tell you, great. Let me come back to you later this year and then we can start. Now you can talk to the agent and it will build most of them on its own. And our team is there to help increase that number over time. speaker-0: Yeah, a hundred percent. think also a previous guest on the podcast also expressed this one very strongly, the idea that integration is gone. Right. It's it's it's it's now the agents could take care of any kind of integration. Right. So that's what's more important now is the context that you really build and the intent around the data. But now speaking of, you know, that idea of really, you know, changing how you operate, there is a lot of friction. mean, identifying an opportunity and recommending, you know, is one thing, but getting people on the shop floor and company to sort of sustainably sort of like implement or use a tool is where I find a big challenge there because you could argue that manufacturers often have no... shortage of even your traditional dashboard. But the challenge is translating those findings into something that you can use to make those improvements sustainable. So how, in your view, can your AI agents help manufacturers move from recommendations to actually changing operations, to actually getting that to stick? change management question then. speaker-1: technical. I think I think it's a great question. Change management is like 90% of the battle, really. I think you could you could bring experts in, look at a line and give you so many ways to improve the process. And then how do you actually make sure it gets adopted? That is a big gap today. So that's what we've built a lot of our, product around. So a lot of the product that Squint has built is actually taking the new process And so say say you do the, you do a time study, you find some improvement, you have new standard work. Squint will take that, it will understand, what was this operator trained on prior? What's the gap, the learning gap? and it will the AI will actually generate training material for them that runs on a tablet and they can use the tablet, it'll walk them through a first time experience. we have augmented reality on the tablet. So you can actually hold it up and the camera will guide you, tell you where to go, how to change your, the operator's motion is very important, right? You if you follow the same path, that's typically best practice. Squint will guide you through the path. So you start to understand the new way of working. and it's better than a PowerPoint. You know, the PowerPoint is very hard to internalize to remember, but learning by doing is much more effective. So Squint does that. And then we create, proficiency exams and all sorts of stuff as a next step. but you have this tablet available that understands the process and you can talk to it. So in the first week you're running, you can actually ask it questions and it will help you because it knows that your process just changed changed. So that's our approach. speaker-0: Awesome stuff. Yeah. So I think before we finish, I want to get into something a bit practical advice for our manufacturing leaders, a chunk of our audience here who may be looking at this opportunity today. I mean, we've covered the technology and you've provided some very impressive real world examples and also kind of like the bigger implications for industrial engineering. for manufacturing leaders who are interested in of like agentic AI or your So solution for continuous improvement, where would you sort of like recommend starting and how should they look at, know, what should they look for to determine whether an initial deployment is, or how should they really go about, you know, thinking about, you know, deploying this solution. speaker-1: Yeah, I think the biggest thing is don't boil the ocean, right? Don't try to say I'm going to go document every process in my company, improve all of them, and then I'm just gonna push it all to the people and have a new company, right? That's not gonna work. So you you have to pick start with a real problem, start with something that is an active initiative that your team is worried about that you're spending time on and use AI in that scenario to help you find, if it's a time study or some sort of process improvement, help you find the bottleneck, help you speed up the process to reducing scrap or, you know, becoming more efficient in process on this line. Whatever it is, bring squint those problems, the ones that your people are spending time on. And let it speed up the process to solving the problem. Because once you do that, people don't want to go back. Right. the customer example I told you about with the fiber optic cable. The alternative to using it was to go watch 16 hours of video in slow motion to try to see frame by frame movement of a cable. That is like, it sounds like a punishment, once they've used this. That team is never going back. There's no way. why would they ever want to do that again? So I think it's all about finding a problem that matters and developing quick wins and then starting to grow from there. once you provide that value, you build trust and people want to use the AI because it's actually the fastest, most efficient way to solve their problem. And then from there you can kind of grow from a team to a facility. multiple facilities and that that's the right way. And I think over the next three years we will see a a massive rise in adoption for AI and factories as a re as a result. speaker-0: Absolutely. And how do you typically engage your customers? you start with the pilot? How do you approach proving value to customers? speaker-1: Yeah. I mean, I really believe that the product just works. So you should come to us. we'll let you use it. You go talk to the AI, tell it your problem. If it solves your problem, we're happy to work with you more closely and and you know, we can then enter a commercial agreement. But if it doesn't solve your problem, I would love your feedback and we're gonna make it we're gonna make it better. But it's free to try. Reach out to us, we can set you up. we can give you some tips and tricks and examples and stuff like that, but ultimately it has to provide value on day one. If it doesn't do that, then we're not doing our job. Right. The AI pilots that run six months, nine months, ROI payback, all this stuff is like maybe a thing of the past because maybe the AI wasn't good enough. And I think now we're at the point where immediately it provides value and it helps you win. speaker-0: Absolutely. Yeah, I will provide links on the description of this video and also on the podcast platforms So as you can get in touch with Devin and and his team Devin, this has been a fascinating, you know discussion Thank you so much for joining me and sharing some of the work that you're doing with grant speaker-1: Thank you so much for having me. This was super fun. speaker-0: Thank you.