From Tribal Knowledge to Intelligent Maintenance: How AI Is Reshaping the Factory Floor

Welcome to the Industry 4.0 Club's fireside chat. The club's purpose is to engage the power of diverse worldwide talent to accelerate the global evolution to Industry 4.0. We accomplish this by hosting and supporting respectful, thoughtful, and interdisciplinary open conversations and creating quality content on wide-ranging Industry 4.0 topics to expand the reach of our community. My name is Mike Ungar, and I am the CEO of the Industry 4.0 Club. I have over 40 years of leadership and management experience Most of that with Michelin in manufacturing. I'm joined today by my co-host and chairman of the club, Mike Yost. Mike, would you mind introducing yourself, please?

Not at all. Thanks, Mike.

Welcome, everybody.

As Mike said, my name is Mike Yost. I am the volunteer chairman of the organization, one of its co-founders, and Mike's co-host here on these fireside chats. I've been in industry almost 40 years myself, first half of my career on the industrial automation industrial software solution provider side, and the last half of my career and counting on the nonprofit side. So really excited that today we have the opportunity to speak with Rhys Mullins, who's the founder and CEO of Metis. We're going to be talking about going from tribal knowledge to intelligent maintenance, how AI is reshaping the factory floor. Medis is an AI company that's on a mission to accelerate American manufacturing by solving the skilled labor shortage. Their product, Jack, is an AI agent that keeps production lines running when people have been doing that work for the past 40 or 50 years retire. Big issue for everybody. So Jack captures tribal knowledge and reasons across manufacturing data sources, including PLC code. work orders, schematics, and live equipment data so they can solve any problem that the maintenance team faces. For a little bit of context, before Metas, Rhys was a controls engineer at Tesla working on robotics and advanced automation. So welcome, Rhys.

Yeah, man, I appreciate it. I'm really, really excited to be here.

Excellent. We're happy to have you.

Yeah, it's good. It's great to meet you as well and have you on this fireside chat. So let's just— maybe we can just jump right in, Rhys. You know, every manufacturer, almost every manufacturer depends on a handful of experienced technicians who seem to know everything about a machine, its quirks, you know, it sounds different, its workarounds, how you handle whatever problem kind of arises, and its failure modes. So as these folks retire and move on, so what are manufacturers really at the risk of losing? And why is capturing that knowledge proven so difficult from your From your experience?

Yeah, I mean, like you mentioned, there's like 10 guys on a maintenance team and 2 of them have been there for 30 years and they kind of hold the weight of the rest of the team, you know, and you've got younger guys who are trying to learn. And, you know, our kind of theory and what we've seen in the field is that, you know, basically if you've got less experience on the floor, less know-how to solve these problems, then when, you know, the line does go down or, you know, something explodes like it does pretty much every day, It takes longer to fix that problem. And it's going to lead to more downtime and cost for the manufacturer. And then you're also not getting as much hands-on mentorship from the older guys to the younger guys. So there's a bit of a cost there. But primarily, if the guys who know how to fix the machines are gone, then when the machines go down, that's a big problem.

And when you say explodes, you mean work-wise and challenge-wise, right? You're not talking literally.

Yeah, yeah, yeah, yeah. I mean, it's just I say that because when I worked at Tesla, I would would come in every day and I'd be like, oh, what happened last night? And my manager would look at me and he's like, the line's exploding, you know? And that was just like our every day.

So— We used to see this, you know, you mentioned that. It's funny about the more experienced folks. We would hire somebody new, in my experience at Michelin, and to main— a maintainer, and we would say, well, it's gonna take 3 or 4 years to train them.

Right.

Because of that problem, right? And then even then you couldn't, really capture everything, but the 3 or 4 years, who's going to actually fund an extra person for 3 or 4 years in order for them to gain those things that don't happen that often? And that was part of the challenge, was justifying the hiring well in advance of the retirement so that you could gain the person, you gain the experience. You just don't, you just don't want to spend that money anymore.

Yeah. Yeah. I mean, I think you're, you're right. Like it, you know, the line doesn't just, I mean, it goes down every day. We, you know, we talk about that, but like there's thousands of different failure modes. that a line might have, and it's gonna take a guy 4 years to at least maybe see half of them, you know, and be effective in that way. So, you know, by us capturing all that and then just giving it to them, you know, we really, that's another big thing for us is accelerating the training and onboarding process for a new technician.

Well, and Rhys, you know, we've talked a couple of times and I think your perspective, you know, coming from the controls arena, even knowing what a PLC is, is different in the Industry 4.0 ecosystem.

Yeah.

You know, a lot of folks that they're talking about AI, they're talking about IoT, they're talking about— and organizations make investments in maintenance systems and MES systems and historians, all sorts of formal applications and stuff. But I do think they don't— a lot of organizations, a lot of people don't have the experience and the perspective that you do. So I'm wondering if you could sort of stitch together for us From your vantage point, knowing what you know, maybe even start with what a controls engineer does, right, and the value of it, and then tie it in with this futuristic AI and all this stuff you're talking about.

Yeah, so I mean, I can just talk about how this company kind of came to be in the beginning. A big part of my job was, I was a controls engineer at Tesla, and I'm in my office and a technician calls me and says, Oh, hey, Reese, you know, the line is down. The FANUC robot, or, you know, at the— this welding cell is, is faulting out and we don't know how to get it running again, right? Then I come down there and I, you know, bring my computer and I plug into the PLC, I plug into the robot, and maybe I look at the SCADA system that we had. I think about, you know, what happened over the past week that I can remember, and then from there I can kind of reason my way to what the root cause is and then make the fix. Maybe it's a PLC change, maybe it's a You know, robot controller change, maybe. I mean, honestly, sometimes a sensor is just, you know, dusty and you just got to clear it off and then it runs again. You know, stuff like that could be very simple. But anyway, that was kind of my job, was to figure out the root cause, tell them how to fix it, or make the fix myself, and then get the line running. And from there, that was kind of the idea of like, well, what if we just built an AI system, like an AI agent, in this case Jack, who was like an AI version of me?

You know?

Yeah.

And so now instead of the technicians calling me, they call Jack, and Jack has access to the PLC code and like the live machine data. It's using the equipment context from schematics and manuals and historical data, and we can actually pull in the literal logic, and the technician is saying, oh, okay, I'm getting a brake chopper fault on this motor. And then it can look through the logic and find exactly what bit, down to the exact bit, and explain it to them in natural language what the problem is. And I think kind of at a macro sense, that's the acute problem that we're solving right now. And then at a macro sense, it's like, okay, all these different things you talked about, the SCADA, the MES, the quality tracking system, just generally a data historian. This, they're all kind of solving different pieces of the pie. Like, I mean, even IoT, they're all kind of solving different pieces of the pie, and none of them are really a complete solution. And we kind of think that like it all starts in the PLC, it all starts on the floor. And if you can basically, you know, get all that data into an AI system, then it can actually like give you insights and not just say like, oh, okay, you're getting a high temperature here. Like the IoT system was like, oh, the temperature is spiking here. And maybe you need to go over here and check it out. Like, we can say, okay, temperature's spiking here, here's what we think the root cause is, here's like the literal exact piece of like logic or parameter that might need to be changed, and then be able to make that change.

So in that example where you're actually in the PLC code itself and it takes you to that bit that has faulted, right, or has triggered the machine down, that literally, back to your earlier point, could be the sensor that's dirty Right, that you could then tell them, right, which you know from experience because you've been there, but then you could teach somebody else quickly, this sensor is the problem. And then the next time they go, probably the first thing they do is they walk over and make sure that that sensor is clean.

Exactly. It'll be like, oh, we're not getting this sensor bit fired. And then, okay, now let me look through the historical records. Oh, looks like the past 10 times this has happened, And, you know, 7 times out of 10, it's just a dusty sensor. So then we can just, you know, just stuff like that that you pick up over years of the job.

Well, and there's probably nothing that would drive people battier than just saying, like, you know, all of these systems and, you know, all the complexity going on, and we got a dusty sensor, right? Yeah, yeah. You know, somebody knows it. So now if Jack can know it and inform the next one. Yeah, yeah.

We used to get— yeah, it's funny you mentioned the dusty sensor. We used to get Frustrated. So I'm going to go back a few years, right? But when something wouldn't run and someone would say, well, did you just knock the, the, um, did you just knock the, the, what was it, the electric eye?

Yeah, the photo eye.

Yeah, photo eye, just like the photo eye off so it wasn't seeing each other anymore. So someone would spend 10, 15 minutes trying to troubleshoot it and so go, oh, here it's a problem. and it would run again. You know, it's like, wait a minute, what? You know, and so being wise enough to— you're exactly right— being wise enough to have that reminder.

Well, so, and I guess to that point though, Rhys, I mean, when you have conversations with your audience that you're talking to, do they recognize the fact that being to that level of integration into the PLC code is important, or are they assuming from others that it that they're doing the same? Like, is it as differentiated for you in your conversations with people as I think it is? You know, because I don't hear people talking about that.

Yeah, I mean, I think so. We come in and I demo them our product, and immediately, like, the thing that really lights them up is like, okay, okay, Jack solved some problem, and he's telling them literally the exact bit in the PLC, the exact wire number in the cabinet, and telling them where in the cabinet this wire is, you know, and how to Okay, maybe check to make sure this is connected. Yeah, I think it's just something that like people don't necessarily think about until we talk to them about it, where I'm like, oh yeah, like you could actually have an AI model that is reading what's the actual logic that's written in the controller. Because most of the time these controllers are black boxes. There's a handful of people that know how to use it and they're just like, oh, it runs and it's run for 30 years. And like, I just, we don't ever look at it. And this idea that we could literally like embed AI into their physical control system. It is very new. And, you know, that being said, it is different. There's a lot of other companies out here that are starting to have some sort of bolt-on AI feature, like, you know, like a CMMS or an MES or, you know, whatever. And again, I think they're— the reason people have gone with us or are solving this problem this way versus with other people is because we can get to that level of detail. It's just, it's something— it's not something I don't think people thought possible. I didn't really think it was possible 5 years ago. And now it is.

So— Just one quick anecdote, Mike. I'm sorry, to your point about it just runs, we don't know what it does. I used to sell industrial automation, used to sell PLCs, and I got called into the boss's office one time, plant manager's office, and he started chewing me out because he got dinged for having spares in the maintenance area for the PLC stuff. And he said, we've had this on the shelf for 15 years and we've never used it, right? Which was like a compliment that our stuff never failed. They never touched it. Why in the world do we have— so I think it plays into your point, the fact that that stuff does what it does and not a whole lot of people understand what it does.

Yeah. Also, I guess one more point. There was this— yeah, this— a lot of this code doesn't change. So it's been on the controller for 15, 20 years. But I mean, I guess over the years it does get changed, but nobody really knows how it's changing. And with one of our first customers, like, we were able to like look through the logic and find, oh hey, you've got a lot of old logic, dead logic, inefficient timers and hold bits that are costing you throughput. And no one actually— like, the people who care about throughput don't know that they're like sacrificing throughput from inefficient logic because they don't know how to actually read logic. So then we can basically go in and democratize that ability so that at the high level you can realize, oh wow, you know, if we made all these big code changes, then we could make parts, you know, at 10% higher throughput or, you know, whatever. These gains can be very significant.

You mentioned something that's interesting to me about AI and the PLCs. I've heard also a lot of folks talk about that you need really, really strong guardrails because you don't want your AI to be messing with your PLCs, right? There's safety, there's quality involved there. How do you put those guardrails in place to ensure that you're getting the benefits of the knowledge that AI can bring, but you're not having AI do something that it shouldn't to the PLC?

Yeah, I think, I mean, that's a great point because, I mean, we're talking about like industrial control systems where like people are interacting with them machine. And I mean, obviously, you know, we don't want anybody to get hurt, or that's nobody's goal. So yeah, I think in the beginning it's all human in the loop, right? Like, you know, they say, oh hey, maybe we want to add some alarming to this program, maybe we want to, you know, change the logic here. Okay, the AI is going to go make the change, but then at the end of the day, the human still has to accept it or, you know, approve of it before it gets— anything gets written to the PLC. But, you know, frankly, like, I don't know, you know, other people you have on here that talk about this kind of stuff, but like, really, these frontier models as far as coding are at a level that like is unbelievable to us. Like, we've got very experienced software developers who have been building code for 10 years, and we trust these agents to go out and build stuff for our product. And it's really unbelievable. It's going to get even better. So it's Eventually, these models will be better than your best controls engineer. And that's— I think that's where we're headed.

Well, that's neat. So it really kind of leads me to this question about— a lot of folks are skeptical of the technology. I mean, what you just said is, for some folks, it's, what do you mean that the AI is better than I am? I've been doing this for 30 years kind of attitude, right? So how can manufacturers introduce AI in such a way that it earns the trust of those folks, of those 30-year employees and the frontline teams, as a way to really improve the way things work rather than replacing them? Or so can you talk me through how do you do that to make sure that you both win with what the technology offers but you don't leave the people behind in doing so?

Yeah, no, I think it's a great point. And like, we I, I, it's my favorite part of the job is working with these guys who have been here for 30 years and just, they've got so much experience and it's so cool to watch everybody work and learn from them. And I think at the end of the day, like just plain and simple, we're a company like us or kind of most of these AI companies are not necessarily trying to get rid of people's jobs. And I don't think like realistically we will. We're just trying to basically, you know, give them almost like, I don't want to say superpowers or something, but basically amplify their skills, upskill them to where they're performing at a higher level because You know, the whole theory here is that we've got this skilled labor shortage and it's not even about like reducing headcount. It's just like we can't find people to fill these jobs, you know? So it's like their job actually might become even more valuable. And our goal is to kind of take away some of the monotony of putting out fires all day so that now your technicians can focus on more preventative maintenance or like higher value maintenance. And, you know, even to the controls engineer, you know, comment. Like, there's so many factories out in the US right now that literally don't have a single controls engineer in the factory. And they have, like, they have, you know, they would love to have a guy who could come in there and work on all the PLCs and make the changes that they want, but they get a guy to come in there once a month to make these changes. And so now we're, you know, you're offering people labor that they don't even really have right now, as opposed to, like, reducing any kind of labor. And that's— we're not really seeing that in the labor market. I understand the skepticism. I've heard it all from different technicians that we've worked with just because I think there's been a lot of fearmongering in the media about AI, but I don't think the job loss is really being reflected right now, and I don't think that's the direction that we're heading. So, great.

I just want to take a minute and remind everybody that we are talking with Reese Mullins, who is the founder and CEO of Medicine. We are talking about moving from tribal knowledge to intelligent Maintenance and how AI is reshaping the factory floor. I'm wondering, Rhys, with the conversation so far, you talk about the agent, the AI, having access to in-depth levels of code and resource manuals and all that kind of stuff, which somebody needs to know in order to go out and perform maintenance. So is that the primary way that that you approach this as far as capturing all of the available resources like that? Or is there a way for you to capture the knowledge from the workers themselves, right? Are you actually asking for input from the maintenance folks or the controls engineers or the operators that becomes part of the knowledge that the agent has?

Yeah, I mean, it's a little bit of both, right? So like, I think the key insight is that like, okay, a lot of these answers to these questions, like maintenance, is not just in the manual. Like, if it was in the manual, then like everything would be easy. We just look it up, oh, just do this, and then great, line's always running. So I think the manuals are great, the schematics are great, all this stuff kind of builds to help or works to help build basically a mental— you can kind of call it mental model of the factory, like the factory ontology, like a How the model interprets the physical reality of the system, and the PLC is very important in that too. But yeah, to your point, a lot of this stuff just lives in people's heads. Like, they know the quirks, they know the workarounds, they know the stuff that isn't written down. And for that, we have kind of 2 different ways of capturing that knowledge. One, with just like— we call it persistent memory. So anytime a technician is interacting, or a controls engineer, whatever, is interacting With our software, we are looking for information that is not stored anywhere, and then we capture that and store it. And so now, like, I guess a good example was— I guess this isn't really tribal knowledge, but it's the first thing that came to my mind. We had a manager who said, hey, I want to make sure that anytime a technician is doing a change in a cabinet, he does a live-dead-live check on his multimeter to make sure that the multimeter is working correctly. And so now, from now until like the end of time, every time somebody opens up a cabinet, Jack tells them, hey, make sure you do the live-dead-live check.

Yeah.

And, um, you know, that's just kind of one example. And then as we've also been doing, uh, experimenting with like agentic interviews where we can sit down with older technicians and we look at all of their work order history that they've done for the past 10 years, and we can say, hey, like on this hydraulic press, like why do you— like how do you know that, you know, when there's ram drift that one of the valves is there's this broken seal versus like a solenoid issue or something like that, and they can kind of walk through their thinking. So we're, we're doing a few different things to capture this tribal knowledge, but it is super, super important to build, you know, because we're basically building like a brain trust so that whenever these guys leave, a company doesn't really lose any knowledge. It stays with the company.

So.

That's great.

So we've talked a lot about maintenance, specifically kind of focused in on the results that we get from a maintenance perspective. But because it's such an important thing to justify the investments and think beyond just that, what are some of the other gains that you're seeing in terms of results with the manufacturers you're working with?

Yeah, I mean, primarily, like, our biggest ROI is from just like helping them reduce number of downtime hours, which typically comes from maintenance. But some of the other applications have been, yeah, rewriting PLC code is really important. We've been working to solve quality issues like, okay, this, you know, part that we're making is not within spec. What are the parameters that I need to change on the piece of equipment to get it back into spec? So improving quality, reducing scrap, these kinds of things. But yeah, I mean, we primarily, we build for the maintenance team and, you know, with maintenance we kind of also incorporate controls and some other things. But right now that's our main focus. And then I guess the next level up is kind of management. So from management, you know, they— there's— okay, I guess this is an interesting point that we kind of came across. And this even goes back to when I was at Tesla working on this with my manager back then. But we want to basically close the information gap between management and what's happening on the floor. It's like you're a manager and you're like, okay, I'm running this factory and I have an understanding of what's going on, but then the reality of the floor a lot of times can be different. And so we basically, because we're pulling all this data off of the line, we can use AI to then like deliver, you know, insights and reports and close that gap so that they say, oh hey, like we've replaced the, you know, cooling tank 4 times this week. Like, there must be a problem here. And so, okay, now let's go figure out what the issue is because maybe it's on different shifts and they're not communicating.

Yeah.

Um, so giving— for the technicians, okay, those are those big gains. But then for management, yeah, we're really trying to close that information gap so they can make more informed decisions, plan for downtime, improve their PM procedures, create work instructions, um, you know, just for routine tasks for operators or whatever.

So yeah, if I could just like really quick just comment on that too, from going past, you know, when you— unless you're just doing simple assembly, quite often the reliability of the equipment's impact on quality is quite significant. So, you know, one can imagine that in a lot of different manufacturing processes that the reliability of the equipment is key for quality. Um, you know, when it comes to things like wear and, and, and those sorts of wear issues, they don't happen instantaneously quite often. They happen over time. So there's a tremendous benefit to see, to see how well the machine's not running quite right, but to know that the impact on quality is going to happen if it hasn't already happened. And so this can play a really big role, I think, and also quality results, as you believe too. Yeah, exactly.

It's like maintenance and quality are kind of you know, related, you know, their brother or sister, if you will.

Yeah, absolutely.

Well, that was— that's from your last point, that's where I was going to go, right? You have maintenance and quality that are really interconnected, but they're also different roles, different departments, right? And so you have all these different departments, and then you talk about management. So as you see AI continue to evolve, do you see each department having its own agents? Do you see it rising up a level to where there's agents that can work across the different silos? Like, what's your vision of what a future state looks like in leveraging AI?

Yeah, yeah. I mean, to me, it's just so, so exciting to think about. But yeah, I mean, it's like all these departments are sort of acting independently and you're relying on people to communicate with each other and all these kinds of things. They've got different softwares. And I think, like, you know, when you say, like, different agents, I mean, there could be thousands of agents working on a factory because even with our product, like, we have an agent Jack, but really it's one main agent with a ton of different sub-agents that we're kind of orchestrating all the time.

Okay.

So, but yeah, I think it'll be— you'll have— we basically need to bring all this data together, normalize it so that we can understand, you know, all the part numbers match and there's no— because when agents have They can rectify mismatching part numbers, but it just makes things more complicated. But basically, like, instead of having this huge stack of different softwares and different teams communicating with each other, like, you really need to consolidate into one system because they're all like— process is related to controls, which is related to maintenance, which is related to quality. Like, all of these things work together so that at the end of the day, we can have, you know, make parts off the assembly line that we're proud of at a low cost. Like, that's— we all have the same goal and we're all working to get there. And the way that we measure that, I guess, is like OE, you know? And so I think like consolidating, having agents that can work across apartments and departments, I think will be really huge. And then being able to eventually recognize a quality issue, figure out the root cause. Oh, this parameter needs to be set in the PLC. Okay, now the agent's going to go change that parameter. And then that can really be the closed-loop autonomous system that I think, you know, probably not for another few years, but like that I think we want to get to.

Yeah, I appreciate that.

That's fantastic. Well, we've been talking for almost 30 minutes now, and as we get ready to wrap up, I wanted to just ask Mike for a final thought and Rhys as well, so, you know, that our audience could benefit from. So, Mike, you want to start us off?

Yeah, sure. I mean, we've drilled in pretty good on the, you know, the value of the controls perspective in here. I think that's really unique. Unique for people to understand. And, you know, the point I think that Rhys brought up early on is that, you know, there can be inefficient logic that's done for a reason, right? But then it's costing you throughput and things like that. So the ability to even understand the interconnectedness of all these different components, you know, functions within a plant, systems within plant and stuff and the opportunity for AI to help us understand those and stitch 'em together. And I think it's really, really a powerful, positive future state to look at. And like I said, oftentimes we hear people coming in from a top-down perspective. And so it's really refreshing for me to hear, you know, Rhys talk about it coming from the, you know, the actual level 0 process control level on up.

Ground floor, yeah.

Ground floor, yeah, exactly.

Exactly. Reese, your thoughts as we get ready to conclude for our audience? What would you like to share?

Yeah, I mean, I think it's a really exciting future that we're heading in. I think that, like, I've got a super big passion for manufacturing, so it's great to see a lot of new companies, like, kind of coming up and things are changing and people are starting to embrace these technologies. I think we'll be able to make a lot more parts in America. We'll be able to have a lot Cheaper parts in America eventually. And yeah, it's just a really exciting time where we're really— I'm really loving what we're doing and we're, you know, doing stuff that I didn't— if I'm like, oh my gosh, if I had this tool that could like write PLC code when I was a controls engineer, like I would, you know, kiss somebody, right? And so it's really cool to see people starting to use this and these things starting to pick up. So.

Fantastic. Well, thanks, Reese. Thank you for taking the time out of your schedule to join us. Talk a bit about how AI can capture tribal knowledge and help manufacturers to manage through our, you know, this labor shortage we have, and even we know a more challenging retirement bubble. So thank you for that.

Yeah, thank you.

So the club, yeah, the Industry 4.0 Club would like to encourage you to take action. Listen to a webinar like this, read a— attend a workshop, read an article, but do something to advance the deployment of Industry 4.0. even develop a strategy for your facility or encourage your facility to do the same. We'd love to hear from you, hear what you're doing to, to make a difference. So drop us a note at our website at www.industry4oclub.com or on our LinkedIn page. We also hold these conversations weekly. So if you'd like to be a future guest on one of our shows, we'd love to have you as well. Again, reach out to us on our webpage or our LinkedIn page. page and let us know of your interest. And with that, Mike, thank you. Rhys, thank you. And we really appreciate the time today.

Thank you, guys. Really enjoyed it.

Thanks, Rhys. Appreciate it. Take care, everybody.

Bye-bye.