← Industrial AI Podcast
Industrial AI Podcast
How a chemical engineer built an agentic manufacturing troubleshooting assistant
00:41:58
Discover how domain expertise and AI merge to solve complex plant problems – no coding required. Can experience be replaced by intelligence?
Loading summary
Transcript44 lines
- [00:00]
A
This podcast is presented by nxai, your partner for time series foundation models and physical AI. Hi there. Welcome to a new episode of the Industrial AI podcast. My name is Peter Seaberg and I'm your host. Today I'm going to be talking to Scott Duncan. Scott is a chemical engineer and Scott and I today are going to be talking about the agentic manufacturing AI troubleshooting assistant which Scott built, by the way, without having any code experience. Hi Scott.
- [00:45]
B
Hello, Peter.
- [00:46]
A
How are you doing?
- [00:47]
B
I am doing good. I'm excited to be here. Appreciate you having me.
- [00:50]
A
Good. Please introduce yourself to our listener, Scott.
- [00:54]
B
Sure. Yeah. So Scott Duncan. I originally from the Jersey Shore, chemical engineering background, ended up moving south to, to Houston and I have been in chemical manufacturing pretty much my entire career. So around 10 years or so by now. So grew up in the oil and gas industry in the plants. So working for a petrochemical plant. So got my experience and you know, just troubleshooting plant issues and understanding how plant operations worked whatnot for first five or so years of my career. Then I, then I transitioned more into a, you know, energy efficiency carbon neutrality role where I was helping write the carbon neutrality net zero roadmaps for large chemical company. And then, you know, this next part of sounds probably cliche but you know, like a lot of people, you know, used chat GPT for the first time when it came out the, you know, November of 2022 and I was just pretty amazed with what it did and I, you know, I just, I, I needed to learn everything there was to learn about it. And you know, I guess I just kind of saw that. I just had a gut feeling the world was kind of go moving in that direction based on how advanced the technology was and just have been trying to orient my career in that direction ever since. And eventually after a couple of years of just messing around with it, I had this idea for this agentic troubleshooting assistant that you mentioned, Peter, which I'd be happy to go into as much detail as you want.
- [02:34]
A
Very good. I'll certainly do that. Okay, thanks for that. So before we get into the details. Yeah. Let's maybe quickly talk about and I'm not sure about in how far you know any details in discrete manufacturing because discrete manufacturing is what most of the time we do talk process manufacturing as well. Would you agree? If you do know that in roughly we talk, we, we use different terms. So we have a PLC in discrete and I believe you have a DCS in, in process manufacturing use recipes to produce chemical fluids, oil and gas products, whereas in discrete, we produce discrete products. Would you roughly agree to that?
- [03:23]
B
Yeah, I would. I'd say so. You know, my background is mostly in the continuous manufacturing space. So. Yeah, I mean, you're still, you know, in discreet. You still have like PLCs and, you know, you need to control your processes, but once you move on to the continuous, you need, you know, DCS systems and.
- [03:41]
A
Good. So let's get into our topic. With the help of Claude code, you built an agentic manufacturing AI troubleshooting assistant. Tell us, please introduce what this is about.
- [03:54]
B
Yeah, so as I mentioned, in the beginning, in the beginning of my career, I was just, you know, working in the plants and I spent a lot of my time just troubleshooting issues. So, you know, there might be an issue with a boiler or a cooling tower or a reactor. And it was a very manual and tedious process. You had to gather a lot of data yourself. You know, you had to find the data from all sorts of different areas. And then once you got all the data, you had to reason across all of that data. Maybe you had like, maybe you were pulling trends onto a graph and you were trying to piece together some story about why the, the problem was happening. It took such a long time and it was a. Quite frankly, they were difficult problems to solve. And I, what motivated the, the troubleshooting assistant I made was just thinking back to my time at the plants and thinking, man, I. How much of this troubleshooting in plants can we automate now? The AI is here. And so I kind of built this whole architecture inside Claude code that ended up being, you know, way better than I expected it. You know, what I learned was that the technology is further along than I had anticipated. And you know, as you mentioned, Peter, I, I built it in Claude code. And you know, the only coding experience I have is taking a couple MATLAB courses in College about 10 years ago. And I'd imagine that, you know, some of your of your viewers are familiar with Matlab and they probably even took a class or two in college and probably forgot most of it like I did as soon as they left the building. But, you know, I think, you know, before AI came out, you, the only people who could code were people that knew how to code, that they knew the syntax. Now that CLAUDE code is out there, that's not really the case anymore. And, um, there's, you know, there's starting to become this new skill set of. Because, I mean, through Claude code, all you're really doing is, is you have an idea in your Head. So I had this idea for a troubleshooting assistant. And all you're doing in CLAUDE code is you're speaking in plain language, just like you would to a normal, normal large language model. And CLAUDE code goes ahead and, and builds the code for you, it executes the code for you, it builds all the files and you know, you go back and forth with it enough times and over, I guess, a couple hours, you have a working product. So now, you know, before, you know, before, before CLAUDE code it was, you know, you needed to know how to code to make software. Now we're approaching this time where CLAUDE code makes it possible, where as long as you have an idea and you can articulate it in natural language, you can build a pretty good product. And I think just one thing to add in there, because I, I think it's important, is that there is still a huge need for software skills. I mean, you know, when I'm building through CLAUDE code, a big, I guess, gap is that, you know, I have limited knowledge in how the code works. Right. So I, when I go review what CLAUDE code made me, I'm not going to be able to analyze it like a software person would. And there's a difference between making something that works locally on my computer and something that's, you know, maybe enterprise grade that you'd use in a manufacturing facility or just a big company. Right. That's got to get through the IP systems and all that. So I just, just want to make it clear that, I mean, the, the tools make it possible to build things without coding experience. But I'm not going to say that you, coding experience is not important anymore. Right. I think there's a big distinction to be made.
- [07:54]
A
We'll get into that later on as well, into any objections, arguments, discussions that you have had on LinkedIn.
- [08:01]
B
Yeah, I'm excited for that conversation.
- [08:03]
A
Yeah. Right. Or as you said, the coding language is the English language. So maybe at the end we're going to be talking about what that means, what it means to education, what it means, what it means. I mean, as long as you are a domain expert and is that domain expert, does that person, starting with education, does that person, he or she need four years, six years to eight years, or, or, or at some point in time, is that person going to do, you know, a training on the job for a year or two? I don't know, I have no idea. But we can discuss that later. What should we do first? Shall we talk about, can you let us know the elements of what your solution consists of? Or maybe you first want to just walk us through the use case like, you know, give us an example of a question that you would be asking and then explain us what it is that you see. And then I just, before I forget, I was just looking at your YouTube. Is that one, is that a thing that we can reference to in the podcast as well for people who want to be looking at later so they can then be following on YouTube what it is that we're talking about here. So maybe you, maybe. Why don't you first walk us through asking a question and tell us what, what you're going to be seeing then what is going to be happening?
- [09:21]
B
Sure. So what I essentially built, I'll get to the question in a second. Was a, I built this application on top of a public data set that was for a coal fired boiler in China. So it's, it's, it's real data, you know, it's an actual boiler that existed. And yeah, one thing to keep in mind is that it's using public data sets is a little bit limiting because you know, best case scenarios you'd be using, you know, actual, your data and your actual plant. Right. So you're, you are limited in the data that you have available. But a use case for this troubleshooting assistant would be, you know, let's say there is a, an incident where the boiler is experiencing low steam temperature which could have all sorts of downstream impacts to your plant equipment. Essentially you'd be asking the troubleshooting assistant, you know, what is causing the steam temperature to be low? You know, because in, in a plant, you know, it's not always obvious. Let's say the, the, the operator gets an alarm, you know, hot, low steam temperature, high steam temperature, whatever the alarm is, there's so many of them always going on, it's not always obvious, obvious what's causing that alarm. So that often causes a pretty long investigative process that I was kind of leading to in the beginning of this video. So what I was kind of hoping was that in best case scenario, you know, if this existed in a plant, I wouldn't be asking after the fact. It would be catching the, the issue as it happened and then trying to solve the issue immediately where the, a main bottleneck of my tools that I don't, I don't, I don't have real time data coming in. It's just a prototype that has, you know, data from historian and I'm solving the problem after the fact, if that makes sense. So the user would ask, you know, why is the Steam temperature low. And the assistant would essentially go and grab whatever data it needs to solve the problem. Right. So if I'm thinking about how a problem is solved in general, it doesn't matter if it's manufacturing or if it's medical or really anything in life. You have to first, like you articulate, you know, what is the problem, what problem are we trying to solve? Pretty much. And then after that, okay, so what data do I need to solve that problem? So the assistant will go get whatever data it needs, you know, from however many databases that you have access to. In, in my tool you have, it has access to a historian, right? That's just time series data from sensors out in the field. It has access to alarm logs, any alarms that were coming in through the console, which are very useful to have here. It has access to a knowledge graph which essentially is just a graphical representation of how all of the assets in your plant are, are connected together. Right. This, this equipment is upstream of this equipment which is upstream of this equipment. You know, this valve controls the level in this tank. This sensor controls the, or measures the flow rate through this pipe. Right. So you have, if you have all of that represented in, in a form a large language model could take in. Now the large language model has an understanding of how your process works, which you will need. And then you have a vector database, which, not sure if that's a new term for anyone, but it's associated with rag. It's essentially just a database that contains some static documentation from your plant. So every plant's going to have SOPs like startup, shutdown, maintenance procedures. It's going to have equipment data sheets, it's going to have P and ids, pfds. My goal here was that the troubleshooting assistant would have access to all, all of those databases so that when it to comes, goes and solves a problem, it figures out, here's the data I need, it grabs it and then once the data's there, it, you know, reasons across it.
- [13:19]
A
Sorry. So would you say you give the, the application access to any information the human operator would have access to as well?
- [13:28]
B
I, I would say so. I mean, I think best case scenario is that whatever a human would, whatever information the human would have access to, I would want the assistant to also have access to. I mean, I would draw a line between, you know, having access to information and actually able to make changes to the process. Right now I would say that technology and the level of trust in the, in the industry is not there, or, you know, the autonomous operations aspects of
- [13:58]
A
large language models and then you ask it to, to tell you kind of what it's doing or I mean, so you are sitting there looking and the app will tell you, you know, I'm going into whatever database, XYZ, etc.
- [14:16]
B
Yes. No, I'm glad you brought that up because I mean, what. The way I designed it, one of the updates I made was that every time the, the assistant does something, whether it goes and grabs data from a database or it makes a hypothesis or it does like a reasoning step of some sort, it. It logs that step in a report that an SME can review.
- [14:41]
A
Very good. Now you already refer to some of the elements. Maybe we can spend a little bit time on some of these. The knowledge graph, that is an important element. I think you already explained. The knowledge graph basically shows the relationship between all the elements available. How do you produce such a knowledge graph? It's a knowledge graph, maybe I believe in, in, in the majority of, you know, environments, production environments. I assume the same also for your environment that has not necessarily been available as such. So assuming you would, you would need to produce that. How, how is that work? Does that, is that easy these days to produce a knowledge graph?
- [15:29]
B
So I would say it's very much depends on the situation. So for my, for my assistant, which is a little smaller scale, just to give people a sense of scale of my project, I like the historian. The database that it has access to has about 30 tags for the historian, so 30 different variables that it's measuring and about five days worth of minute data. So a good amount of data points. But that is not nearly as large as what you'd see in a large chemical plant, right? I mean, in a chemical plant you have thousands and thousands of tags and pieces of equipment, right? When you add up all the valves and sensors and pumps and reactors. So I mean like creating a knowledge graph for the project that I had was pretty straightforward, I would say. I largely just used a large language model for it. I had a description of the process that I found in a research paper and I had all the tags and I just put a prompt into the large language model, into Claude code and I said, just build me a knowledge graph. And it did a great job, right? It's small, it's, it's small scale. And obviously when you have a knowledge graph, you should review it, right? If you're going to, you, if you're going to have a large language model, do anything, always, always check the work. So I did that and it did a very good job. But now, now when we start talking about a, building a knowledge graph, let's go a level up here. Let's say we're now talking about a, an entire production unit at a, at a chemical plant. Right. This is a much larger effort, right. That I, I can't sit here and say that, oh, just plug it into a large language model, it'll do great, right? I, it'll help. I, I certainly think that the large language model will speed up the process, but the, the quality of data that exists in most plants is not as, not as clean as what I had access to. So you're going to have issues there with the cleanliness of your data and the availability of the data. So if you truly want to build a knowledge graph for a facility or even a part of a facility, there's going to be a lot of manual effort. Right now I would say, but right now I'm of the belief that there's a lot of boring work that needs to be done before the industry can really get benefits out of AI. And the data part that I'm talking about, like building knowledge graphs and cleaning up the data to help with knowledge graph construction, it's, it's necessary and I, I would just, I would urge companies to start thinking about that. How are we going to execute that step? Because it, it's not going to be done for you. And I think if we just pause here and don't do anything, I think then the industry will fall behind.
- [18:17]
A
I was just thinking, I had shared with you that a couple of years back I was in the United States and used and did a presentation for the OPC foundation that was about autonomous manufacturing, I believe. But the point here is not sure that you're aware of or work with OPC way like an information model. I'm thinking of. Is there like an alternative for. I mean the knowledge graph is one thing. Does it have to be a knowledge graph? Could it be something similar like for example an OPC UA information model where all your attacks. I think you have maybe a couple of thousand in a real chemical plant where at least all of them sit in a hierarchical structure. Not sure that then you're lacking the links between them, but at least in a hierarchical structure you define what they are. Is that something you could use if you would have that as a basis to build then a knowledge graph on
- [19:14]
B
top of or I would say I am certainly not married to a knowledge graph and only the knowledge graph. Right. I think that there's gonna be a lot of changes as like AI progresses. And I. The way that I think about it is that we just need a way to represent the relationships between all of the assets in the plant in, in the best way that a large language model can, can ingest them. Right. That's a knowledge graph for the, the solution you're proposing or really any, any other solution that someone way smarter than me comes up with one day. I, I think that that's, that's where we should move.
- [19:55]
A
So then you're using a time series database. Is that, you know, is that what you typically have? I mean, that's what we have in discrete manufacturing. Maybe the difference is not sure. In discrete. We may be measuring milliseconds, typically not sure if that's the case in process as well. But a, a time series database is what you, in a chemical manufacturing environment would always have available as well.
- [20:22]
B
Yes. Yep. I mean, I don't want to say always. I would say most chemical, most modern chemical plants and refineries will have, will have a historian that logs time series data. I would say that by now that is like a critical element to operating a plant. But you will see some older facilities not have a historian. And it makes problem solving, I mean, infinitely harder. I mean, if you can't pull past data to show what, what was happening in your, in your process, just. I'm sure you can imagine how much more difficult it makes problem solving.
- [21:02]
A
Yes. Almost not imaginable. Then I see that you've been using an MCP server. Now we've been talking MCP a couple of times here. We're following what's happening. It's an entropic solution. Right. Maybe you can share two, three lines on the MPC server. MCP server. Sorry, at the beginning it wasn't clear if it was going to be at the moment. Again, from where I sit, it's looking like it has got to be some kind of standard. Can you share with our listeners what the MCP server does? What did you have to do to. Could you use. Was the one available? Did you or CLAUDE need to produce one?
- [21:47]
B
Yep. So the purpose of the MCP server is just to act as a bridge between your application and the external databases that you're linking to. So it's essentially like a. The, you know, your application, like the client side, like Claude. CLAUDE code needs to speak the same language as the databases you're, you're communicating to. And that's where like the MCP server comes in. So for my particular use case, I had Claude Code build me an MCP server to Communicate with all four databases. So inside the MCP server you have a list of, of tools that you use essentially to go grab information from those databases. Now, from my understanding in, you know, if you were going to build this for your company, hope and you know, maybe some, some larger databases or if you're trying to connect to external software that maybe like a PI system that is more of a, you know, enterprise product, hopefully that, that external software has its own MCP server that you can connect to instead of having to be very creative and create one your own on your own.
- [23:13]
A
So I would, I would almost also assume that if within an organization, you know, there, some body, some organization will decide to build an MPC server, one or several ones, depending on what it is, and maybe that will be published within the organization. That's very similar to, I believe OPC UA works. So I mean, at some point in time, if this will be one of the potential standards, then you would know. Or if you work with Claude, maybe CLAUDE would be then smart enough if you allow it to first go and look for the availability of MCP servers but before it would build one itself. Does that make sense?
- [24:00]
B
Yeah, yeah, I agree. I mean, I think right now it's kind of a little bit like the wild west when it comes to this stuff, but I think, I think it's at least good that we have people, a lot of software folks that are smart enough to do it on their own. Like I think I know a lot of people who, you know, when the problem is important enough, you know, they'll figure out how to build one themselves. And then as this AI boom progresses, like you said, I'd have a hard time believing we're not going to start standardizing this. It sounds like taking a large language model application and connecting it to external software or databases is such an important thing as, you know, right. And developing a best practice system or standardizing that I think will have to come.
- [24:49]
A
Let's look a little bit at the state of the art of where we are, which I agree you said is really, really impressive. Nevertheless, and that is very important. You already mentioned one or two, of course we need to look at what you've been doing is building a demo. You have not yet built a solution in the real world. So let's talk a little bit about what were the pushbacks, what were the arguments, what were the objections? Maybe in the discussion where I could see it on LinkedIn, maybe you've talked to people, what are the things that we need to take care of on the past towards whatever in the next couple of months or year, until somewhere, somehow, we're going to make a real life solution out of this.
- [25:42]
B
Yeah, um, there are plenty of pushbacks I've received, and I mean rightfully so. I mean, it's, I think the demo is cool. It, it demonstrates, you know, what is potentially possible one day. But, you know, I think one of the biggest ones has been like, it's great that it gives you an answer to a question. Like if you ask it for a root ca. Root cause and it gives you an answer. But how do you, how, how do you know that it's correct? Like, how can you trust it? And one workaround I found I kind of alluded to in the beginning, which plenty of improvement can be made to this is, you know, you, you need to have someone like an SME subject matter expert. You need to have them review the answer. And it's important to have more than just an answer. Right. There should be a report generated that walks through here is exactly how the assistant got to, to its answer. Right? Because with, without that, that work shown step by step, I mean, personally I would not trust a large language model answer. So I think that helps. But then you, you start asking, well, like, what is actually what kind of report is necessary? Like, as I've been experimenting with this feature of my tool, it has outputted, you know, 10 plus pages of all of the work it went through, which not many people have time to go through that. Right? So it's, what is the, what's the correct way to go about this? Right. I think over time the industry will kind of align on something kind of like standardized, like we were talking about with MCP servers. So that's, that's one big thing.
- [27:21]
A
Can I, can I stay with this for a moment? And then we look at the other ones? What, what is the subject matter expert? Sounds like what I've been calling in general for whatever kind of job, a domain expert. What is the subject matter expert typically? Is that coming back to, again, to the experience, to the education and experience? Is that the person who has done a university, maybe not as much as that. More. Is it a PhD? What is that kind of person? Is that the person who has been writing code in the past, or is this a person who maybe was also depending on other ones and now could do more him or herself with the help and the support of an assistant like the one that you did?
- [28:05]
B
So I would say that subject matter expert is. I don't think there's a one size fits all definition for it. I think it very much depends on the situation. I wouldn't say it necessarily means it's the most educated person, like with, with the university, you know, depending on the problem you're solving, I would just say it is the most knowledgeable person about that specific problem. Like even if you don't have a college degree, right. If you, if you've been working with that equipment for 20 years and you just know it better than everyone. Right. I mean, I, I'd say you're, you're the, you're the subject matter expert. Right. But I think it depends on, you know, the problem we're solving. If we start needing to solve a problem that is like highly, highly technical in nature and maybe you did your, you specialized in this problem during your master's or your PhD, then. Yeah, I think maybe that additional education might help there. So, yeah, I mean, I would say the subject matter expert is going to very much depend on what problem we're solving.
- [29:07]
A
I'm with you. I'm with you there. Yeah. And I'm really asking also for, in general, not only in the process manufacturing, discrete manufacturing, know United States as an example, reshoring, you know, we want to build and maybe we, or you do not always have the people. And I'm looking for or same thing, you know, on the European side, other parts of the world, many people, you know, have worked now for whatever 40 years will stop working. We don't have the people. And I'm looking at this way of saying, you know, even if the people then do not have the four, six, eight years, whatever, but maybe less and a half years on the job with the help of the assist of the assistant, you know, can they come up to the level of a person which maybe as you say, has been doing this work in this environment for, for 20 years?
- [30:06]
B
I would say no, I don't think these AI or large language model solutions can replace decades of experience. That, that's, that would be a jump. And I, yeah, I definitely shy away from making that argument. I think where, where the line is though, is, is tough to draw. You know, how many, how many years of, of schooling, let's say, or like how many classes would. Would equate to having a AI capability enough to solve the problems? I think that is a really difficult question to answer. I mean, I think about these sorts of things a lot because it's certainly, you know, having access to a capable AI certainly increases your ability to solve problems, like quite a bit, I would say, and get work done. But I would Say, my prediction right now for the next handful of years is that, you know, as the, the industry and the world in general continues to adopt AI into their, you know, their, their company processes and solving domain problems, I think the most valuable person will be one with, you know, one, they have a lot of domain expertise in whatever it is. It could be chemical manufacturing, it could be medicine, it could be economics. And they understand the AI side and the technology side, they understand like how to build solutions. Because right now, I mean, I'm sure you'd agree that it's most people only have the domain expertise and, or they have the technology side. And then when you start, when those two people start communicating with one another, there's a pretty big barrier in between because both worlds are very complicated and you know, it's hard to communicate across that barrier. But hopefully I'm hoping that as time goes on, more people start developing the knowledge across both sides. Because if, if you had the domain expertise and the ability to build these solutions, I personally can't think of a more valuable skill set.
- [32:24]
A
I'm with you. So yeah, become, be, become the domain expert and then we talk about education. I am convinced that we're going to be spending a lot of time on that very topic around the world and then, you know, use, use the AI, the Claude code or any other to build a solution. What were any of the other pushbacks you would say that you had in any discussions?
- [32:52]
B
Yep, I would say the next main one was, you know, it's great that you built this on a small public data set, but it will be way more challenging to do this on a, on the level of a full production facility or even a full production area of the plant once the, the scope gets large enough. I mean, you have all sorts of challenges that come in. You have, you have to work with the IT departments to make sure that your data isn't leaving the company boundaries. Right. I mean, right now it seems like I don't know many companies who have figured out exactly how to, you know, just give any data to a large language model and trust that it's not going to be used to train the future models. Some companies are working on it, but that's a big issue right now. And then also, you know, like I said before, my solution right now is a local system. I'm actually about to put it on GitHub for those that want to download it and use it themselves. But in, in reality, in, in, in the manufacturing industry, you're going to want enterprise grade tools that are, they don't break at 2 o' clock in the morning. Right. That you can use them whenever you want, anyone can use them. So that is a, an additional challenge. So I'd say that was the next biggest one which I don't have like a one size fits all answer to that. I think it's more, you know, that's a very valid concern. It's, it's, there's going to be growing pains but I think that we have a lot of smart people in the industry and the different organizations like manufacturing it legal, we all just need to just like any other software tool that we, we make and give our employees access to, we're going to have to work together to, to come up with the best solution.
- [34:55]
A
Yeah, and I saw there was a discussion around benchmarking formats for evaluating so that all of us, you know, if you're going to produce one, somebody else to, is going to produce one so that we can kind of test what it is that we are producing before we're going to put it into the real world.
- [35:14]
B
Yep. Yeah, no, that's, that, that's kind of a little bit tied to the first part of just how do you, how do you trust them? And I, I would, I haven't, I'd be curious to read more about the benchmarking you just mentioned because I haven't, I haven't read about that. If you have, if you can share it with me. But the, I would, I would hope that as we start introducing these AI solutions to the industry that you know, we start small. You know, we, we start on helping solve problems that don't really have a safety element to them. You know, if, if something goes wrong, then it's not the end of the world. Right. Then you, you kind of prove them out there, right? Get like, let's prove that the tool can provide an accurate solution consistently and then once we've gotten beyond that point, we might start, you know, expanding the scope.
- [36:10]
A
Very good. Let's, let's try to bring it all together. Where do you see this moving? Are we going to be moving towards lights out factories, discrete or process? Or do you say no? As we've been discussing today, we say the subject matter expert is the person to decide, to decide what he or she, what decision they're going to be making. So is AI going to be taking over in the next 10 years or do you say no? As far as you can see, there's always going to be the human, maybe there's going to be less humans. Where do you see this Moving.
- [36:51]
B
So I don't see anything like a lights out situation in a, in a continuous chemical manufacturing facility anytime soon or, or ever. I mean, I, when you're talking about the, the time horizon of forever, it's hard to predict what's going to happen in let's say 100 years. Right. But like, you know, in the next 10 years. No, I, I don't see anything like that happening. I, the way that I see this shifting is I think the, the industry will slowly start adopting solutions like the one that I provided. And I'm sure I'm not the only one making these right now. I think many, many people are, are creating similar types of solutions as the one that I spoke about today. And you know, I, I've noticed a, a pretty big shift in, from 2020, 2025 to 2026 in terms of AI adoption. I think up until the end of 2025, it was a, a, a big buzzword that people didn't really know how to use AI or what to use it for other than coding. I think the software industry figured out pretty quickly that there's plenty of coding capabilities with the training data and these large language models and they went, they dove into it pretty quickly. But the rest of the industries, I, I don't know, manufacturing, I don't think there was a huge understanding of what can we actually use these tools for 2026, I'm noticing a shift here that, oh, okay. It sounds like we can use these large language model tools for several different things. Like, you know, the troubleshooting assistant that I'm, cause I'm creating, it can help with data analysis, you know, many other things. I think over the next handful of years I think we're going to start trying to implement these solutions and there's going to be a little bit of a slow period where like, you know, there's going to be some growing pains and figuring out how to consistently get it to give you correct answers. And then hopefully as we start implementing these solutions and they are able to give us some measurable impacts, we might start scaling them right from, you know, from pilots to, you know, expanding them across to more than one production area or even more than one facility. Right. And I think the large part of that decade will be kind of figuring out how do we take it from this, you know, pilot mode that we're in to scaling it. So I'm hoping by year 10 we'll see tools like the one that I'm suggesting just be commonplace in plants. That is my hope. There's a long way to go I think. But I see that being possible and then you know, once that happens, I think we will be looking at what, we'll be looking at plants that run better. I think, you know, as long as we are keeping the human in the loop and not just automatically doing what the AI says without checking the work, I see the plants having less downtime and even if they do go down, I see us getting the, the plants back up more quickly and which leads to more production hope and then you add in some of the energy efficiency gains that you can get from the plan from these tools. So yeah, I, but I, I do want to say I don't, I don't see these, these tools just replacing people. I don't see like you know, 10 years from down the road we now have half the amount of engineers and operators in our facilities. I just see that we are going to be seeing that the employees use their time differently. You know, they use their time to solve more challenging problems rather than firefighting.
- [40:22]
A
Great. Thank you very much Scott. It's very impressive what you built. It is a demo and you listeners can look at it if that's what you like Scott to share on YouTube. 45 minutes I believe it's very impressive. It is a demo nevertheless you built it without real well coding experience. You don't have to code. Coding is in the English language. Thank you very much. And yeah, looking forward to be following what it is that you're doing what others are doing in this area and, and see you know when, when the first real life AI based solution is going to be working in, in a chemical plant. Thank you very much.
- [41:13]
B
Yeah, for sure. I do want to say real quick that I do have a couple of YouTube videos but I have not made one on this specific tool yet. I am in the process of doing that so I just want to make it clear that I don't, I don't want to confuse the listeners.
- [41:26]
A
No stress. I think there's one that works more with gene models I believe but at least people then people can and visit you on, on LinkedIn I guess and then they can follow you and who knows, maybe you're going to be making a video on this topic as well. Scott, thank you very much for your time. Have a great day. Bye bye.
- [41:45]
B
Yeah, you too.