The Startup Defense
The Startup Defense explores the intersection of commercial technology and defense innovation. Callye Keen (Kform) talks with expert guests about the latest needs and trends in the defense industry and how startup companies are driving innovation and change. From concept to field, The Startup Defense covers artificial intelligence, cybersecurity, mission computing, autonomous systems, and the manufacturing necessary to make technology real.
The Startup Defense
AI Defense Manufacturing, Enabling Versus Eliminating, and the Problem-centric Approach with Terry Miller
In this engaging episode of "The Startup Defense," Callye Keen delves into the fascinating realm of AI and generative modeling in the manufacturing and defense sectors with Terry Miller. Discover how AI technologies are reshaping industries and the challenges they face. Dive deep into the vital importance of defining clear problems before seeking technical solutions.
Topic Highlights:
[00:00] Introduce Terry Miller
Callye Keen introduces Terry Miller, Executive Director-Global Advanced Analytics at Johnson Controls, is an experienced Analytics service leader with a demonstrated history of working in the industrial automation industry.
[01:23] Unlocking Manufacturing's AI Potential
Explore the uncharted territory where manufacturing becomes a playground for AI, transforming it into an engineering, IT, and cybersecurity job.
[08:00] Machine Learning Engineers: A Scarce Resource
Terry Miller highlights the scarcity of machine learning engineers and their crucial role in scaling AI solutions in industrial applications.
[16:00] AI as an Augment, Not Replacement
Delve into the role of AI as a valuable augment in production environments, reducing downtime, enhancing productivity, and mitigating staff shortages.
[23:00] AI in the Defense Industrial Base
Examine the unique challenges and vast opportunities AI presents in the defense industry, driven by a multitude of small, specialized businesses.
[27:00] Defining Clear Problems for AI
Understand the importance of having a clearly defined problem before seeking technological solutions, and how this clarity can lead to impactful outcomes.
"In industry, you'll find examples where someone has realized the benefit of technology but then discovered that they didn't necessarily need a new technology; they could solve the problem with existing capabilities." Terry Miller
Callye Keen - Kform
https://www.linkedin.com/in/callyekeen/
https://twitter.com/CallyeKeen
Terry Miller - Analytics Leader
Experienced Analytics service leader with a demonstrated history of working in the industrial automation industry. Skilled in "Big Data" applications, IoT, IIoT, Predictive Modeling, Manufacturing Analytics, Engineering, Business Development, Project Engineering, Process Automation, and SCADA. Strong sales and marketing professional with a Master of Science (MS) focused in Predictive Analytics from Northwestern University.
you find that a really clearly defined problem doesn't need some new technology. You have the ability to solve that problem internally and you can use heuristics or some other means to do it with your existing capabilities.
Speaker 2:Welcome to the Start of Defense. My name is Callie Keen. Today I have Terry Miller. Now Terry, you're an expert in AI and ML in industry 4.0. So that's a fancy term for smart manufacturing for everyone, and the listeners know my background's in manufacturing. I am obsessed with actually building things. We got connected from a mutual friend who's been on the show. So, Russell Adel, let's talk manufacturing. Let's talk about why this is important, what interesting things are happening in the market and, more importantly, to kick this off, what are you passionate about right now?
Speaker 1:Probably everyone's familiar with chat, gpt and the commercial applications of generative AI. I think, as is the case with a lot of these technologies, there's a lot of hype and a lot of promises of what the art of the possible is, but I think in execution it's much simpler than that. But I am bullish on generative and manufacturing, specifically Any sort of industrial application where it can be an augment to service technicians or production supervisors. I think there are a lot of benefits. But yeah, I'm passionate about generative. Beyond the hype of the commercial implications, I think there are a lot of really practical utilitarian use cases for generative and industry specifically manufacturing.
Speaker 2:How do you feel about generative in CAM solutions, so computer-aided manufacturing, programming, CNC machines.
Speaker 1:Yeah, so I think we can even pull back further than that application, but sort of the broader programming field. I think as an augment it's not going to replace programmers and I think not in the same way it's not going to replace artists. I think that's a concern for a lot of folks. But I think as an augment for code debugging, for sort of rudimentary simulation package or ways to test code snippets before trying them in production, I think there are a lot of benefits Troubleshooting or problem solving with machines that are actually in production. I think generative has a huge potential impact in that respect. I know we used it my team used more in an enterprise setting, but we were training models for an enterprise application and we used in our ML Ops platform. We used generative AI. It was a new feature of the platform but we used generative to debug our Python scripts and it was a very powerful tool for that. Very powerful tool Not a panacea, but a very powerful tool To give us some context and just you can disclose what you can.
Speaker 2:But where have you implemented Industry 4.0 or ML AI applications in manufacturing in the past?
Speaker 1:Yeah, this is a fun topic because I think it highlights a few things. So I got in this, I think, fairly early. I finished a graduate degree in this space and I started in 2014 and finished in 2016. And sort of while I was doing the graduate degree I was working full time but I was working for a major automotive manufacturer. That wasn't my employer, but that's, I covered them for my employer. I started just doing traditional. I started helping them design in traditional vision inspection systems, laser profile measurement systems, quality, traditional automation systems. That helped with quality. Well, they you know some of the engineers that I had close relationship with figured out that I had could write code and train algorithms.
Speaker 1:So we did a couple of what I call Skunkworks projects where we just used, you know, sort of on-prem data ingest technologies that I had available and did some projects and one of them I won't get into the hairy specifics but one was testing motors to go into cars, obviously, to test for improperly seeded or missing components in the motor by their acoustic signatures or their acoustic profiles. So they had test cells set up where they would test them and Bob, who's been doing it for 40 years, would listen to see if it sounded like the motor should sound, but they had quality issues that were sort of embedded in the process because obviously that's not a foolproof method. So we did sort of proof of concept with just off the shelf microphones but a little filter, preprocessing and then some algorithms to evaluate what does normal sound like and establish that as a baseline and then catch if there are different acoustic signature profiles from what normal was. We could catch that algorithmically. We were able to do that. They wanted to get as specific as being able to tell what components were missing and that would have been a much more involved project and we never got around to that.
Speaker 1:So that was one that was really interesting and enjoyable. The other was and we didn't get to execute this because it turns out robot OEMs don't necessarily want to turn over access to data that comes out of their arms. We wanted to evaluate the relationship between a six-axis robot's path so the yaw pitch roll the path of the robot and premature servo motor where by access or for all axes, I was able to do that on a test cell. But in production it's much more challenging for a variety of reasons, but the main one being it's hard to get all of the OEMs to agree to turn over data that's captured in the arm. Those were a couple of projects that I did. There are others, more quality based, but those were the neat asset based or mechanical system based projects that I worked on.
Speaker 2:There's this big push in manufacturing and it seems as long as I've been in manufacturing, there's this push to get people interested and get them into the industry. Defense has the same issue hey, let's get people in here. We want smart people in, we want to get people into this skilled trade or into this space. And it's difficult with manufacturing to find a machinist, to find a fabricator, and instead of going, hey, I'm going to skip college and I'm going to go into manufacturing, which by all means is a good path to a good job.
Speaker 2:I think we have largely missed this other avenue, which is that manufacturing is becoming a engineering job, an IT job, a cybersecurity job. It's becoming a very high level, interesting technical trade, and so the types of projects that you're talking about, this sound attractive to an engineer, to a programmer. Does somebody that's interested in the sciences to come in and say, maybe I didn't really think about this space, I thought about doing a tech startup of X, yz, but maybe I should go look at these opportunities there in the defense industrial base or there in the automotive manufacturing space, because I can apply and level talent, you can pretty much. It's a big white space, right, and it's a huge dollar space, but those sound like very interesting projects, right. They sound like something that I would want to do. I don't tell me what you think is like the impact of doing something like that. It might seem minor, but when you're talking about working with a large scale automotive company, the small changes might have dramatic results, right?
Speaker 1:I cannot agree with that point more and I've seen both sides. I have a good network of folks that have, you know, work in silicon valley and the vc community that's active in silicon valley. Here's the dirty secret and I don't know why anyone would choose this. Really smart people I do know why. Really smart people to graduate with technical degrees from the standards, the m. It's the caltechs, the places where their funnels and to fang and tesla and the technology companies. The dirty secret for those people is, outside of the really special engineers, a lot of those people spend a lot of their time doing just really rudimentary tasks like a, b testing bits of software and just things that are that I haven't done it. So this is an assumption on my part, but they have to be Boring.
Speaker 1:I would assume that a person with really intricate technical skills that not a lot of folks have would want to solve these big problems with work on fun things like robots, industrial robots in a manufacturing environment. So I would say I totally agree with you. I think the cash a and the pay Of the big tech companies is what lures the brightest minds, but I think there's much more interesting work to be done In the automotive wing of the southeast, the defense industrial base, a lot of our manufacturing and energy spaces. There are some really interesting projects to be had, problems to be solved, that I think those folks, if they had the opportunity to see that, it would give them pause. Now the other aspect of that is I don't think that the pay is comparable to what they get to a b tests little bits of software in silicon valley and I think that's a gap that has to be filled. But the work is definitely interesting. I have seen that.
Speaker 2:I have seen both sides of that and I think your point is a good one yeah, I would say that If somebody is thinking about creating a startup, the upside is there because the market value is there. But I definitely understand that. Yeah, if you go work as a middling engineer at facebook, you could be making two hundred or three hundred thousand dollars in san francisco locally here in the dc area. They're probably gonna put you To work a little bit more to make that money. But if you were saying I am a specialist in this area, some type of data analyst specialist, you could be making that money or a little bit more. So to make the leap into manufacturing, manufacturing has to look at this like how do we create a startup? So I push back on a lot of the manufacturing industry leaders that are saying, hey, this is a good job and people should come in and see what's available here. I think it's our kind of our obligation to train people up but also to create competitive jobs when we can hire and attract top talent and eat.
Speaker 2:There's. There are a couple of venture act manufacturing companies that have applied mass intellect and even, just recently, ones that arguably are just really just Tech enabled machine shops. They've attracted tens of millions or, you know, hundred million dollar plus of investment and I look at it as somebody grew up in A machine shop. I'm like that's just a machine shop. They have robots and they have talented engineers, but their plan is great. But really I understand the premise is they're saying there's not enough people here on the ground level. You're not. I can't go out and hire five machinists. I can't hire five machinists this year unless I want to fly people in from across the country. But in this area I could hire three or four data scientists. I could hire a roboticist if I wanted to. I could hire a roboticist more easily than I could a churning man machinist. I see what they're doing, so you bring up an interesting point.
Speaker 1:I think there's a distinction between a generalist data scientist and a machine learning engineer that can execute projects on industrial assets. With the latter, russ and I talk about this a lot on our show. There are only, I think, in the US. I think there are, and of course a lot of these we've brought in from abroad, but there are only 22,000 PhDs in machine learning or AI in the US and I think globally the number is not much higher than that, like 26,000. Maybe hire 37,000. Last I look it's been a while. So the point is that the domain and really that's the choke point that's the bottleneck for scaling AI solutions in industrial applications presumably the defense space as well Just aren't enough people that know how to do it, and I think data scientists can do linear regressions and train linear models and do things with cleaner data sets.
Speaker 1:But industrial data is really really, really hard to wrangle. You have considerations of the sampling frequency of the controller and if it's every 100 milliseconds, you're getting a reading every 100 milliseconds. The cost of storing data at 100 millisecond resolution is astronomical. So how do you the data ingest part? How do you? You have to write some compression logic and you have to do things. You really have to know the space to take into account the nature of industrial data.
Speaker 1:And the second point about that is industrial data is very non-linear. If I'm training an algorithm on who's most likely to buy Widget X in an enterprise, that's pretty standard dynamic or static data. It doesn't change a whole lot. But in an industrial setting and I've seen this firsthand you can have machines side by side on a production floor and installed the same, made by the same vendor, the same integrator, and they behave differently. And it's just voodoo. You could have different air feed to the machine and the filter in one line is different. The installation torsion could be different on one of the two machines. There are a number of things that could make the two machines behave differently and you have to be able to capture that in data. You have to be able to capture that to get models that make sense and it's a real challenge.
Speaker 2:It's definitely interesting problem space to pursue and kind of to fold the conversation back on itself. Where do you see generative AI then supplementing some of those challenges and making use of the data, so going from that data acquisition to data intelligence? Where do you see that playing in the ecosystem?
Speaker 1:I see two spaces and both to me I'm really really interested in and how I go about getting involved is still a little murky. So one is an augment for technicians. So you have the typical production environment that I've been in and manufacturing. You have hundreds of vendors. If you're taking into account servo motors and robots and cylinders and power supplies, there are hundreds of vendors. The maintenance staff is not going to be expert in all those vendors. You can train a generative model on all of the vendors and assets that are in the production environment and really create sort of a technician assistant. That would presumably reduce downtime, improve productivity and potentially even mitigate some of the shortages that I know a lot of these folks have in maintenance staff. It's hard to find quality people and a lot of them are retiring out. That's another problem and I think a generative model would be a real augment in a complex production environment to boost productivity with the maintenance staff.
Speaker 1:That's one the other, and we talked about it earlier. The other is and I'll take the automotive manufacturer that I work for. They had a proprietary PLC that they use for all of their global locations and it's a repurposed brand that everybody knows, but it's specific for this manufacturer. Well, the number of people that are really good with programming, that PLC they're hard to come by. It's not something, it's not a common brand, it's unique to that company. So as a programming augment and it's not limited to that PLC, it could be any asset that's in the production environment as a programming augment, how do we get a new program, a new line in with more efficiency, up quicker?
Speaker 1:I have to be careful here because given away all the secrets but there's also probably a path forward to take faults from assets and feed those into a trained generative model and get a sort of a work order protocol based on what the faults are. So there are again. I would think I would. For anybody that might hear this, or for just manufacturing generally, the use of AI is best viewed as an augment. It's not replacing anything. Right now it's not sophisticated enough and that's where the hype has to be. We need folks that understand that distinction between the hype and what's executable. But AI as an augment is incredibly powerful and to boost productivity with existing staff, reduce downtime, reduce the consumption of raw materials things that are big cost drivers in the production environment AI can really help with that.
Speaker 2:We have a piece of equipment and it has a very simple ML feature. It just listens to the cut and so it's Pretty intuitive. If there's supposed to be a tool cutting and there's no sound, probably a broken tool. Some very simple things. But you know more what it does. It reduces chatter right, moving rpms up and down so you get better tool finishes. But that's a fairly simple application. But also it has a fairly trainable data set.
Speaker 2:Whereas a lot of the Pine, the sky, blue sky, thinking that people have for manufacturing and they walk through, this is what the new bmw plant is going to use Omniverse for complete digital twin and all these things are connected. That is fantastic. But I work in the defense industrial base and the vast majority of machine parts and all parts that are for the defense industrial base are produced by very small companies, generally their Owner operated companies with under thirty employees. So there's thousands and thousands of these that are run by. Somebody was a machinist, they started their machine shop. Somebody is a fabricator they started their contract manufacturer. So there's thousands of these companies and so they handle the low volume, high complexity work that makes the Defense industry actually go right. There might only be fifty of a part ever, or there might be fifty and then two every year for ten years. That's pretty common in the space. Of course there's Volumes that are higher, but making thousands of something is considered fairly high volume in manufacturing, whereas in automotive that would be a pilot right. There would be like a test quantity.
Speaker 2:When you take these blue sky, oh, I bet it could program any part. It could show me exactly how to set this up. It could Show me the right fix string or how to design an end effector for this part. You know right off the bat it's really it's an anything to anything to anything like multiple, difficult problem because the parts are variegated, the machines are variegated, all those businesses kind of operate their own way. There's no unified way that they they operate and all of the inputs are from external companies which they have different ways of designing parts, creating packages, creating models. So it's very interesting that you're taking this and I don't know if you can see my sign behind me. My son says and height and height was my old podcast.
Speaker 2:You know I am a technologist, I love new tech and I like to track kind of where it goes. But I think years ago went to. I am TS and One of the larger vendors. They were showing off voice control for their mill. I thought that's the cremos. The craziest, most useless thing I've ever heard is cool. I mean, you know it's cool to Shout at your machine and get us to do something, but what scenario is this gonna help me?
Speaker 2:On the other hand, I want digital twin to work right. We've had Tim shinbara on right from amt. We've had a beast code on it. Does digital twin for destroyers and jets and for military equipment? And I'm very interested in the space because I like the idea of having a digital connected world and be able to pull sensor data and usage data off the train people yeah, three print replacement parts. But in in practice, the defense industry has been really difficult because there these bespoke systems or very low volume systems, and the problems are very gated since, like everything's a Million or billion dollar system that there might be two hundred of. So there's only five people are ten people know how to work on something. I see the progress, so I'm really bullish on it as well, but I just am really anti Hi about industry. Four point oh, about AI in general, what you describe is exactly correct and it has its parallel counterpart.
Speaker 1:Problem with the ability here is sort of the these are the guardrails for Technology, including AI for industrial systems that's an umbrella term most of the applications themselves. So the production, as you explained it, especially for defense, you have unique applications that aren't necessarily scalable from an application standpoint and then you have the compounding problem that there are very few people who know how to scale or execute the technology part of that and there's no, it's not like Siri, where you can train it and they still have to retrain that algorithm. That's the big misconception. That's not like you build that once and send it out of the wild. That's continuously retrained. It's why you agree in terms of service. When you sign up with your iPhone, you agree to send them data because they want to retrain those algorithms.
Speaker 1:That doesn't apply in industrial settings, because there's not one model that's going to work for even 5% of applications. There are too many distinct applications and not enough people that know how to do the bespoke solution. So I would say this you'll find examples where CompanyX has really realized the benefit of TechnologyX, whatever that may be, but they've invested in a team or a person or they've spent the money to bring in the resources that know how to uniquely engineer that solution for their environment, and they can't then just scale what they've done to some number of manufacturing sites or plants around the world just because you have to replicate that model every single time. And I think that's what industrial companies trying to sell platforms, like some of the ones I've worked for that's the problem that they've run into is, the consultancies have told them you need to build this platform and you can scale it to all your customer base, but they left out that somebody has to engineer this every time, and that's the challenge that they all run into.
Speaker 2:On a really positive hey, let's move forward perspective. Let's ride the hype a little bit. So people are interested in AI, they're interested in ML, and so when somebody I talked to a startup and they're like, hey, we're trying to pursue this, I do steer them towards manufacturing applications and defense applications, because I feel like there's just way more white space there. There's less people competing and the value of an actual solution is much higher. So if they're trying to start a new company or they're trying to look at what's a knowledge space I could explore, hey, come over, look at this thing, let's talk about what's happening in this sector. Let's talk about what's happening in this sector, but there's such a high barrier to entry because, unless you're in the industry, this really not like, it's not as immediately intuitive or accessible. But what are some ways that people can explore this concept? Because we have general technologists listening to this, innovative people. What's a way that they can kind of get involved, dip their toes in the water and be like, maybe this is something I should pursue?
Speaker 1:I would say and this is the way I start all my projects so I've found and it doesn't matter where I've worked that if someone's coming to me they wanna see what AI can do for their thing, their application, the first thing that I do is have them quantify what metric is this problem? What is it now? So if you're currently consuming X kilowatt hours of power per year and you wanna reduce that by 3%, so that's the problem, or that's the outcome that they wanna achieve, Well, what problems need to be solved to deliver that outcome? And then you work backwards but to attaching technical solutions to those problems. So I would say the best way for and I think this is a non-negotiable, I think the best way for someone to dip their toe in the water is to really hone in on what outcome they wanna achieve.
Speaker 1:What are the big strategic pillars of the enterprise? Is it reduced energy consumption? Is it improved quality productivity? What are the big strategic pillars that they feel would impact growth of the organization or are strategic for the organization in one way or another? And then what problems need to be solved to deliver that outcome? And so, if reduced energy consumption is the quantifiable outcome that they want, well, what are the problems leading to? Why can't you just use less energy today and what would need to be solved, and then have a really clearly defined problem for someone to work on and then hopefully you can find smart people that can attach a technical solution to that clearly defined problem. I found, given all the constraints that we've talked about previously, I found that the biggest hurdle to get over is a clearly defined problem to solve. Most of the technology discussions are fishing expeditions and things that folks have seen work in some other environment, and would that work for us? What is your problem? And I think a real, clear problem to be solved lends itself more readily to a technology solution.
Speaker 2:Yeah, I love that approach. It's very pragmatic and when we run accelerators we see generally the problem that people have is they haven't identified an actual customer and then they haven't figured out what problems that the customer actually has. So when we look at what it takes to raise capital, I just look at those two things and we have some very simple actions that they need to take. And if they don't do those, it's kind of like why even bother with tech scouting and IP this and looking at grants and partnerships? If you can't figure out what you're actually trying to solve, why are we doing all this other work? That's kind of irrelevant.
Speaker 1:You know, what's interesting to Cali is if that exercise is undertaken properly, if that exercise of what are the big strategic goals of the organization, what are the objective all sort of metric based outcomes that we're trying to achieve and then clearly defining what problems need to be solved to deliver those outcomes and some non trivial percent of cases, you find that a really clearly defined problem doesn't need some new technology. You have the ability to solve that problem internally and you can use heuristics or some other means to do it with your existing capabilities. Not all I think a lot of times it's just getting that clarity piece right really defines what, if any, technological solution you need and you can do have a much more targeted approach to solving problems that way.
Speaker 2:Yeah, I'll pick on crypto, because everyone's kind of sobered up from crypto. I have friends that had very large crypto projects. I have had people in different accelerators with crypto enabled this and that and inevitably when you pull back the curtain of what they're actually trying to do, most of the time it wasn't something that even required blockchain as an approach, and that was the problem why I would say hey eventually. I don't think this is a good avenue, or you really don't need to promote this in your literature, because the problem should speak for itself and one Unifying fundamental fact and you can take this as a physics baked into reality part of life is that people have problems and organizations are made out of people. So If you want a really great problem, you can go out and talk to people. You will find a problem worth solving, because if you solve those problems, there's just secondary and tertiary problems. There's downstream problems, upstream problems. There's always problems, right, anytime you're trying to do something, there is a problem and by talking to people, you can better define what the problem. Is very difficult to define what the solution is from somebody that's experiencing the problem, but People are really good at complaining and they're really good at describing pain. Our memory, our brains are wired to remember pain, right, and so, yeah, usually when we go down this road, I think of it as like can this be solved by a brick on a string? You know what's that? Usually can be solved very quickly, which is Great for somebody getting started.
Speaker 2:Whatever the tech equivalent of that very simple solution is because you can start validating with customers. You can start actually building a business or learning the technology without having to make this Ultra elaborate cutting edge solution that you need to hire the people that get paid three or four hundred thousand dollars a year and you need the best bd and the best sales, the best marketing. In general, you don't, because if you're solving a real problem, it's like, yeah, how do we get three percent more efficient? Well, if I can demonstrate that I can keep going down that road with a customer and, you know, eventually Build some amazing cutting edge, mind blowing technology. But in general, you just find a lot of the solutions are like they're really don't involve Using technology the way people want it to involve. You know, you, it sounds like you need a more sophisticated search or maybe like a I driven tagging, but I don't think you need an LLM to do this. I think you just need to find the answers in your documentation faster.
Speaker 1:The example of that is, you know, one of the big applications that I've seen emerge from generative and not big, but I've seen it emerges as data querying with generative. What an industrial setting. What could you do with a generative model that you couldn't do with elastic? That's where AI and machine learning finds itself. In a lot of instances, especially in industry, there are possibilities. The possibilities Are very, very real for all of the things that AI and machine learning have been shown on a powerpoint to be able to do. Those things are real, they're hard and it requires real clarity To connect to those things, to problems that need to be solved that will make an impact on an outcome in an enterprise, and I think getting all those pieces together has been a big hurdle for a lot of the man, the industrial base, all to say, the industrial base generally there has been really great.
Speaker 2:I hope this jogs some people's desire to look at different spaces for different problems, including manufacturing, which is close to my heart. We don't often get to talk about it on the show, even though Manufacturing is at the heart of what defense is about. Somebody has to make the thing that we invent right. Somebody has to turn it into reality. So I appreciate you coming on the show and talking about manufacturing and talking about tech, so it's really good.
Speaker 1:Yeah, thank you. Thank you for having me. I enjoyed it. Hopefully we cross pass again sometime. I name is Cali Kean.
Speaker 2:This has been the startup defense.
