Episode Summary
The episode "Operations Meets AI: Data Intelligence in Manufacturing" delves into how generative and agentic AI are transforming manufacturing operations with TD Engine's Jeff Tao and OpsMateAI's Howard Heppelmann as guests. TD Engine specializes in big data and cloud technology to help manufacturers optimize their operations, while OpsMateAI focuses on leveraging generative AI for automation and enhanced decision-making within factories.
Key insights from the conversation highlight the transformative potential of AI in manufacturing. Howard emphasizes that user passion is driving adoption, with employees eager to integrate these tools into their daily work, making their jobs easier and more efficient. Jeff underscores the importance of agility, noting that startups like TD Engine can quickly implement new technologies compared to established PLM companies, which often struggle due to slower internal processes and marketing cycles.
For PLM and engineering professionals, the episode underscores the necessity of embracing AI to stay competitive in an increasingly data-driven manufacturing landscape. The key takeaway is that integrating AI tools can significantly enhance operational efficiency and decision-making, making it crucial for organizations to explore these technologies despite potential challenges.
Full Transcript
Michael Finocchiarowe're live. This is Michael Finocchiaro on the AI Across the Product Lifecycle podcast. I'm joined today by Jeff Tao of TD Engine and Howard Heppelmann of OpsMateAI. It's gonna be a great conversation. Thank you guys for joining me. I'm really happy to see you. Do you guys wanna introduce yourselves real quick before we dive into some of the questions?
Howard HeppelmamnYeah, Jeff, go ahead.
Jeff Tao @ TDengineSure. Okay. My name is Jeff Tao. It's my pleasure to join this webinar. ⁓ I'm the founder and the CEO for... I started this business more than eight years ago. We focus on industry, internet, and also IoT. We try to use big data and ⁓ cloud technology and AI to help manufacturers for their operations. Okay.
Michael FinocchiaroThanks.
Howard HeppelmamnYeah, and ⁓ Michael, same here. Thank you for having us today. And Jeff, nice to be here with you. ⁓ I'm Howard Heppelmann. I'm the CEO and co-founder of Opsmate AI. And in my previous life, I was the general manager of the ThingWorks business at the industrial software company, ⁓ PTC. I really focused on manufacturing and how we could help manufacturers bring their data together to get better insights into how to run their factory. And ⁓ we'll get into the conversation, I'm sure, but with generative and agentic AI, ⁓ I saw clearly that there would be a different way to go after that problem that would create more value for the customers and break down some of the barriers that have held us back over time. And so that's why we started OpsMate AI and thrilled to be here today to share our story with everyone.
Michael FinocchiaroI'm really excited. This is going be really cool. ⁓ And that's an interesting topic, right? The Thingworks ⁓ acquisition by TPG so recently. ⁓ Lots of movement in our space these days, is interesting. ⁓ So ⁓ we're a couple of years into the AI revolution, right? Because GEN.AI and OpenAI kind of burst the gates in ⁓ 2022, 2023. ⁓ I'd to know from you guys' perspective, would you guys really bearish when it came out? Were you immediately jumping on it, saying this is going to be the future, or were you a bit like, this is a bit hype, maybe I should wait, is it really going to pan out? I mean, even now, we've got a bit of a bubble going on, right? You've got all this crazy cash transfers going all over the place, which is making people feel like there's going to be a bubble burst. So we'll talk about that later, but so anyway, like... What was your impression like when it first started two to three years ago? Bullish or skeptical?
Howard HeppelmamnYeah, I'll start there, Jeff, if that's OK. ⁓ You know, as someone who was running a large business, it was obvious to me that generative AI at the time was going to change how people interfaced with data, you know, especially in manufacturing, where any given factory has 15, 20, 30 different systems that frontline teams have to rely on. ⁓ you know, as an early adopter of Copilot, I saw that, you know, there was going to be a different way for people to engage their data. I think the big unknown ⁓ and pleasant surprise really was how AI evolved. know, originally it was a gateway to your data. Today, as we know with the Gentic AI, it's really a tool for automation. You know, the idea that agents can... reason, think, and act in addition to just presenting data to you is a very profound concept and it's one that I believe is going to completely change the industry as we know it today. So yeah, it started with a recognition that AI, ⁓ generative AI in particular was going to have a transformational impact on the end user. But I think as it evolved, what we see today is that it's going to have a transformational impact on everything. in the factory processes, users, automation, et cetera. So really excited, Tom.
Michael FinocchiaroThank you. How about you, Jeff?
Jeff Tao @ TDengineYeah, three years back when the generic AI came out, I'm totally impressed. But I think the manufacturing or the whole industry is still far away from it because the cost is too high. But after almost three years, the cost is less than 10 % than before. So at least one year ago, I think it's time to catch up. You have to apply AI to those traditional industries to help manufacturers to get insights from the data quickly, an efficient way. So I embraced AI more than one year ago by a whole team. I think we need to apply AI to our product.
Michael FinocchiaroYeah, because I suppose that, I mean, even in ThingWorx, you guys had ThingWorx Analytics, you were doing ML. So I mean, some of the concepts of AI have been around for a long time. So as you said, Howard, it's more the fact that you can actually converse with your data that's revolutionary, right?
Howard HeppelmamnYeah, converse with the data and now drive agentic automation. that's the idea that in my old world, ⁓ we strived to provide frontline teams with insights, but that entire layer above those insights was left to humans to interpret and act on. And this new way of thinking about these challenges where humans can be pair it up with their digital teammates, and those digital teammates have the same kind of ability to reason, think, and act, ⁓ really is where the big transformation will ultimately come from. Short-term productivity gains, absolutely. Longer term, it's going to show up in how we fundamentally rethink and reimagine what the workforce looks like when you can have ⁓ digital teammates deployed with the same kind of ⁓ conceptual capacity that humans do.
Jeff Tao @ TDengineYeah.
Michael FinocchiaroYou agree with that, Jeff?
Jeff Tao @ TDengineYeah, so from my side, ⁓ more than like one year ago, I'm thinking about how to apply AI to our time series database. Because our TD engine is originally a time series database. ⁓ Many software ⁓ or business intelligence software try to get insights from our ⁓ time series database. So I'm thinking how to use AI to ⁓ dig the insights from our tensors database automatically. ⁓ Last year we got a very good idea. I can talk about this later.
Michael FinocchiaroOkay. So, another thing that's a bit, it has felt a bit hypey in the beginning and it's a lot less hypey now is ⁓ the use of AI in development and developing the product. ⁓ You guys are both running companies with many developers and engineers. Has this already been part of their workflow as engineers? Are they already using cursor on a daily basis or how is AI? ⁓ impacted the way they develop and has it accelerated the entire process in terms of maybe even automating pieces of the DevOps tool chain and stuff.
Howard HeppelmamnJeff, do you want to take that to start with?
Jeff Tao @ TDengineSure. ⁓ From TD engine team, ⁓ think almost all our software engineers are using AI tools like CURS or Copilot ⁓ in Every day, it boosts the efficiency, especially for unit testing. Because our software is like a database. It's kind of kernel code. ⁓ It's very good for... or you need testing. But it's still hard to generate. Some engineers still use AI to generate some code, but it takes time to debug sometimes. So it's hard to say ⁓ how much boosts our efficiency. But definitely for testing, for review, it's much, much better. It saves us lots and of time for testing and review code.
Michael FinocchiaroMmm.
Jeff Tao @ TDengineand also for our marketing and even sales, it helps us too.
Michael Finocchiarobecause you can do a mock-up really quickly, whereas before you need a little while for an artist or somebody to come in and do it. Sorry, Howard.
Howard HeppelmamnYeah, I think. Yeah, I think Michael, this is an area where, ⁓ you know, it's not hype. The reality is lived up to the hype for sure. ⁓ It doesn't just change what you build it, at least for us, completely changes how we build it. And if you think about, at least for me, the old world was we would have product managers who would spend a lot of time documenting usually in word format, you know, what the requirements were for the product. And then you'd go off and using kind of older tools in a slower process, you'd try to build some mock-ups, etc. I mean, today, you know, we've really flipped that process where we start with a prototype of what we want the actual system to be, and that's what we show to the customer. And, you you can generate a true POC in hours, and then that becomes the pivot point around how you think about the feedback you're collecting and what you want to go build. And then when it comes time to building it, whether it be automated testing where the testing function has really switched from people writing code to now people sort of overseeing the code that's written. ⁓ That's all just absolutely amazing. ⁓ A story I can share is my first realization of this, and it goes back already six, seven months ago in our life cycle. One of our ⁓ developers, we had a review one morning and we were working on our multi-tenancy system. And we got on a call and he was so excited because he said, hey, I wrote 6,000 lines of code last night to finish this project. And of course he didn't write 6,000 lines of code. He had Cursor and these other tools write most of that code and then reviewed the code and tested the code, et cetera. But the capability that AI is providing in terms of the acceleration of the product itself is simply unbelievable. It's amazing.
Michael FinocchiaroIs this a capability you're giving to all the developers? I ask that because I talked to someone last week who said that ⁓ he was not giving those tools to the junior developers because he wanted them to still have development skills and rather than delegating everything to the AI, he wanted it to be more for the senior developers. ⁓ I don't know. It was an interesting perspective. I hadn't thought of it that way. As management, as the CEOs of these companies, are you just like, open bar, use whatever you like as long as you get your objectives done? Or is there a bit more thinking in terms of growing the person as a developer, as an engineer gradually? Just a question.
Jeff Tao @ TDengineIn my point, I think for fresh graduates or junior engineers, better to use AI because AI is much better than their work. No matter the quality of the code and even the style of the code, even the design, it's much better than junior level engineers. But of course, they lost the chance to learn some details, but that's fine.
Michael FinocchiaroHa!
Jeff Tao @ TDengineThey don't have to ⁓ waste their time on something, you know? Just like 30 years back, I even read assembly code. Okay? But now I don't need to work on assembly code at all. Okay? If you know how to write assembly code, you know some skills, of course. But it doesn't matter. Because now I can use advanced language to write my code. Much, much faster, right? So...
Michael FinocchiaroMm-hmm Yeah. Hmm.
Jeff Tao @ TDengine⁓ I think for junior level engineers, they pick up the AI tools from first day. ⁓
Michael FinocchiaroOkay.
Howard HeppelmamnYeah, I was just going to say that ⁓ first of all, you we're still a relatively new company and relatively small, so we don't have a lot of junior engineers. Our team is pretty experienced. And but I think I tend to agree with Jeff here. You know, at the end of the day, in any business scenario, you have an objective you want to accomplish and you want to accomplish it in the fastest and most effective way. I think these tools are going to become the tools of the future.
Michael FinocchiaroOkay.
Howard HeppelmamnSo, you I think it's more important that people understand how to use the tools they have in front of them to be the most effective they can be, however that shakes out. But at the end of the day, you measure it by, you know, productivity, output, quality, et cetera. And if you can get a better result and get it faster, then, you know, that's what we're going to encourage our team to do.
Michael FinocchiaroIt makes me want to ask also, like, you you guys have tons of experience and there's a lot of anxiety, I think, in the younger engineers that people are even still in university ⁓ about AI taking their jobs. So do you have any perspective or any advice to those people on what to focus on in order to stay relevant in this AI-powered workplace?
Howard HeppelmamnLearn AI. that's that's that's my you know, I guess I'm a believer in a phrase. I didn't come up with it. But many people say, you know, the the jobs that will be at risk are those that are held by people who don't know AI. And, you know, these tools are amazing. You just have to if you're going to be part of the economy, whatever market you're in going forward. And it's an economy that
Michael FinocchiaroHa ha ha ha!
Howard Heppelmamnis affected by AI, take out restaurants, whatever, probably been there as well, you need to know AI. So I would just say I have two young kids and I wake up every day telling them you need to get familiar with this because it will be the determining factor for how you engage at least in the business to business workplace of the future.
Jeff Tao @ TDengineFrom my point, because I'm a software engineer too, and also I have so many friends, they are software engineers. Some of them have ⁓ anxiety, you know, they worry about losing their jobs. But I think ⁓ the whole world requires more and more software engineers, okay? Because AI loads the barrier for software development.
Michael FinocchiaroYou think so too, Jeff?
Jeff Tao @ TDengineMaybe many more people can become programmers or software engineers, you know? And also the system becomes more and more complicated. Just think about it. When I joined the industry 30 years ago, I developed the software for mobile phones. Our team has over 10 software engineers. Our software is only 128 kilobytes the size. But it requires more than 10 engineers. But now, think about the ⁓ application on the mobile. It's always over maybe 20 megabytes or 30 megabytes, even 100 megabytes. The system becomes more complicated, because much, much bigger than the older phones, right? Now, with all the AI, how can you build so much complicated system, right? So don't need to worry about losing your job. You may get more opportunities. ⁓ I think the basic way is just pick up AI. I think just like how would you say it? Yeah.
Michael FinocchiaroSo you guys both have pretty awesome products that are exploiting AI in different levels. And I wanted to understand where it was implemented. Is it something that the user sees in the UI with the chatbot? Is it something that's underpinning in terms of how you're doing your analytics to bring those insights or do those dashboards in the case of TD Engine? Where are the points of contact with the user and AI and between your product and AI?
Howard HeppelmamnJeff, feel free to start there.
Jeff Tao @ TDengineOkay, I can start. ⁓ For our product TD Engine, we just released a new product. It's called IDMMP, Industry Data Managing Platform. It's based on our time series database. For this new product, you really can enjoy its benefit from AI because based on the connectivity data, our new product IDMMP can generate the panel's visualization charts. reports automatically for you. Now, ⁓ chat ⁓ GDP or even Germany are so powerful, but you still need to how to ask questions. In the AI age, the big challenge for people is how to ask questions. Most people don't know how to ask questions. Now, with our tool, they can generate the visualization charts or reports automatically for you. You don't need to the only thing. It's just like a TikTok. You watch the video. If you like it, you just watch it. If you don't like it, just skip it. So life becomes much easier. ⁓ So we solved this problem. Of course, if you know the business, you know how to ask questions. You can say, please generate a report for me for the last days ⁓ power consumption. or some calculation, they can do it. But you got to have domain knowledge, right? So with our product, we give a name, it's called Zero Query Intelligence. ⁓ so it's a paradigm shift for the data consumption. ⁓ You want to get insight, you need to query the data, right? Now, the insights is pushed to you. by our data platform. So that's a big, big change by AI. It lowers the barrier. I think many, many data analysts got to worry about the jobs. So this is all new for us.
Michael FinocchiaroYou're already
Howard HeppelmamnYeah.
Michael Finocchiaroseeing a bit of problems at Accenture and McKinsey, right? You're already seeing bit of the foundations being shook by this stuff, indeed. ⁓ So basically, AI is redoing the user interface and bringing up the insights from the time-scissor database in an intuitive way. And then the user can kind of use a tender interface of, no, I don't like that one. I like that one kind of thing, right? That's cool.
Jeff Tao @ TDengineYes
Michael FinocchiaroHow about with Ops Mate?
Howard HeppelmamnYeah, maybe let me let me go back to sort of the foundation again of what we're doing. First of all, so, you know, in the old world, I ran a business where we were using low code technologies to connect IT and OT data sources together to give people real time visibility to their factory operations. ⁓ In the new world, I like to say what we do is we give manufacturers the opportunity to take everything they have and know about their operations.
Michael FinocchiaroYep.
Howard Heppelmamnknowledge, data, and both IT and OT data, and automatically ingest and contextualize that into the AppSummate platform. So now what does that mean in terms of how it shows up for users? The first thing is we're building this platform not for specialized developers ⁓ and AI experts or coders. We're really building it for the frontline team. So an individual, much like they use Excel today, can collect their own data information and knowledge and bring it into the system and build their own agents on top of that and build their own assistance. So it's really number one about democratizing these capabilities in the hands of all of the frontline teams, not just those that the IT capacity can ⁓ conserve. And then I think the second big thing is, you know, with the generative AI capabilities, the vision we have in the future, is that somebody walks into their factory and they're having a conversation with their factory. That could be through a chat interface, that could be in the future through ⁓ literally a verbal conversation back and forth ⁓ with the data. And we like to think of that as then embedding AI in the flow of work. Everywhere where somebody needs to get their work done and they need information to get it, they're doing that through this sort of user-centric but AI-powered, generative AI-powered ⁓ user interfaces. But then there's a whole piece of AI that they don't see that lives below the surface, where the focus is really on ensuring the accuracy of the information they're getting in the context of the specific work they're doing. And you know, there it's all about AI capabilities, we automatically generate and build a knowledge graph, we have ⁓ tools that allow AI to capture human knowledge in the flow of work. ⁓ We have ⁓ agent scaffolding built in at many different levels to ensure the accuracy of the data that the user is interfacing with, going all the way from deterministic when need be to ensure ⁓ accuracy in a highly sensitive use case to more flexible ⁓ probabilistic capabilities when you want the agent actually to be part of that exploratory and decision-making. ⁓ investigation. ⁓ just to say there's really two levels here. mean, AI for us is everywhere. It's both above the surface in everything that the user does to interface with their data in their factory. And then it lives almost at every process below that to ensure the accuracy ⁓ and what you would expect to be the requirements that need to meet the high stakes demands of a manufacturing environment.
Jeff Tao @ TDengineyou you
Michael FinocchiaroThat's interesting. Thanks, Howard. Some of these calls... Go ahead.
Jeff Tao @ TDengineYeah, Michael, I can add something more. TD Engine IDMP ⁓ can work with OpsMate in a perfect way. If a customer combines two products together, a magic happens, Some magic will happen. Let me explain. Maybe I take one minute to explain. ⁓ In our TDM engine IDM pin, we just released a new feature. ⁓ We built our own 10-Six foundation model. It can detect anomaly automatically. You ⁓ can use our tool, say, if anything wrong with my pump, with my boiler, please raise alarm. You don't need to the rules at all.
Michael Finocchiarosome audio issues. There he is.
Jeff Tao @ TDengineYou can just use the literal language to say, only machine wrong with my device, please, raise or not. Then, OBSMET captures this signal. Okay, let me find out what's the root cause. They'll give you solutions, you know? So we only tell the customer something wrong, you know? Then OBSMET, how about this software can say, let me analyze that.
Michael FinocchiaroMm-hmm.
Jeff Tao @ TDengineTell you why it happens.
Howard HeppelmamnYeah, think what ⁓ Jeff and I have had some interesting conversations about this. But if you think of a typical factory environment today, there's basically four questions that come up every single day. What's happening? Why is it happening? What do we need to do to fix it? And how do we prevent it in the future? And what Jeff just described there is when you connect the foundational data platform like TD Engine that's able to identify that first level issue of what's happening and be proactive about it. And then you marry that up with agents that have the capability and capacity to reason, think and act. You're able to automate a lot of that. ⁓ I should say automate away a lot of the lost production time today that sits at that middle layer of, you know, humans needing to number one, find the issue and then number two, figure out.
Michael FinocchiaroRight.
Howard Heppelmamnwhat are we actually seeing here and then take all the actions to fix it. So yeah, I love what Jeff was just saying there. You can see a future where a lot of this is a closed loop, ⁓ agentic automation capability. Now, humans need to be the ones that train them and guide them and,
Jeff Tao @ TDengineYes, it's a closed loop. Closed loop, yes.
Michael FinocchiaroI ⁓ Yeah, that's cool. That ties in well to the articles I was writing last week about the industrial metaverse, because all that is about putting the design data into simulation in a closed loop, also bringing in the operational data in order to improve the original designs to make sure that the errors are caught and corrected almost in real time. Actually, I have two other questions in the same vein. I'll do a short one real quick. Are you guys focused on Are you doing both process industries and discrete or are you focused only on discrete manufacturing?
Howard HeppelmamnYeah, Jeff, I'll hit that real quick. For us at OpsMate, we're really focused on discrete and batch industries. yeah, the reason for that, what I should say is the right answer is we're focused on industries that tend to align with the ISA 95 or the Purdue model, because that whole concept of the plant hierarchy, the plant model is very germane to how we build out our solutions. know, at the end of the day,
Michael FinocchiaroUnbatch, okay, not continuous, gotcha, okay. Right.
Howard HeppelmamnIt's very, very difficult in this AI world if you took an area of a factory that has 35 similar machines and one of them's having a quality problem to figure out how to empower a technician who's troubleshooting that problem with precisely the data about that one machine, not the 30 that came before or the five that came after it, but that one machine. And so to do that, there's many, techniques that we leverage to engineer in that accuracy, but the plant model, the plant context is key to that. And so, you know, we tend to stay true to that sort of ISO 95 ⁓ construct for how we think about creating our solution.
Michael FinocchiaroOkay. And you, Jeff, are you us the same or?
Jeff Tao @ TDengine⁓ For us, our technology can apply to any industry, ⁓ this great edge will continue both works.
Michael Finocchiaroto all of them. Okay. And so ⁓ what I found interesting when I was talking to ⁓ Karan Ataladi of First Residence is I was asking about, you know, saying, well, it's what must be difficult is as engineers, we want everything to be deterministic. And in the engineering world is very deterministic. And we're throwing these probabilistic LLMs at these solutions. And it's a bit disconcerting, right? Because you think And yet, what Karan said, I thought was pretty insightful. said, well, the thing is manufacturing is also ⁓ probabilistic. It's possible that the operator comes in sick. It's possible he throws a switch wrong. It's possible he trips and breaks his leg. And then we've got an, there is some probable, you know, probabilistic elements to it. But I'm just wondering, you guys have any perspective on that? And like how we're taking this probabilistic LLM stuff and applying it to these very, very deterministic. know, processes and ⁓ so forth.
Howard HeppelmamnYeah, maybe I'll start there. Just two ideas. One you kind of alluded to already, but I was talking to a CEO of a company and he said, hey, ⁓ we make mistakes all the time every day. All I need to do is anything that's better than our current performance, meaning better than the current mistakes we make is value add to the business. And if it means we can move faster in all of that, that's great. So I think Logically, people want to go to it has to be perfect. The reality is it isn't perfect today. It only has to be better than where you're at today. Generally, I would say that's the sort of first level one. Now, at the end of the day, there are certain use cases that we deal with where you do have to be perfect. And that's why really the blend here isn't all AI. It's a blend of when you need deterministic approaches, you need to actually write code and rely on the code, then the agent's job is to know that it needs to default to a 100 % deterministic approach. In other areas where you may want it to look across many documents and say, hey, what could be the problem? And maybe there's three answers. Okay, it's okay to have a little bit of leeway there in terms of what the agent comes back with and recommends. But we're designing a system that you can make it 100 % deterministic, or you could allow it to be 100 % probabilistic. And it's really the use cases that need to dictate based upon what tolerance you have in those answers. Safety, zero tolerance. Setting up a line with an expensive material run, zero tolerance. what might be the root cause of this. Okay, I can live with three answers coming back that I have to apply my own human intellect to.
Michael FinocchiaroOkay, how about you, Geoff?
Jeff Tao @ TDengineYeah, for our side story, a little bit different. It's a combination of deterministic and probabilistic. Once we show the visualization charts or reports, the data is from our database. Definitely it's 100%. Correct. It's deterministic. But what kind of visualization or chart or what kind of report is probabilistic? Maybe it's not right. The KPI generated by large language model, maybe not right at all. Or maybe the formula to calculate the KPI not right, maybe. It's probabilistic. But once we apply the formula to calculate the KPIs to our real data, definitely the data is 100 % right. Because it's from database itself. So, use a combination.
Michael FinocchiaroOkay, I like those answers. Thank you, was cool. So then now we're three years into the GEN.AI revolution. We're heading towards the end of 2025 and 2026. We're likely to see even more changes, perhaps a bit of deflation of the bubble, you know? And I was wondering like for you guys, how is...
Jeff Tao @ TDengineYeah, okay.
Michael FinocchiaroHow has your opinion around AI changed? How has your view on its utility, its limits, its possibilities that haven't been explored yet? How do you feel about AI today compared to three years ago? What has changed in your perspective?
Jeff Tao @ TDengineGo gate, hold up.
Howard HeppelmamnOkay, yeah, I think I I alluded to this at the beginning in my opening statement, but to me, it's really around the automation capabilities that AI provides. ⁓ When you think about AI agents that can reason, think and act, again, very similar ⁓ to how humans do, you can start to reimagine the entire way you're conducting and allocating your workforce and things like that. And we're already working with some very advanced customers who would, know, who are researching all of this. just met with a customer last week when I was at the Microsoft Ignite event. And that customer said that they've already done an analysis and they believe that 80 % of the decisions that today are made in their manufacturing organization can be done in the future. by agents. And when you go into it with that kind of analysis and thinking, what is abundantly clear is that this is not just a productivity tool alone. It's really a way to reimagine how work is done, the speed, the agility, the flexibility that all of that adds to a manufacturing company, which today has been, for no fault of any worker in the factory,
Michael Finocchiaro80 percent.
Howard Heppelmamn⁓ Today, just way too many of those processes are human-centric, even when humans are not the best people to be performing them. Just think reading and contextualizing and intensifying data is not a strength of humans. Maybe assessing the results of that might be, but ⁓ we can't process 2,000 pages of information in a second and then come up with a... a synthesized response to that. so anyway, ⁓ I think that's the biggest learning again is that it's really the big, big, big unlock, at least for what we're building, will come in the form of the agentic capabilities that sit on top of the generative capabilities. But it's really about how you deploy digital teammates, a digital workforce that can reason, think and act alongside their human counterparts, both to lift Their human counterparts give them the tools they need to be the absolute best they can be, but also to offload non-value added human workloads to agents who simply, in some cases, are far, far better at performing them ⁓ than humans are.
Jeff Tao @ TDengineOkay, from my side, I think after three years, ⁓ one thing I can tell very clearly, if you know how to use AI technology, you can become a superman, okay? ⁓ Because you can have how many digital workforce or digital teammates you can manage. Then it means how much...
Michael FinocchiaroOkay.
Jeff Tao @ TDengineproductivity you can deliver. So if you can handle 100 digital teammates, perfect. You can become one man person company. You can even with one single man, you can generate 10 million revenue or 100 million revenue is possible right now. And another thing in my mind after three years, I think in the AI age, there are two how to differentiate ourselves from other people. The key traits, two traits. First one, how to ask questions, because it's very, important. You got to know how to ask questions, okay? You got to know how to raise very good questions, right? The second thing is you need to tell the solution is good or bad. the solution provided by AI, you need to have a very good taste to tell if it's the right one or the one. If you know how to raise questions, if you know how to tell good or bad, you can win. Otherwise, probably you lost your job for others.
Michael FinocchiaroWell, that's where the basics and being an engineer or a software engineer are so critical and being able to determine whether the response is valid or invalid, right? That's great. I also liked Ayoah Berry's framework that he used at PTC, the advise, assist, automate. And I see it as like the assist was the very first phase of co-pilots just trying to give you more information.
Jeff Tao @ TDengineYes.
Michael Finocchiaroand advise is sort of like, okay, here's one, like basically what you were saying, Howard, here's three possibilities, which one do you want to do? And then an automate is, you here's a workflow. And I think I agree with you guys. think that we're in 2026, 2027, we're moving into that automate phase where we're going to give more and more pieces of a workflow to AI. One thing I had thought about at a workshop and a conference this year was that Maybe we need some my amenities agents, some agents are like, I don't know. No, I'm not really sure. You're not the most beautiful, brilliant guy on the planet. I'm not going to give you 17 other things I want to do. I'm just going to say, I don't believe it. You know, the response to that other agent. No, no, no, I don't believe you because you know, at the moment, the agents are just tuned to tell you how great you are. They need to be tuned to be a little bit more skeptical, I think, you know. So one last question in terms of AI before we move to your customers. Do you guys think that as we go forward, you're going to be training like your own ⁓ SLMs or LLMs or LLMs ⁓ on operational data in order to solve these problems deterministically? Or do you think you're going to continue to rely on the off the shelf ⁓ LLMs from Anthropic and everybody? And then you're just going to use RAG to augment the context in order to answer. Where do you see that going? different answers from different parts of the industry.
Jeff Tao @ TDengineyou
Howard HeppelmamnYeah, I would say when we started, we thought we were going to have to. I think for a couple reasons, until you really have to, that's a dangerous ground to walk on. So far, we haven't found that we need to because these underlying models get so good. But I would also say, there's not one model. We sort of use this idea of a model garden. There's always... the best model for the best task when you want to balance cost performance and, you know, accuracy, et cetera, et cetera. So, ⁓ you know, I don't want to get into a scenario where we're training ⁓ models, particularly that, ⁓ you know, might compete with the next version of ⁓ Google Gemini or chat GPT. And so far, what we see is those models are improving at such a rate that ⁓
Jeff Tao @ TDengineThanks. you
Howard Heppelmamnyou know, just incorporating those capabilities with all the agent scaffolding, not, you know, not just reg, but the deterministic guard rails that exist. All the ⁓ other capabilities that we build into the platform is really where we can wrap them and then deliver that model. We reserve the right to change that, you know, ⁓ if it looks like there's a need in the future. But I would say if anything, when we started, we thought we would have to.
Michael FinocchiaroExactly, Of course.
Howard HeppelmamnAnd today, we're wondering if that will actually ⁓ be true down the road just based upon the pace and progression of the underlying mega models that exist.
Michael FinocchiaroHow about you, Jeff?
Jeff Tao @ TDengineOkay, definitely I don't want to train my own larger language model. but for operational data, we do need to train our time-sets foundation model because the time-sets data is still a little bit different from natural language. And also for each sub-line, something different. We already built our own time-sets foundation model.
Michael FinocchiaroMm. Yeah.
Jeff Tao @ TDengineBut for different industries or different customers, maybe we still need to tune the model for the JSC-9000. That's possible. But for the domain knowledge, definitely, I don't want to train at all. I think OpenAI or Jiminy are doing very, very good job. ⁓ For Kansas Foundation model, it's much, much smaller. The cost is much lower.
Michael FinocchiaroYeah. To your point, Howard, I think that there's a lack of maturity in the DevOps chain terms of LLM ops, right? Because that's gonna be more more challenging, as you're saying. It's just accelerating. And if you start doing your own thing, it's gonna be obsolete because Jimini 5.1 just came out. So chasing a moving target, right? So let's...
Howard HeppelmamnYeah.
Jeff Tao @ TDengineYeah, yeah, yeah, yeah, yeah.
Michael FinocchiaroJust switch wheels. We've got another 15 minutes. I wanted to talk a little bit about when you guys go into the real world to customers. So the first part of the question is when you go in, ⁓ when you, and you don't name any names, obviously, but overall, when you're going to your industrial customers, where are they today in terms of digital maturity? I think of digital maturity on a spectrum of one to five. One is like, we're still on email and doing Excel. Five is like fully agentic digital twins with adaptive AI and, know, autonomous robots and everything else. ⁓ I'm suspecting that most companies are between one and two, maybe between two and three. Is that the same that you found?
Howard HeppelmamnJeff, do you wanna start your?
Jeff Tao @ TDengineOkay, we actually our TD engine already has a huge has a big base of customers. First of we have like almost 1 million free customers, but I don't know how they work. But we do have over 500 paid customers. So I visit the customers almost every week for the digital maturity. Some customers are really still in very old days. They still count on like a spreadsheet, okay? Even just email. But some of our customers are already in very advanced stage. They use AI tools. They use all kind of modern data platforms like ours every day. So maybe like 10 % can be say it's a...
Michael Finocchiarobetween two and three.
Jeff Tao @ TDengineYeah, it's 10 % is like a five. most of the things, in 22, yeah, yeah, there is a big, big gap. Okay. And also some customers already connect lots and of data and clean the data and the data quality are pretty good too. For example, one of our customer
Michael FinocchiaroFive. Okay, so there's a big gap between them.
Jeff Tao @ TDengineconnect the data ⁓ like 5 terabytes a day or 5 terabytes a single day. Okay, that's a huge amount of data. And also they keep the data very clean. So they can use AI tools to get insights. ⁓ Okay, very, very fun. I mean, in a very good way, like our tools can help them a lot. But for some customers, even they have data, the data quality.
Michael FinocchiaroHmm.
Jeff Tao @ TDengineare very very bad. AI has no way to help them, even we provide them very good tools. They have to clean their data, they have to improve their data quality first. So there is a long way for them to go. Even you have a very powerful tool. from my point of view, many many many factories, ⁓ many many plants, ⁓ maybe two or just one.
Michael FinocchiaroYeah. All right. Okay.
Jeff Tao @ TDengineOkay.
Michael FinocchiaroHoward, is that your experience too?
Howard HeppelmamnYeah, I mean, I would say that I'm probably even a little more, I don't want to say pessimistic, but maybe if I just think of the world that I came from in the past and, you know, there are all these tools out there that supposedly are there to give people real time insights to what is happening. You know, if you go to a company that has 100 factories,
Michael FinocchiaroMm-hmm.
Howard HeppelmamnThey might tell you about one or two that have those capabilities deployed, ⁓ but they're not everywhere. It's not pervasive. By far, the most popular tool in the factory today is Microsoft Excel. By far, that's just the reality when it comes to problem solving. People are in Microsoft Excel. Now, I think that ⁓ if you look at just
Michael Finocchiarois Excel.
Howard Heppelmamntools like ChatGPT and generative AI in general, these tools for the first, this is like one of the first technologies maybe other than mobile, where users are driving the adoption, not companies and governments, right? So I'm sure all of us, the four of us here use ChatGPT probably daily, just in the sort of year plus that I've been at this, when we first started talking to customers, you know, last, ⁓
Michael FinocchiaroMm. Right.
Howard Heppelmamnlast summer and fall in 2024, we would always ask them the question just as a bellwether to where we were in the conversation, do you use tools like ChatGPT? And the answer originally was maybe 30, 40 % would say they do. When we asked that question today, the users are actually saying, yeah, I use it all the time. By the way, don't tell anybody. I'm even using it sometimes in the factory to... take a picture of something and ask chat GPT or, you and by the way, that should make the IT organizations very, very nervous because these are generally the free versions of these tools that people are running around the factory snapping pictures with their phones, et cetera. And you got to wonder where that data is going. So anyway, I only say that to say that in this particular case, the desire, you know, nobody wants to show up
Michael FinocchiaroYeah ⁓ yeah.
Howard HeppelmamnI don't want to leave my version of chat GPT behind when I walk out of my office here. I would want to take it with me into my new ⁓ job role, whatever. And I think companies have this, you know, users have this ⁓ desire. And so I think this is going to move at a much faster pace. Again, the model we have in our mind is not that we're replacing Excel, but that we're building a tool set for users. that's like Excel. They can use this to collect their own data, solve their own problems. I think just in general on the question of maturity, one of the first questions, Michael, you can imagine people ask is, well, do I have to have perfect data to get started? And ⁓ that's always sort of been the idea behind AI is that your data has to be perfect. ⁓ What we like to say is, what are your people use today?
Michael Finocchiaroyou
Howard HeppelmamnIf you have people troubleshooting quality issues on the factory floor today or downtime issues, that data isn't perfect. Give them better access to the data that they have and then allow the system, the AI system, the closed loop process to cleanse that data over time. Our approach is very simple. You start with what's your bottleneck process that's causing you the most issue? What's the corpus of the data information and knowledge that surrounds that? Put that into a system. give people access to the same data they're using today, but in a ⁓ more instantaneous and reliable way, and then allow feedback from the factory floor to flag where there's either gaps in data or data needs to be improved, et cetera. It's more of this sort of bottoms up, use case driven, value driven ⁓ approach to how people get started. So I would summarize all of that just to say that I think the appetite is really strong for solutions here. And ⁓ the idea that users can pick up AI tools like OpsMate that empowered them to solve their problems, I think, breaks through a lot of the barriers that have held back ⁓ adoption historically, least certainly based upon my experience.
Michael FinocchiaroActually, that's ⁓ great because one of the comments in the live chat, ⁓ Jim fan and saying, what are some of the reasons or barriers that manufacturing space is slower? That it's making the manufacturing space slower and adopting AI. And I think we've addressed some of those. It's just legacy data, know, manual processes, safety, regulatory compliance and things like that. I like the there's another question from Devendra Aravinda. ⁓ And I like this one. This may be a good one to. Before we close, do you see any patterns and the kinds of problems that manufacturers are looking to solve using ⁓ data analytics and AI today? Which sort of speaks to the use cases that ⁓ ops made and TD engine are implementing today. What are some of the ones that are just resonating the most with your customers?
Howard HeppelmamnYeah, I'll start there. When we formed OBSME a year ago last May, so in the spring of 2024, one of the first things we did was put together an advisory board of blue chip manufacturing companies. And we basically sat down with them over several months and asked them, ⁓ given what's possible with these technologies both today and likely in the near future, what are the use cases that you think would bring the most value around what's feasible and ⁓ valuable. And one of the use cases that we settled on, which is by far the number one use case our customers ⁓ look tops make today to help solve, is around troubleshooting loss production time. Again, it goes back to this very basic concept of, know, factories are complex, they're all different. lots of different systems every day, you know, things don't go as planned on paper. And so the questions are what's happening, why is it happening? What should we do to fix it? And how can we prevent it? And so that's really the area that we focused on because there's there's certainly a productivity value to those people. They, know, typically engineers, technicians in a factory spend more than 20 percent of their time ⁓ just simply searching for information and and then maybe that much again performing administrative non-value added tasks. And so if you can give them sort of one second access to the data information and knowledge that they need to make better and faster decisions, that has value. But the real value, as we can all appreciate, is if you're, you know, depending upon what industry, of course, but from food and beverage to, you know, heavy industrial or automotive, ⁓ know, downtime can cost anywhere from probably on the low end, $10,000 an hour up to hundreds of thousands of dollars an hour. So if we focus on ⁓ that particular ⁓ use case, either downtime related to equipment failures or downtime related to quality events, and really look at how can we help people accelerate the meantime to resolution, that carries a outsized ⁓ financial benefit to those organizations.
Michael FinocchiaroHmm. Nice. Have you seen different use cases, Jeff?
Jeff Tao @ TDengineFor my use cases, I found many small and medium size manufacturers, ⁓ they don't have full time data analyst. So ⁓ that's why think AI can help them. That's why our new product IDMP can help them. Because our data agent or AI agent is just like their data analyst. ⁓
Michael FinocchiaroOkay
Jeff Tao @ TDengineThere are a few million small-sized or two-media-sized manufacturers in the world. They cannot afford full-time data analysts at all. So AI really can help them to get insights from their SCADA system. And also I found that most of the manufacturers already have automations, but there are a lot of digital ready or digital maturity is not there.
Michael FinocchiaroMm.
Jeff Tao @ TDengineThey even don't have the ⁓ data center, you know. They only have SCADA or SCADA or DCS system for automation. But now we need help them to get the data out from the SCADA or DCS system and use AI to provide insights for them. I think that's big opportunity for a company like us.
Michael FinocchiaroRight. So when you guys have put in OpsMain and TD Engine, has it had a ripple effect on the entire organization? Has there been a bit of a aha moment by management saying, oh my God, look what we can do if we actually had some initiatives in place to improve our data maturity and break some of the silos between the different departments? Has there been a ripple effect or maybe it's just happened in a bubble and you know? Just wondering what kind of impact it is when you put in a kick-ass AI software.
Jeff Tao @ TDengine⁓ ⁓ I can give you ⁓ a very good example. Just last month, we signed a very good contract with a coffee company for the roasting factory. They don't have a data platform at all. They only have a SCADA system, automation system. So we only spent one single day to set up everything for
Michael FinocchiaroOkay.
Jeff Tao @ TDengineNow they have visualization charts, reports, just in a few days. ⁓ In the old days, it takes a very, very long time, a few months to build this whole system. Now it only takes them, I think, less than one week. ⁓ Even they don't have domain knowledge, because AI generates all the reports and the visualization charts for them automatically. They don't have data analysis at all. They only have automation engineer. So I think that they benefit from AI really.
Michael FinocchiaroIt's a powerful story. How about you, Howard?
Howard HeppelmamnYeah. I think, you know, the way I would describe the ripple effect today is my entire career, you know, I've spent building software for either engineering or manufacturing functions within manufacturing companies. And ⁓ it's always been relatively easy to convince management of the value. You know, what would it mean to you if you could have real time visibility to what's happening in your factory versus not having that? You can get the leadership excited. AI obviously is even that on steroids because generally the CEO of the company today is poking around saying, what are you doing with AI? What are you doing? I want to see a plan. I want to see some action. The thing that for me is radically different is when you get to the end users. The end users of, I'll admit, every system I've ever built before.
Michael FinocchiaroAll
Howard Heppelmamnyou know, when management would put it in front of them, they'd kind of say, well, what's in it for me? Right. I mean, I understand why this might help you, but how does it help me do my job today? AI is completely different in that sense. Again, I mean, we're all adopting it in our own lives. It makes, it makes our own lives, personal lives better. It makes our business lives better. And so when we do a proof of concept, which usually is something we can do, you know, just within a day, much like what Jeff was describing. and.
Jeff Tao @ TDengineThank
Michael FinocchiaroYeah.
Howard HeppelmamnPeople can see that they can bring, know, sort of a, I hate to say chat GPT, cause it's way more than chat GPT, but they can bring that concept of take everything I have and know about my department, about my factory, about a specific bottleneck process, load it into your system in hours and then be able to query it. And in one second, get the answers back to what's happening, why is it happening and what should I do about it? makes their lives and what they're asked to do so much easier. And so this is the first time where I think the user passion is outstripping the management passion and the management passion is still very high. ⁓ I think the big challenge for industrial companies is just, you they're slow. I mean, it's no offense to any of them. I love working with them. I always have. but the clock speed of an industrial company is about the slowest business clock speed you can find. And so everything from, know, how do we validate to, you know, how do we just mobilize to whatever that can be the real hold back today. think that some of the companies we're working with understand there's an urgency around this technology because it does have the opportunity to reshape the competitive landscape. And others, as they're just learning about it, tend to be very conservative in how they do that. ⁓ But I think the big difference is just the user poll. ⁓ Users want these tools, they have them in their personal life, they don't want to check them at the factory door, and it makes their lives a lot easier when they have them in their back pocket.
Michael Finocchiaroespecially when you use TikTok at work with TD Engine, right? Well, it's been a really fantastic time. I really appreciate you guys taking your time today. It was a lot of fun. ⁓ And ⁓ I'll be back next week after Thanksgiving. I've got a couple more of these with some other exciting startups. I think also with respect to what you just said, Howard, the industrial companies, I'm thinking also of...
Howard HeppelmamnYeah.
Jeff Tao @ TDengineYeah, ⁓
Michael Finocchiarothe big three PLM companies that ⁓ struggle to keep up because you guys are far more agile. So you can just take something with the eye and run with it. They've got to put it in their process and redo the marketing. And so that opens up a huge opportunity. And I've found so far 380 startup companies like you guys just doing this amazing stuff is absolutely stupendous. think my latest kind is 19 billion site and market in terms of market cap for these companies. It's just incredible. So I, I salute you guys, fantastic job. It's really awesome to see what you're doing. Always very inspiring. I hope I can get you guys on again sometime, maybe six, eight months, and see how things are going. So with that, I just wanted to say thank you and ⁓ wrap up. Do you guys want to say goodbye real quick? Howard Heppelmamn (1:00:43) Yeah, Michael, thanks for having us and Jeff, pleasure to be here with you today. And for your audience, you know, excited to see where companies can take it. And we're obviously excited and willing to engage anybody who wants to learn more about OpsMain AI. So thanks a lot. Michael Finocchiaro (1:00:59) Awesome. Jeff Tao @ TDengine (1:01:01) Yeah, thanks Michael, thanks Howard. ⁓ The holiday season is coming. Best wishes. Okay. Yeah. Okay. Michael Finocchiaro (1:01:08) Yeah. Happy Turkey Day. Happy Turkey Day, everybody. Okay.