922: AI for Manufacturing and Industry, with Hugo Dozois-Caouette

0:00Jon Krohn:Welcome to episode number 922 of the Super Data Science Podcast. I'm your host, Jon Krohn. Today's guest is Hugo Dozois-Caouette, co-founder and CTO of MaintainX, a startup that has raised over $100 million in venture capital to bring AI to industrial operations in an intuitive way. We haven't had many episodes on industrial applications of AI, despite industry generating oodles and oodles of data, making this a great episode to hear about the enormous amount of opportunity applying data science to this space. All right, enjoy. Here we go. Hugo, welcome to the Super Data Science Podcast. It's a delight to have you here.

0:43Jon Krohn:Where in the world are you calling in from today? Hey, thanks, Sean. Thanks so much for having me. I'm calling from Miami in the south of the US. Excellent. And like me, you're actually a Canadian originally, but you're from the Francophone part of Canada from Quebec. And you studied computer science, computer engineering at the University de Sherbrooke in Quebec, right? Yes, that's correct. Studied computer engineering and then made my way to the U.S. afterward. Yeah, yeah, yeah. A bit of a stint in the Bay Area before founding your own company. That's correct. Was in the Bay for about six years, founded Maintain Ice Fund there.

1:20access to network in the Bay Area is definitely something that is not present anywhere else. And that's why I wanted to get to the Bay Area in the first place. And then like many during the pandemic, made my way east. And then a lot of our company is based on the East Coast. So actually it makes it very convenient. I can make my way to our Montreal office frequently and it aligns all the time zones. So it makes it a lot easier day to day.

1:42Jon Krohn:Fantastic. Yeah, that Bay Area network in tech in general and then in AI specifically, it is unrivaled. There's nothing like it. I'm certainly, every time I visit there, I'm like, man, I really should have considered moving here or should move here at some point. But yeah, I don't know. I'm an East Coaster somehow, end up in New York. But yeah, Miami as well, where you've been drawn to now. Obviously the weather, particularly in the winter, fantastic. And so another place that calls my name sometimes. But let's talk about your company. So you founded MaintainX in 2018 after that Bay Area stint.

2:25Jon Krohn:And you've been CTO there ever since. Your website describes MaintainX as the undisputed leader in CMMS. And obviously that marketing message is targeting a different person than me. Because when I go and I see that on your website, I'm like, what's a CMMS? So Hugo, fill us in. Yeah, for sure. CMMS stands for Computerized Maintenance Management System. At the end of the day, what does that mean? Very simple. It's work execution software for industrial manufacturing. You can think about it's a digitized version of a clipboard. Everybody in the plant going around with clipboards that have procedures, inspection, all kinds of different things that they need to fill.

3:06We provide a digital version to get reports and be able to connect to machines as well so that we can gather more data and bring more insights.

3:12Jon Krohn:So when you say that, is it like it's kind of like an iPad or something like that, like a tablet that people walk around with? Or is that just an analogy? Yeah, that's more of an analogy. So for us, we have a web app that people can use on their computer and we have a mobile app either on a phone or a tablet, depending on what they decide to use. But really, that's where the analogy comes in, where on the app, we have ways to build procedures. We have ways for technicians to go around and fill them. And we want to make it feel like what they're used to, because at the end of the day, change management is probably one of the hardest problems.

3:48So how do we reduce that? And it's making it feel like what they're used to, but just in a digitized format and make it easy for them as they're going around.

3:56Jon Krohn:Gotcha. So in any kind of industrial setting, like say I'm a manufacturer and historically I've had a bunch of employees keeping tabs on how things are going on my manufacturing floor. historically they would have been walking around with a clipboard and just like you know writing down notes manually like in there and they're probably horrible handwriting uh they would have been just putting check marks in boxes because i suspect there's a lot of kind of like process of making sure that kind of everything's going smoothly so you're kind of like here's my checklist and you just go through and you walk around the floor and kind of check the checklist it's probably something like that historically that now you've replaced with an app yeah 100 100%.

4:36Most customers that we see oftentimes came from paper. So they were printing work orders. Either they had them in their own custom work orders in Word, they might have had some solution that was allowed to print, but nothing in the field. And then we've seen customers tell us like, to handle high priority work order, for example, they have a red stack of paper right next to the printer, they change the paper in the printer. And if you see someone going with a red sheet around the plant, that means something is really high priority. And that's the industry we're coming in and trying to really help and make sure that we have ways to capture that data because there's so much valuable data that they're capturing every day.

5:10And unfortunately, it was in file cabinet all along.

5:14Jon Krohn:Right. And so you guys have had a lot of success. You, you know, to be able to claim that you're the undisputed leader in this space and see MMS, you know, that that's tremendous. And so it sounds to me like a big part of your success was creating an application that looks and feels to these workers, to these people on an industrial plant floor to make it feel like a consumer-grade application, to make it feel like they're on LinkedIn or Amazon or any of these other kind of consumer-grade applications that people are used to. Is that right? Yeah, that's correct. You know, growing up, we always heard enterprise-grade, enterprise-grade everywhere.

5:58And And when my co-founder and I came into the space and realized what enterprise-based software looked like, we were like, okay, there needs to be a massive change here. Because all the big social media company, they got to billions of users without any training. Why does it take eight weeks to train a new worker on their maintenance application? in the world that we are right now, where personnel is retiring at an alarmingly fast rate, where there's people that are new to the field, machines are getting more complex. There's no time for us to take eight weeks to form or train a new worker. You need something that people can just pick up and they just understand it.

6:38And so for us, it was mimicking a lot of the consumer patterns, whether it's in the way the app is navigated or the inputs or the way the screen feels. You really want to bring something that feels familiar so that they can pick it up right away.

6:50Jon Krohn:Makes perfect sense. I hadn't thought about it from that perspective of intuition, where with most consumer grade applications, the expectation is that you can just pick it up and use it. And it's just going, you know, you're going to be able to figure out all the bells and whistles kind of in an organic way instead of needing, like you described there, eight weeks of training, which I guess was the standard in CMMS before MaintainX came along. that is definitely an interesting point. And then, so how do you work in things like data and AI? How do you integrate those kinds of capabilities into your MaintainX platform?

7:28Jon Krohn:And maybe to kind of color, to give us some color, maybe you could kind of walk through like a common user story of a user of MaintainX that involves some kind of AI element. Yeah, for sure. I think to preface this, I think just to talk a little bit more about data, when we talk about industrial data, there's different sources of data that come into play. The first and foremost sources of data that's the most common will be just human generated data, like technician walking around a plant, doing repairs, putting data in the system, creating reports. We're still in a world where a lot of the machine have maybe sensors, but they're not connected yet.

8:08And so that's why we invested a lot in making it as easy as possible for people to be able to come in and put data on the platform and making sure that we capture all this information that was in people's head before. We have technicians that have been there for 20 years, 30 years, and they have all this knowledge that has been accumulated. And when new technicians come in, they cannot access that information. So we're really trying to find ways, whether through voice, typing to make it as easy as possible. And then you can get into the next level of maturity, which is all those machines have a ton of sensor on them.

8:44And there's a lot of interesting information from temperature to vibration analysis to current data. There's so much we can do with this. And so it's then capturing all this information and making it accessible because you, from vibration data, you can actually know if a machine is about to break. You can know what type of breakage actually is about to happen. It's not just knowing that it's about to break. You can see if it's a misalignment. You can see if it's a problem with the oil. You can find all kinds of different issues from this information. And so when we're thinking about data, for us, it's like the capture part is probably the most important piece.

9:20I feel like oftentimes people talk about data and they assume a state where all this data is already pre-accessible and on the platform. But when we think about manufacturing, one of the big gaps that we've seen is that we have to make sure that the machines are instrumented in the first place, that users are able to bring this data in. After that, then it's like, what can we do with it? So first part I mentioned, like predictive maintenance is a super interesting use case with machine information. We can detect fault, classify them and get into prescriptive maintenance, which is kind of like the holy grail of maintenance, which is how do you take the fault information, mix that with work order history, with the manuals of the machine, with maybe OEM, the manufacturer's information as well, and be able to tell the user what steps they should be taking or recommend actions of when and what they should be doing to make sure that this machine doesn't go down.

10:17So that's one use case. Second use case that we see in the field as well, at the end of the day, right now, there's a McKinsey study that says that about 60 % of the technician's time is spent on researching, troubleshooting, and writing reports. So for us, when we're thinking about AI, how do we reduce this? The first step is researching. Like right now, most of these machines manual are still on paper. So when something goes wrong with the machine and you see an error code, that's like FD34 on the machine. And you're like, what does that mean? Unless you've been in the field for 20 years and you've seen this code a hundred times, you have no clue.

10:53You have to go back to the library, find a manual, go to page 175, find that error code and try to solve it.

11:01Jon Krohn:When you say library, you mean a physical library of books? Yes. Internally in the company, they'll have a book room with all their manuals just sitting there. That's actually one of the big problems we've seen with customers is how do we get those manual digitized? And we've actually had customers so excited about some of the new A capabilities because one of the things that we offer is allowing people to bring those manuals digitally on the assets and be able to search through them with regenerative AI, building kind of a copilot assistant for all the Zeta. They've actually hired staff, temporary staff over the summer to digitize those manual by hand and get them in the system.

11:41But yeah, this is this is where we we stand in this industry. It's like a lot of this data is still on paper. And so we really have to focus on how do we bring this data in an efficient fashion? Because I think there's so much we can do with it. Wow.

11:53Jon Krohn:All right. I think I might have interrupted your second use case there, but maybe you kind of got to it. Yeah, it was very much aligned with this, which is like, OK, everybody's walking with a supercomputer in their pocket. all day. So how do we make this data accessible to that supercomputer? I think generative AI brings a lot of new capabilities on this front about summarizing searching information that would have been really hard to find before. And so people can now be at the machine directly, search their error code, get some information or ask directly. Maybe they don't know what the name of that piece is.

12:23You know, how often does it happen? You look at a car and you're like, oh, this piece is wrong. No clue how to even search for this. And so I think generative AI brings a a lot of new capabilities to align language and making sure that no matter how you call it, they're able to find it and bring that information to you.

12:39Jon Krohn:Nice. Well, fantastic work, Hugo. It's amazing the technology that you've built over these years. Tell us a bit about the tech stack that you use for building technology like this. Yeah, for sure. Our tech stack is fairly narrow because we want to make sure that our team can work on it in an efficient fashion, but we always invest where we need. So main application tech stack would be mostly TypeScript, Node.js, React, React Native, TypeScript. So a very aligned stack on the product side so that we can move quickly. And then when we think about data, we're powering a lot of our data insights through Databricks.

13:19So we're moving all our data into Databricks warehouse. Python heavily on the AI front, RAG pipelines are still one of the best tool that we have out there to be able to get insights and actually bring information to users. I think this is, sometimes people will say, oh, Ragus is outdated, but to me, it's still one of the best tool out there when you're looking to bring information to people and you live in a world where there's a lot of data. Models have increased context, but it's not an easy solution. You can't just pass the data along the network all the time.

13:54Jon Krohn:Yeah, when you're talking about situations like you were there with, once you digitize that library and you have all of those manuals converted into a digital format, RAG is exactly what I had in my mind, retrieval augmented generation, because that seems like the perfect kind of pipeline for pulling out relevant parts of say a large number of books that you've digitized and then using some kind of chat interface, some kind of LLM to pull out the relevant information from those particular pages from the particular books that you need and be able to answer somebody's question in natural language.

14:32Jon Krohn:I think that's a, it's a no-brainer use case for RAG. Yeah, exactly. And that's where we see kind of the first use case, the first foray, because I think it also, it's about helping the user educate themselves about how to leverage AI. It's new for everyone. And we want to make sure that we make it transparent to them and easy to understand what are the implications of what they get. And then the part we're investing now is around background agents. You know, as mentioned, a lot of the time is spent on researching, but also writing reports. So I feel like agents are a great use case that can work in the background.

15:06Once you finish a work order, that can work for you and bring that report to the finish line and then let you approve it so that it's sent to your manager automatically. A lot of the work also is like, you know, parts inventory management. Going on five websites, picking the best part at the best price is not something that you need a human to do. And I'd much rather the part manager think about is part strategy and think about what's the future and have agents that can work for them and really make sure that all this information is done in the background, captured, and then brought back to them.

15:38And I think this is a case where agents can really come into play and it gets really interesting. we're seeing a world where agents should be kind of colleagues to the technician where all these kind of jobs are where we automate the the non-human work out of the loop right humans should be doing human work and so all of this repetitive work can we have an agent that's able to do it and human is able to orchestrate this is capabilities that's coming off in time I feel like in the tech world and a lot of tech companies and are available but haven't made their way in the manufacturing world yet.

16:12And I think this is where the big opportunity is around the app.

16:16Jon Krohn:Excellent. And then so when you're talking about things like the predictive analytics that you're doing, so you're talking about how critical it is to be collecting data on vibrations, say on some piece of equipment, so that you can predict when some piece of equipment is about to fail instead of after it's failed, which I can imagine for your clients is a huge money and time saver because a piece of equipment actually going down as opposed to being about to go down and fixed, it can make a huge difference in terms of downtime in say a manufacturing setting. So I can see how valuable that is when you're building those data pipelines or the analytics, maybe building an AI model to do predictive analytics.

17:00Jon Krohn:This might be proprietary, so maybe There's nothing that you can disclose here, but if you're able to, what kinds of frameworks are you using there? Is it Databricks alone or are there things like scikit-learn involved, other kinds of regression modeling technologies or something like that? Yeah, it's a mix, to be honest, because one of the main challenges, you don't have the same amount of data across every customer. Some customer, you work with them and they have three years of historical data. You can train some pretty advanced models there. You can go into deep learning techniques, which allows for like more cross machine training instead of being one model, one machine.

17:42But not everybody is there, right? So you have to think about how do I get started quickly? And oftentimes some of those simpler techniques will work really well when you have limited capability of limited amounts of data. And so that you can get at least get going and provide some value, right? At the end of the day, what's most important is time to value. And so we use different techniques depending on the amount of data. And that's something that we're refining as well so that we can have kind of like a maturity curve of like having maybe some more generic models, simpler that works with you at the beginning.

18:14And as you progress, we can train some more advanced models with more advanced techniques that really brings you to deeper insights on this front.

18:25Jon Krohn:Nice. Thank you for that answer, for giving us some insight behind the curtain. And then in terms of this is a question that I know you can kind of go into a lot of detail on, and I'm not asking for any kind of secret sauce here, is when you're building a tool like this that is supposed to look and feel intuitive, when you're building some kind of predictive or AI capability into that tool, what kinds of considerations do you need to have to ensure that a non-technical user, You know, a lot of our listeners are technical hands-on practitioners, but if you have somebody that's used to having a clipboard on a manufacturing floor, and now they have all this AI functionality within an application, what kinds of considerations do you have to make that AI intuitive?

19:16Jon Krohn:Yeah, intuitive for the users and just feel like a good experience? Yeah, no, that's a great question. And I think that's something we've been spending a lot of time. So a couple items related to this. One, it's making sure that you're clear when there's something that is AI generated. You're working with, you want to make it clear when something comes from AI so that you have clear sources as to where that information comes from so that it can be double checked. You know, in a lot of cases, if you're writing content or emails with AI, if some of the content is hallucinates a little bit, you're not in a dangerous situation.

19:51but people are working with machines that have high voltage, high frequency. A lot of things can go wrong. You want to make sure that when you are recommending steps or actions that somebody wants to take, that it really is grounded on manual information or authoritative information. And so we have a lot of checks and balances in place to make sure that we reduce hallucination. We have multiple models that will check the answers also afterward and validate what was generated. But then the final step is really making it transparent to the user, really being clear with them that this is something that comes from AI.

20:26And here's the sources so that they can double check and validate. Because at the end of the day, safety is really, really important to us. And it's important to our customers. And we want to make sure that we keep that in mind. So that's probably the first principle. Second, it's making sure that AI will be where you expect it to be. What I found from the best tools that I've used on the software side is that maybe it's the settings or the button or the action that I want to take next is always kind of in my workflow and in line. Not something that I have to go check three menus down the line.