AWS re:Invent 2024 - Transform troubleshooting in manufacturing with AWS gen AI services (IOT204)
- Welcome to our session, Transform Troubleshooting and Manufacturing with AWS Generative AI Services. I'm Dan McGrath from Cognizant, I lead our smart manufacturing technology development. - And I'm Barry Bonar from Trane Technologies, a director for enterprise, or (chuckles) not enterprise architecture anymore, but emerging technologies. And look, thank you for all coming into this session.
I don't know if you chose this session on purpose, or if you couldn't get into any other session, 'cause I hear it's hard to get into them, but thank you anyway for being here. Trane Technologies, who we are, hands up just in the room. A show of hands, who knows or has heard of Trane Technologies? Okay, quite a few. That's good. For those that haven't, Trane Technologies, we are a global climate innovator. So what does that mean? We're a HVAC company.
Heating, ventilation, and air conditioning. You might have heard of some of our brands like Thermo King in our transport business. We have Trane in our commercial and residential business, and we do a lot of work these days keeping data centers cool, so it's good business for us to be in with the expansion of data centers and AI. We are headquartered officially in Dublin, Ireland, which is where I'm located, but our operational HQ is in Davidson, North Carolina.
Globally, we have about 45,000 employees, depending on how you count it, and about 45 manufacturing facilities worldwide. And our chair and CEO talks all the time about our purpose when he's talking publicly. And our purpose is to challenge possible to innovate for a sustainable future. So we're big into sustainability and everything that we do and that challenge possible and innovating for that sustainable future, part of that is what we're doing here, and we're gonna talk to you about today, how we're challenging possible and pushing the boundaries with generative AI and manufacturing.
- And I'm Dan McGrath from Cognizant, and we engineer modern businesses to improve everyday life. I'm gonna tell you what that means to me. I am a long time OT manufacturing worker that has engineered and control systems and even set up IoT, you know, to the cloud systems over the past 30 years. But you know, too often in my companies, we've run into problems that slow down progress. Either it's skills gap, you know, technology issues, and we just kinda run out and they turn into just POCs.
Well, at Cognizant, you know, since I've been at Cognizant the last six years working with manufacturers, I've seen how a very large organization like Cognizant with 350,000 plus employees globally with deep expertise, with passion to solve problems and strong partners like AWS, can really change that game and solve problems. So today in today's presentation, I'm hoping to provide you with a few takeaways to help you drive value in your organization, especially with generative AI. So first off, here's our agenda.
And I wanted to ask, is anybody actively working in manufacturing in this room? Could I see a show of hands? - It's actually kinda hard to see the hands here with the lights in. - Yeah, it's hard to see with the lights, but yeah, a good number. How about manufacturing IT? All right, very good. So that's really good.
So, you know about assets and how organizations are trying to provide uptime. So today, we're gonna go through how generative AI can make a big impact there and how to address the challenges that exist and why this is different than some of the other technologies like IoT, that sort of, maybe didn't totally pay off. And we're gonna show you a working demo and take you through steps that you can use in your projects. So starting with those challenges, as I was mentioning, you know, the skills gap, when you're trying to engineer systems, but think about the people on the plant floor with the range of equipment that they have. There's some plants with very old technologies that don't have very much sensors to new machines that arrive with new operating systems and interfaces, and then the line stops.
How do you solve that problem? You try to call somebody. Where do you even find the expert? You know, a lot of manufacturers are having big problems, even hiring people. You know, I could ask for a show of hands of how many people have open recs for over a year for certain people. I mean, it's very common.
And then next, the pressure to reduce downtime. This is, I think, like a $35-billion problem. Really. The value at stake, a manufacturing plant is like $22,000 a minute. If it's a pharma, you know, if you lose a batch, it could be a million dollars in loss. So there's a huge value at stake and staffs are only getting leaner. So these are some of the problems and why we we chose this problem to talk about today.
- Yeah, and some of the process challenges then that folk face, not just in Trane Technologies, but generally across all manufacturers is the IT-OT challenge. And generally, you know, folk in IT have better, more mature practices and processes and standard work in place than some of those in the OT side. And when both meet, 'cause you know, there's a lot of IT components involved in the OT side, but when both meet, we get that sort of friction at times. Now, luckily, in Trane Technologies, we've got a good relationship with our OT teams. We have good collaboration with them and we work well together.
But still bringing new IT solutions to an OT problem or trying to figure out, you know, the accountability, the governance around that, we haven't quite got that all figured out yet, but there's a lot of work to do there. A lot of what we're gonna show you today hits that boundary between IT and OT. So you're gonna see some, you know, OT-type equipment with IT-type layers on top of that. So trying to figure all that out or challenges as well that we see in these existing operating models. And the last point there about, you know, don't fix what ain't broke.
If I had a gun with one bullet and I could shoot one person and get away with it, it would be the guy or girl that came up with that phrase, but not breaking, you know, don't fix it if it ain't broken. You, of course, you should fix it if it's not broken. And that's what we do. We challenge possible. We look at things that have been in place for a long time and just because they have, doesn't mean there's not a better way to do it.
So we challenge some of those mentors as well. - So, you know, we're up together on the stage today because of collaboration between a client and a provider like Cognizant. And on this slide, we're really talking about how do you innovate? At Cognizant, we've set up a manufacturing innovation center in Chicago at a place called MxD that's focused on digital transformation for manufacturing and cybersecurity. So we have a lab space there where we've staged various use cases including generative AI, advanced analytics, MES, and other technologies. And so, we have a great platform there to work with clients and solve problems, so that led to some work with Trane. - Yeah, at the same time, we had set up, we had two different labs.
We have the IT, the middle box here, or emerging tech lab. So we do a lot of prototypes and proof of concepts from an IT perspective, and we showcase those in what we call our emerging tech lab. And we get a lot of traffic through there 'cause again, this is based in Dublin where I sit, as I said, we're officially located HQ in Dublin, so our board of directors have to meet there every, at least once a year in that place, in that facility. So they get to come and see this emerging tech lab and how we're innovating and advancing from the IT side.
But now we're also adding on that extra OT layer and that's the way in which we bridge the gap getting from IT into OT. And on the right hand side, you can see our OT labs that we've set up. So we have three of these called MTDCs.
They're manufacturing technology development centers. And we like, from the IT side, to incubate things early and do small things in the emerging tech lab and then move them into the MTDCs. And they try them at a bigger scale in the MTDCs, so they have actual manufacturing kind of scale before we roll it out into production. So anything that we're trying in a POC-type fashion, and I think, Dan, you mentioned earlier, problems getting from POC into production.
I think we've probably all faced that, we can do POC purgatory, but we believe this is a good model to help us get past that POC purgatory where we have somewhere else to go after POC before production. And we try this out in our MTDC labs as well. So, Dan showed us a wonderful toy that he's been working with, a motor fault simulator that layers in generative AI and manufacturing. So I worked with Dan, we worked together, and we got that set up in our emerging tech lab in Dublin. And you have one in your lab as well.
But from the IT perspective, we've done a lot of work with generative AI. My job is in the emerging tech function. I have to constantly look past where we are today and see what's coming next. So when the buzz broke with ChatGPT about two years ago, we knew about it. We knew about GPT and we were ready to adopt at that stage, and we knew why we didn't want to adopt ChatGPT, but we were very quick to build our own internal GPT platform.
We called it EVA or Employee Virtual Assistant. We then incubated an AI foundry to own that product, and they're adding a lot more use cases to it. So now, in my team, we go and look for, well, what's next? Where do we move next? And clearly for us, the next move was generative AI and manufacturing.
And that's what we're working with. I'm gonna show you some demos here with Dan in Cognizant. - So at Cognizant we've been pursuing generative AI for some time and there was a natural fit working with Barry on that. And also, digital twin is a key foundational element that's very popular in manufacturing because you're able to get all that contextual data on AWS and do analytics, and it's a natural for doing more productive generative AI and solving these problems like troubleshooting. Automation and robotics are also a related trend area that we're bringing to our joint labs because robotics are, you know, have dramatic impacts on productivity.
And now with generative AI technologies, there's even ways to help do simulations and planning there. So at Cognizant, we've been working with AWS on our generative AI suite of solutions. So if you look at this slide, you can see that manufacturing can be impacted in many ways, starting with information assistance. So from basic summarization, so many plants have all kinds of standard operating procedures, troubleshooting guides, you know, you walk up to a machine, you'll probably have some clipboards with multiple-step troubleshooting or things taped on equipment, even, you know, when you start looking at, you have a lot of legacy older systems that you wanna update and migrate. Using generative AI for code management and engineering the code, auto generation is a key use case.
And then doing that, setting up that digital twin data contextualization, all of this is building up to real value with insights and workflow transformation, whether it's smarter standups. So instead of having these meetings in front of the whiteboard, having a digital dashboard where you can see the KPIs and what happened overnight, drill down into the problems, create service tickets, much more dynamic. An advisor approach, that's what we'll go through today with a live demo of a troubleshooting agent we built and even, you know, predictive analytics, predictive maintenance.
The end game here with generative AI will include expert agents. And that is really where you have multiple agents helping you do a variety of workflows such as monitoring for problems, triaging them, even implementing fixes. So, you know, whether it's a cybersecurity incident, whether it's some IT outage or OT problem, the automated guidance and the value that generative AI can bring is really exciting. So at this point, I don't know, Barry, which ones do you see most? - Yeah, well, again, the advisor.
But the key thing for us is, you know, we didn't want to introduce yet another application for our factory worker operators to use. We wanted to meet them where they were with their standardized MES and MOM and SCADA-type systems. So that's where we wanna integrate this generative AI layer inside those systems. And that's where we're working with Cognizant, getting generative AI capability built into Ignition, our standardized system for SCADA built into Aprisa, our standardized system for MES.
So the users don't have to, or the operators don't have to context switch, and they get the power of that generative AI with the content and the contextualization of the motors of the sensors and the data coming off those sensors. So like Dan said, he's gonna show a demo on the troubleshooting, you can ask questions, and it knows about the specific context of that motor or you know, the data coming off of that motor at that point in time to augment that with the model from the generative AI. - So when you take a step back, how do we avoid getting into the POC, you know, purgatory situation? It's really picking the right problems to solve that have high ROI and are scalable, creating an infrastructure where you can build and scale.
And that's where we find a lot of value working with AWS and finalizing that tech stack. So we're gonna walk through some of the architectures, but when you're talking about a lot of these older legacy systems on the plant floor, older documents that may exist in manufacturing, you really have to puzzle out where is it most cost-effective and scalable to set up these data flows. So now, we're gonna get ready to go through the demo, you know, we really worked on in our labs first to show the template and how we can set up these workflows and data flows for generative AIs using LLMs and intelligent prompting. So, you know, we'll talk about how we deployed at the Dublin lab and you know, in general, this idea of test it in a set up your sandbox, set up that environment will lead you to be able to make a scalable solution.
- Yeah, and look, it's easy to have IT systems in there and demonstrate them, but when it comes to OT, you know, we need physical devices and motors and things like that, it's hard to get access to the plants to try things out and do POCs there. So the approach we took was bringing this motor into the emerging tech center, putting on our own sensors on that and collecting the data from it. So the challenge we face then with the IT-OT teams are saying, well, you know, we haven't standardized yet on the sensors we wanna use to get the data off the motor. We haven't, we don't know how to get the data off these machines. We don't know what we do with the data.
We we're not ready to start yet. But, you know, in the true innovation sense, we always think about starting before we're ready and not the standstill because someone says our data foundations aren't right, we're not ready to start. We've gotta start somewhere, start small, and show the outcomes, the progress, the art of the possible, and then use that to start to push in and help the other teams, you know, prioritize getting the data foundations ready to unlock this value that we can show through the demos and POCs. - So the, you know, you can see our pictures there.
Over the last year, we've been collaborating the, on the left, there's a shot of Cognizant's Innovation Manufacturing Innovation Center in Chicago at MxD. And we have an array of different types of demos that we'll talk about a little bit, but we have a machinery fault simulator and you know, even a working cup filler machine putting M&Ms in a packaging equipment. So it's an area that we can explore the real world use cases and manufacturing and how to leverage IT technologies like generative AI.
- Yeah, I think the picture on the right doesn't really do it justice. I don't know who took the picture, but you can't really see what's in there. But trust me, there's a lot of cool stuff in there that we showcase when people come to town. - So here is a shot of some of the demos that are common to both labs.
As Barry was saying, the challenge when you're trying to stage and create a problem that mimics a real-world manufacturing issue can be, you know, a problem at the stage. So on the right you see a machinery fault simulator. This is a motor that has two bearings and two discs. And the idea is you can create a misalignment, you put weights on those wheels, create an imbalance and even put in bad bearings. So when you ramp up the speed of that machine, it generates, you know, noticeable vibration and sound. Kinda like if you got an old car that has, you know, a bad problem with the wheels and so on, you'll feel that well on this equipment.
You'll feel that. We've put on IoT sensors. So we monitor vibration, temperature, motor current, RPM, streaming all this data through an edge gateway through AWS and the cloud. We're able to do analytics on it. And then on the left, you can see what looks like Legos and, you wanna describe that? - Yeah, it is Lego and I got a lot of grief when I said I wanted to put this in place. People saying, "Do you not have a real job to do, you wanna play with Lego?" But they soon changed their mind when they saw what it could do.
And really what it is, is it showcases, a lot of people in IT have never visited one of our plants where you can actually see what we make, the products that we make and how we make them. This is a way to bring that scale down. So within that, we've got, you know, the high bay warehouses.
There's two stations on a manufacturing line. There's a sorting line, there's vacuum gripper robots that move things and moving things up and down the belt from one station to another. But underneath that, the IT systems, all the IoT and the MQTT protocols and the OPCUA protocols between the PLCs and to move the robots, that's the same technology that's used in our plants at a bigger scale. So we can try things out small here and make changes and see how things work. And as our factory teams and OT teams introduce new standards, we can then, you know, incorporate them here or in some cases, we can inform those standards by trying new things here fast in a small, small way.
But like the graph showed earlier the three different charts, this is in the engineering lab first of all, and then it moves to the MTDC, and we put it in place in a bigger scale before going to full production scale in our plants. - Yeah, I'll just add to that, that tabletop factory has industrial PLCs, Siemens, and it it has raspberry pie with node red, video analytic, RFID. So it's a real, scaled down model, and we're able to bring that data and do analytics and generative AI on it.
And we'll talk more about that here where we're showing the high-level architecture diagram. You can see that we have those assets connected to ignition edge. So from there, it's forwarding data to Ignition Cloud. We also have a Aprisa as an MES and in the middle there, you can see AWS.
We have a data flow for doing, you know, the LLM chat bot that will walk you through today that generates intelligent prompts and orchestrates generative AI answers on troubleshooting. Included in this high-level diagram, you see NVIDIA Omniverse. So the concept here is we can bring in 3D models of the equipment and along with the live data streams and do, you know, a 3D digital twin and even simulations.
So this high-level solution architecture brings together, you know, something that are in both of our labs in Chicago and in Dublin. It's actually funny, I'm just remembering now that how we got the idea for that factory in the first place was actually from Amazon. They used to bring these factories to their shows as well and integrate them into their systems. So that's where we got the idea from to bring it to our lab. And now, we're integrating into our own systems, Ignitions and Aprisas with that as well. - So zooming in on that box in the middle, you know, you wanna look at the architecture of how we bring the data in from the edge.
So we have those PLCs, in this case, we have some IoT sensors through a edge gateway and we send that to Ignition Edge, which then with MQTT, we're able to forward to AWS and so now, we're able to bring it into, you know, we have a Ignition in the cloud. And then you could see how we have the manuals and SOPs loaded into Amazon S3. And with OpenSearch and a vector database and RAG approach, we're able to develop the prompts using a configurable prompt template and chain-of-thought prompting, so that you're able to come up with, you know, a question that a plant floor worker might ask about, well, how do I troubleshoot a vibration problem on my machine today that's causing me downtime or potential downtime? And it's able to join the manuals and SOPs along with the real-time alarm data that it can read through an API from Ignition.
And, you know, included with this is we're also doing AR-VR apps where we're able to bring that to the field worker, so they can have an iPad, or you can do a virtual reality with a headset and have some unique ways of interacting with the data. But at a high level, this architecture is, you know, key for enabling value on the factory floor. - Yeah, and it's a flexible architecture, so we haven't standardized yet on the factory side in terms of what sensors we're gonna use. It doesn't matter. We get the data off those sensors, put it in the same place.
It's accessible via an API into Ignition that can access that data, same with Aprisa. If that system changes out or Ignition, even if that system changes, it's flexible and Amazon mightn't like to hear this either, but we can change the model behind this. We don't have to use Amazon or you know, the models behind Bedrock or LlaMA. We can plug and play and put different models in there as well. So we kept it scalable and flexible because we know a lot of standards haven't been fully defined yet.
So when they are defined, we can easily make changes and accommodate that. The other piece I'd say is when we enter the prompt, it's not just the prompt going to the model and you know, a bit of RAG happening, that chain of thought that you talked about. It interrogates the prompt and understands the context of the question.
So if the operator's asking about, you know, a particular data that came from it, it knows then to go to the API and to Ignition and get that data. Or if it's asking about standard operating procedures, it knows to go to S3 and get that data and augment that and delivered. And then in terms of hallucinations and whatnot, it's operating on a very defined data set. So the chances for hallucinating are a lot, lot, lot lower in this case as well. - Yeah, thank you.
As, you know, look deeper into it, you can see that Orchestrator framework using Amazon Bedrock and LlaMAs is what we used for the demo we'll show you in a few minutes. So setting up for the demo, you know, we wanted to explain that there is a ignition dashboard that's used in Trane, and we also have at our innovation center to view the streaming data from our machinery fault simulator and from our Agile tabletop factory with the Lego-type system. And we're gonna show you how we've implemented a Factory Whisperer to answer some of those troubleshooting questions. - We're not a hundred percent gone on the name, Factory Whisperer, we're open to changes.
So anybody got any good suggestions? So we'll listen to you about Factory Whisperer. That's what it is for now. - Yeah. So at this point, we'll transition over to a live demo - [Barry] And when this was developed first, what you see here, the APEx dashboard is, you know, it was another application, and that's what we were trying to avoid, and that was one of the pieces of feedback we got from our board when they came through the innovate, the emerging tech lab. They were like, "Well, why another application?" Can we not integrate this into it? And that's where the idea came from to build it into Aprisa and Ignition. So Dan's gonna show you the differences here.
- [Dan] Yeah, so this is Cognizant APEx which is an accelerator that runs on AWS where we can bring in all the factory data, create dashboards, do analytics, and in this case, I'm showing our manufacturing innovation center in Chicago. And you know, your typical hierarchy, ISA-95 style for like a unified namespace that you'd create. So in our space, we have this cup filling machine, we have, you know, some cobot, machining, mills, packaging equipment, modeled here. We have the Agile factory which we mentioned that you know, the tabletop LEGO system.
So you can see different parts of that equipment that include a high bay warehouse, a drilling station, a milling station, a dispatch, a quality assurance station. The unit that we have in Chicago, it actually has a little AGV that operates. So when you load a RFID tag piece into the system, the AGV picks it up and moves it around. In the older unit, you have like a crane that moves it around, right? - [Barry] Yeah.
- [Dan] And it's sort of a, you know, you wanna do troubleshooting. Well, this unit helps you do troubleshooting 'cause it does run into problems, so like a real factory. And then the machinery fault simulator, again, that's that unit that you can purposely create excessive vibration. We're showing a dashboard we created that has the bearing temperature.
You can look at some of the vibration from some of the sensors here, like a frame sensor, and you could see if we're getting excessive vibration or look at the tachometer and the speed of the system. But, you know, trying to discern, you know, where you have a problem. Too many dashboards on the plant floors. Sometimes it can be information overload.
So this is where this concept of the Factory Whisperer comes in. You can ask a question like, you know, I have staged here, tell me about major alarms today. And it said, we have one major alarm today, a high vibration alarm. And you could ask it things like, well, you know, give me the the reasons for this.
And you know, it'll come back with a response. You can actually create tickets. And in this case, we've integrated with ServiceNow, so when... - [Barry] It never fails during a live demo. - [Dan] Yeah. So we can like open up a a dashboard here and look at, you know, we're getting vibration simulator alarms that are, have been generated from opening a ticket here and, you know, we can drill down.
So this is our APEx version of the demo. - [Barry] And just on the ServiceNow piece, why that's interesting to us, we use, on the IT side, we use ServiceNow and ITSM. But with ServiceNow coming out with the OTSM piece, that integration and what we need to do there opens up a whole new, you know, a lot of doors for us on that side as well. - [Dan] So when you talk about Ignition, how many people are familiar with Ignition? I see a few hands. It's a very popular software in manufacturing today that has attractive licensing and ease of use.
- [Barry] Almost too easy in some cases where people can go and build full ERP systems inside of which we don't want, but we've standardized now on, using it just for the right purposes. So it's a very good platform on the plant floor for, you got these older equipment or even newer equipment, and you wanna create your information dashboard. So in this case, we have that same inspector Motor Machinery Fault Simulator, and the idea was let's create a built-in native app to Ignition, so you're not opening up another webpage. So with Factory Whisperer, again, we have this, and we could ask it to share the alarms and you know, it'll come back with an answer. I've kind of, you know, pre-stage it here, hopefully the WiFi permits, but there it goes. So it just came back with the answer.
Yes, we had the high vibration alarm with a certain criticality, and I can ask for reasons and prompts. So the user interface is a little bit different than doing it in React or building your own app like we do in APEx. But this is the platform that there are many engineers in the OT space building dashboards. So the idea here is we'll prove out how we can build an app that makes these API calls to AWS and generative AI for troubleshooting and make those part of the arsenal of tools that engineers can use on the plant floor. - [Barry] And again, Dan's just clicking on the sort of pre-suggested prompts.
You can type your own prompts in there as well. So these are just helpful type prompts like you get with ChatGPT or others, but you can enter your own prompts as well. - [Dan] Yes, and you could have hyperlinks in there to open up documents, videos.
Oftentimes, there's videos on how to troubleshoot or calibrate or something in your company and SharePoints. So having a collection of documents, your troubleshooting guides, any documents of that sort, we're able to join up and make part of that corpus of documents that we can use as generative AI prompting against. - [Barry] Another key thing there as well is generative AI for us in IT, we wanna keep the number of apps again down. We don't want a new app for every generative AI thing that we do. So we're trying to, you know, keep that under control. In most cases, we've set the principle that if we already use an enterprise platform and generative AI is embedded in that platform, use that in platform.
And the same holds true here. Ignition, if Ignition had that native capability built into the platform. We wouldn't be working with Cognizant doing this, we would use that in platform, but I think Cognizant have been talking to inductive automation about this as well. And or do you wanna talk about that story there? - Yeah, yeah. So the idea of making an app that's part of a product, it helps if you're a partner with those software developer companies. And so, we're working with inductive automation on this app.
And also, when you talk about MES systems, we have partnerships with many of them and this generative AI agent approach is very powerful and we'll be able to do a lot of these, you know, in this case, Ignition, Aprisa, and there can be others. So as we look for the reference architecture for enabling the smart factory of the future with AWS, there's really a broad range of things that you can accomplish. It all starts at the edge with getting the data organized well, and Ignition Edge does have this concept of the unified namespace.
So if you're familiar with that age-old problem of having too many systems with the same asset named something differently, or SKUs, the nomenclature in one system or another. So you may have SAP or Oracle with your asset hierarchy than in your maintenance system. Maybe you got a little bit different and then in your ignition project, somebody named something different. So imagine bringing that all up to AWS and instead of a data lake, you have a data swamp they call it, right? So how do you get out of that? Think about a unified name space where you've done it with standards basis.
And so, in our reference architecture of the future, we're showing how Ignition Edge and working with tools like in this case, HighByte, SiteWise, AWS SiteWise, we're able to create a data lake adhered to a naming standard and do not only generative AI with Amazon Bedrock and related LLMs, we could also use SageMaker and do analytics and do many other different workflows from video analytics. And we'll be doing that in our innovation center and showing how video analytics, cobots, generative AI, can all lead to higher productivity, less downtime, improved yields. And as part of that is, is making that same data set available in Omniverse for doing simulations or into other third-party apps. - Yeah, and we're always, even internally in Trane, what we just showed you is sort of next steps and people aren't almost ready for it yet. Or like, we're not ready, the Omniverse and all that, we're pushing the boundaries even further now looking at this. But that's our job as we said earlier.
But, you know, challenge possible to innovate for a sustainable future. We have to continually keep looking at Omniverse where it's going generative AI, what's next, what's next? And being ready to adopt those technologies when we're fully ready if you get me. So going through those POCs and reference architectures like this help us keep an eye on the future as well. - So a key concept to take away today is that, you know, generative AI can scale beyond doing a quick POC you know, a lot of people are dabbling in this. They can say, oh, yeah, in one week, I just did my own little chatbot.
And you know, and you did. But to make that scalable, secure, and solve more challenging problems, I encourage you to look at this agent-based approach. So the idea here is to have separate agents dedicated to various functions and in this diagram, you can see an agent that helps identify the context of the input query, another agent that looks at what sort of question it is and which LLM to best use for it. And then how do you form a good configurable prompt and use the templates that you've set up for different use cases and be able to use the digital twin and knowledge graph that you've set up. So in this approach, there are a lot of benefits and key features here to call out that by using multiple AI agents, they can work independently, you can revise and update as new technology and services are released.
And you can share though that solid basis of the context, so they have a unified namespace. Each agent can tap into that on a common basis, and it'll really drive scalability and efficiency and avoid maybe that kind of, you know, if there's a new LLM that comes out that is better for certain problems, these agents can help you sort through that and make the right call at runtime. So the goal here is that modularity and configurability. Any comments on that? - Yeah, look, when we talk about agents that are out there today and people are starting to use it, if we wait for the next thing or wait for the next thing, or wait for the next thing before we get started, we'll never start.
So that's why we're already started with what was there, what was available to us. Just plain old Gen AI with RAG and a bit of APIs and augmentation. This agentic architecture, we're looking at that as well, getting aware of it, but we're not gonna stop what we're doing and say, oh, let's wait until this is more real and do the agentic piece then, and then we'll, actually, when we're doing that, there'll be something new. So you gotta start somewhere and put a stake in the ground, and you can always evolve and keep moving on. So my advice would be to start somewhere and start before you're ready and don't get blinded by all the other things that are out there and coming down, yeah, from left and right. Just start somewhere.
- Yeah. So what's next? You know, it's, generative AI is relatively new, but mature enough to use the manufacturing, and we've started on these first use cases like the troubleshooting agent, but there's much more available. - Yeah, and if you look at some of the vendors out there like Boston Dynamics and others and what they're doing, you know, they claim you won't need any people in the factory at all. But, you know, that's, will robots and AI really take our jobs? People say, no, but you know, people that don't use it, the people that do use it will take the jobs of the people that don't use it, but I don't know. There may be, in the future cases where robots could actually take some of our jobs. - So by using yeah, video analytics and cobots, we believe we can streamline the setup maintenance and value of cobots and robots and factories, and we're gonna prove that out in our innovation lab.
And it really feeds into a roadmap idea and vision of a autonomous factory where you use data, automation, robotics, video analytics, to really maximize the value of your production assets and your people. And the ways of working can be changed today with generative AI as we talked about today. You know, instead of having to rely on that one expert in the plant that knows how to troubleshoot something, you can democratize that and make it available to everyone in every shift by building in this kind of capability. So we feel, as Barry says, that there's some quick wins out there. Start today, you know, it's building that foundation with the digital twin unified namespace will help you progress to the next tier to do more with simulation and planning using cobots and robots simulation, physics space models.
And in the end, you're gonna be able to continue to adapt some of these technologies. And you know, the idea is the every plant is unique and with a lot of different legacy processes and that's one of the challenges why manufacturing often lags behind a lot of the IT technologies. But I think the tools are getting smarter and, you know, the challenge here is to keep innovating and I encourage you to learn more about how generative AI can be a key part of that. - Yeah. I suppose I was at an innovation conference recently in Copenhagen and I heard someone talk about, he said, "When the winds of change blow, some people build walls and others build windmills.
Now, he was in Copenhagen, so windmills were kind of apt there. I don't know what would be considered here, but I think the message there is, look, you are gonna face people saying, no, we're not ready. We can't do this. But build a windmill, build the next windmill, build the next windmill, and get ready, you know, start somewhere. So I know we are the last thing standing between these folk and lunch. So we purposely ended a little bit early here, but we do have time for Q and A if anyone wants to ask any questions either now or off to the side of the stage, we're happy to take them.
2024-12-09 20:10
