Video: AI Agents In Action: Modernize Manufacturing Maintenance | Duration: 3832s | Summary: AI Agents In Action: Modernize Manufacturing Maintenance
Transcript for "AI Agents In Action: Modernize Manufacturing Maintenance":
Hello everyone and welcome to today's webinar. I'm super excited today to be, presenting the next webinar in our AI agents in action series. Today, we'll be going to the topic of agents for manufacturing maintenance. So my name is Henry Ekwaro. I'm a senior solution consultant here at Dataiku focusing on the manufacturing industry and I'm joined today by my good colleague, doctor Nicolas Mauhé, who's a senior data scientist in our team. And essentially, our team works on tailoring the Dataiku platform to best meet the unique needs of manufacturing companies such as yourselves. So to start off, I'd like to give a brief overview of what Dataiku actually is, what our platform offers, before we go into the topic of agents specifically. So essentially, in our platform, looking at the overview here, going basically from the bottom to the top, we provide basically, connectivity to all data types or compute engines, wherever it may reside. So basically whatever data strategy you have, our platform is compatible. Also in terms of technology, tech stack and the layer above that, basically, companies can decide what the best tech is for them, what they need today, and also be ready for the tech of tomorrow, by avoiding any vendor lock in through our, tech agnostic approach. And very important also the topic of governance, we provide basically a framework to govern all the different steps you may have in the data, to AI life cycle. Basically, going from data prep all the way through to operations. Everything is fully governed with Dataiku. And, this is also an increasingly important and critical topic now with the rise of agents. At the core of the platform, we offer basically any type of data driven analysis you can imagine. So be it simple analytics, machine learning models, and also now recently, building agents and coordinating them. And basically, in terms of the users, we really offer a collaborative solution that allows basically for all users, to be met at the skill level they have. So be it no code, low code or high code users. And, so while we're discussing agents today, everything that we'll be talking about is basically fully part of the Dataiku Universal AI Platform as we like to call it. This means that our customers really benefit from a proven platform when they're thinking about the topic agents where they can connect agents to their data, their analytics, and the machine learning where needed. And it's really critical to note, in our opinion that agents can't live in a vacuum. So if you really want a business value from them, and we really believe that they'll only be able to be truly valuable when they can integrate with all other objects shown in this diagram and at the same time be governed and controlled in a similar manner. The key value that Dataiku really brings is that our platform enables companies to do the creation of AI agents at scale while providing the right controls. So with that said, let's jump into today's topic of beginning with how AI agents could change the game for manufacturing. So maybe to begin, what are actually AI agents? It's really a topic where you ask three people, you'll get three different answers. But, from our point of view, we basically see three components of what makes an AI agent. So first of all, at its core, it's an LN powered system. Now secondly, it's designed to achieve objectives through multiple steps, while having the concept of going beyond just a simple, single retrieval of one information. So that's how we define agents and when we're presenting today, that's kind of the the basis of of, what they all include. So what is like a typical motivation that we see for for manufacturers specifically to be interested in AI agents? As you've probably also learned yourself, it has major potential for improving the efficiency of your operations, be it by reducing equipment downtime, break of maintenance or optimizing your production processes across operations. Agents offer a lot of potential here. Also in terms of breadth and coverage, being able to handle more production decisions, more quality checks and tasks with the same workforce by enhancing these with agents as a powerful tool. Consistency also a big topic in manufacturing. That's always a high priority. With agents, you can really standardize your approach, be it to equipment monitoring or to process improvements by, having the same basis for across different shifts, across production lines, or even factories. And also, definitely not to be, omitted is the topic of employee satisfaction. So, with agents, it's really an opportunity to to to automate and streamline some of these, monotonous tasks to allow employees to really focus on more high value problem solving, and continuous improvement work that's actually more engaging and rewarding. So what are some of the use cases that we see in the manufacturing space? I collected here just a few examples that we've worked on here in Dataiku or that we've discussed with customers. Some that we have in production, some that are concepts and these I'd like to share. By the way, if there's any questions, feel free to write them into the chat, into the q and a and we'll get around to answering those either by chat in parallel or afterwards. So we kind of see three core areas where agents are really creating tangible value on the factory floor and it's important to note here that in our opinion, it's not really about, it's definitely not about replacing people but more about fundamentally enhancing how we automate processes, how we empower our workers and how we connect our operations to direct business outcomes. And that's kind of a theme that carries through through all of these. So beginning on the left, we have the topic of process automation. This is kind of the most, I guess, simplest area of of of agents. It's kind of the foundational layer for the for the other ones. So here you're targeting repetitive rule based tests that just consume a lot of valuable time and are kind of prone to human error typically and the main outcome here is cost reduction by avoiding these mistakes or automating them. So for example, on the topic of, you can have automated quality assurance reporting. So here you could imagine they see a quality manager who would, normally spend the first hour and a half of, his shift every day pulling data from different systems. So process control data from the line, lab results, manual inspection logs, basically just to build a daily report. Where here an agent could completely automate this so that when he starts the shift or when she starts her shift, you can basically have a report waiting in the inbox within with the, highlight of what the trends were for the night shift, any flags of irregularities, and this could really, yeah, so so save a lot of time for for someone like this. Similarly with with having an agent for example in place for monitoring the safety in a plant, this could really drive down compliant risks and of course prevent accidents by automating this monitoring and flagging anomalies or even ensuring that processes are followed. Then moving on to the center pillar, here we have the topic of worker augmentation. So here, the typical result is massive jump in productivity. So again, this isn't about replacing like a skilled workforce, it's really about enhancing it. And the key benefit here is improved effectiveness. So for example, the second one we'll be demoing today. Maintenance scheduling assistant. Here the idea is that, you have a significant leverage to improve your OEE. So basically, instead of scheduling maintenance in a manual way with an agent, you can really do it an optimal way, get recommendations very quickly in terms of what to prioritize and really go from a very complex, multivariable problem that takes long time to to do ideally to a very quick solution in a matter of seconds sometimes. Now similarly for operating machinery as we have here on the first example, so especially for complex machinery, for example, if an operator is relatively new, an agent can provide suggestions for how to change settings depending on certain conditions that are currently happening or certain products that are being produced. You can give warnings if any certain operations were missed or certain steps were omitted. Then on the right finally, we have what we call Intelligent Business Chains. This is kind of like the most advanced area in my opinion, but also at the same time, the most strategic application of agents. We're really starting to get the most value and here it's really about connecting basically historical and real time intelligence from the factory floor directly to commercial and financial outcomes and then thereby really enabling new revenue models even. So it could be around the topic of optimal processes and operations. So with the production process optimizer as we've written the concept here, the idea is basically to have an agent monitoring the entire plant. So let's say for example, you have critical raw material shipment that's delayed. Here a human planner could take hours to recalculate the schedule. But with an optimizer agent like this, it could assess, the different impacts and evaluate basically which orders have the highest margin, which, sequences, the entire production flow, could be implemented to maximize the profitability despite the, disruption. Basically, give a suggestion for how, if you're assuming of course your manufacturing process is flexible enough, how it could be changed and reconfigured to minimize the the effect of such a supply chain disruption And that could be a major competitive advantage especially in these days where supply chains are so, uncertain and unstable. And then here as last example, we call this machine as a service uptime agent and, for example, we're seeing even more and more, companies especially industrial equipment manufacturers. So instead of just selling a machine, they sell a guaranteed outcome as a service. For example, 99% uptime. And basically, it would be that, for business models like this, an agent could live on a customer's equipment, one that they're leasing or they bought and essentially, not just predictive failure but also orchestrate the entire service response and thereby helping ensure that the SLAs are met. This really helps companies move basically from a like a one time capital sale into entirely new revenue stream. So it's a super exciting idea in my opinion. So basically, as you can see, these three pillars aren't really independent. They kind of build on each other in terms of the complexity and in terms of the basis that they provide. Basically moving from automating processes until really, creating entirely new data driven business models. And, yeah. These are these are some really exciting ideas we see. But today, we'd like to go into more detail on the maintenance scheduling as we've announced. So, at eye level, what is our maintenance scheduling assistant? After I'm done speaking, my colleague Nicolas will be giving a live demo so we'll be able to dive into the details. But essentially, what it does is provides an intelligent schedule and orchestration around your maintenance activities. So basically, analyzes, your equipment health, the production demands, the resource availability to then suggest maintenance operations and which machines to prioritize the maintenance on as well as, a possible maintenance schedule that you could actually use for whatever time frame you've requested. All this is done through a chat interface here in the second point. So basically, maintenance managers can chat with this agent through natural language and really get immediate responses in terms of what questions they may have. Also, of course, the entire time the underlying models for this project can be retrained and thereby continuously improving actually the accuracy of the outputs. So what was our motivation for creating this agent? Essentially, it's the classical challenge you see with traditional maintenance. I'm sure most of you are familiar with these. So having equipment downtime, that's of course a super critical topic. This typically costs in manufacturing intensive industries tens of thousands per hour easily and if you're not able to predict these unplanned events and you can't prevent them, thereby, they tend to occur. Then the topic of this whole challenge with coordinating all these different resources that are involved. So, you of course have technicians that are included, you have to know what spare parts are needed, you have to find the right production window, you know what bottlenecks and equipment dependencies there are. So very complex process to solve manually. There's of course the topic of knowledge dependency risk. So you often, especially for older equipment, have only very experienced and few technicians who may know how to do the proper maintenance there and when these aren't available, that's an immediate bottleneck. And then also this topic of having multiple priorities. So maintenance is a topic that involves a lot of different departments. So for example, the production teams, of course, want to maximize the uptime of the of the production. All the maintenance teams want to have the maximum service window to be able to do the maintenance successfully. Finance, of course, as always, wants to keep control, keeps the cost low and, it's difficult to have a systematic way to to really address all these needs. This is also one area, for example, where classical predictive maintenance doesn't take this into consideration where we found that with an AI agent approach, it's much, better suited to take these different perspectives into account as you'll see in a moment. So with the actual implemented solution, we expect quite a few benefits. So of course, reducing the time you actually need to plan the maintenance since this is done automatically by the agent. And this of course, if actually implemented prevents unplanned downtime which has immediate benefits but also, as in the third point here, you're able to cut a lot of the underlying costs so that kind of follow through with having unplanned maintenance. So you're better able to allocate your resources, probably reduce your spare parts inventory, reduce your holding cost there and reduce the number of emergency repairs you're doing which are always costly. So a lot of these costs can be reduced then also maximizing the resources that you're utilizing and also ensuring more consistency with how you're reaching these decisions. The agents are always using a certain databases as we'll show shortly. This really helps to make the decision process transparent and also repeatable. And the benefit of Dataiku is that it's very easy to very clearly set up your agents so you know exactly what databases they're using to make their decision and make their, recommendations and, can get the full transparency and also, consistency. And if there's any errors that are occurring or unexplainable results, there's many ways to to really trace back how the agent arrived at that decision. So, as last slide before we jump to the demo, here's like a high level view kind of how we've set up the project. So what what does this agent use case actually consist of? So typical starting at the top, typical persona could be like a factory manager doing a maintenance plan for the month, but also a shift manager wanting to kinda prioritize maintenance activities for the next day. Essentially, it all begins with through a chat interface that's native to Dataiku or they can ask a question to the main agent which we call the scheduling manager agent. And as you can see here, we've essentially taken a multi agent approach. We have a system here where this, main central agent coordinates, the activities of different agents that are allocated to specific domains that are specialized for specific domains. So we see for example, floor manager agent here that, is used to get information about, the past performance of the equipment, observations the operators have have made and any settings that the equipment might have. And it basically has this perspective of, someone working on the factory floor. Then we have the reliability engineer agent that gets the past reports of reliability for each machine, production manager agent that has the schedule in mind and knows, what the the order is of the processing steps so we can identify which machines might be bottleneck machines. And then the finally, of course, the maintenance manager agent which at the bottom you see here, additional to taking datasets as inputs. It also takes, for example, a prediction model, remaining useful life model as a tool at its disposable. So we can then, for each given machine, get an estimate for the remaining useful life based on the current operating conditions it has. As well as of course maintenance documentation in terms of the steps to do. Essentially, all these different agents contribute and give their feedback to the respective domain to the central agent which can then use this information to get a response in terms of what maintenance scheduling questions to answer and suggest a maintenance schedule depending on what was prompted. So now that you've seen a high level overview of kind of how it's all working, I'll hand over to my colleague Nicolas to go into the live demo. Over to you, Nicolas. Hello, everybody. So I'm, I'm Nicolas, and I'm, I'm senior data scientist. I I built the technical aspect of this solution along with Nari. I see that we have already a very interesting question from Carl, Hiroshi, and Deva. So I've already answered Carl, but I I tried to illustrate it during the demo. And, Hiroshi, and Deva, I'll try to answer your your question at the end of the demo. If you have any other question during the demo, feel free to send them either in the chat or or the journey. So let me share my screen. So I'm gonna share my entire screen. There you go. So, basically, this is a traditional Dataiku, DSS project. And, with those agents, what we wanted to achieve is to to be able to integrate them in your traditional project. So a project in in DSS is a flow of data transformations. So you can see basically that everything is a dataset, and then you have those really simple recipes that enable you to to transform it your data, to prepare them, to clean them, and to do some more advanced things such as machine learning, or statistical models. Agents in DSS follow the same pattern. You now have new object in DSS that are agents and are are really easy to set up. In this demo, what I am going to show you is how to, easily build a multi agent solution that will tackle a complex issue, which is how do you manage your maintenance, schedule, while using a solution that enables you to to do it entirely using, no code, no code building blocks into your solution. So an agent is simply just an object where you can just double click on it, manage the version of the agent if you want to update it, and just simply tell him two things. The first one is the prompt. So I'm going to use a scheduling assistant. So this one is the scheduling assistant, and I'm I'm telling you him very simply that he has to help me, schedule my amendments. And he can to do so, manage these tools. So the the tools are the the second, very important aspect along with the context, which is, what the the agent can do, in terms of actions. So to to answer Karl question in the chat, what's what's the difference between that and the traditional analytics solution? The agent can choose depending on on what's the task and what the question you're asking him, which tool he can use, and in which order depending on the answer of the tool. For instance, he has access to the other agents as a tool. So if I go back to the flow, I can see that there are links between the scheduling manager and a floor manager that manage operators observations. So each time someone on the floor notice something, he can write it into an Excel spreadsheet that will be used as an input in my project. So I get I get a simple spreadsheet that has been inputted in my, in my project where I have some nonstructured data with simply observation from from operators such as, you know, it seems slightly louder or there are some stories or things that we might want to focus our attention on. Things that traditionally have been very complex to analyze, but now thanks to agents and to language model, we can analyze those kind of nonstructured data. So this agent has access to, to this type of information. We also have all the other aspect that are at stake when we decide to schedule maintenance, such as the managing the production. So the production manager has access to several things. The the first one is is the data set of the forthcoming orders to know that, you know, for for the next month, I have to fulfill this order. So I will need to produce this and this, type of product. You also have a maintenance manager that has access to a little bit more things because, obviously, maintenance is at the core of this solution. So the maintenance manager has access to several tools. So let's double click on it to see exactly which one it has access to. So it can read the statistical analysis that we have into the project. More on that, later. It can search maintenance documents. So, this is especially interesting. It's a it's a folder where you can drag and drop a lot of documents, and it has the the capacity to to search those documents that have been indexed. But, I'm going to show to show it to you just quickly after that. And then you can read traditional dataset that you have. So if I go back to the solution, the entire solution that we have without agents, is a machine learning, model that try and estimate the remaining useful life of your equipment. So if you know maintenance, you know that's a traditional subject in, in predictive maintenance, estimating when the equipment is going to fail, what is the remaining useful life of equipment. That's what, this machine learning, part of the the flow is doing. And you get the result to the statistical analysis, which is what, variable might impact my maintenance performance, such as the manufacturer or such as the the the location of the equipment, and then what is the estimation in terms of remaining useful life. For instance, for certain number of equipment, I only have, you know, nine days of remaining useful life estimated before needing a proper maintenance. So the maintenance manager has access to all those things, and then you get the reliability engineer we, who is more focused on, on, is the machine particularly critical for a given process. Meaning that if I if I if this machine is off for a certain period of time, I will be in trouble because I don't have any, machine able to replace it to produce a given product. So those are a lot of elements you have to take into account, when trying to schedule your maintenance. So why not using a traditional method here? It's because it really depends on the result of all of those processes. It's for instance, I'm, I'm interested in, in, in running maintenance of a given, machine that has a remaining useful life, which is pretty low. I would like to check if, if the if there are any orders that will come in and that we need this machine or not. So I I would need to ask the policy manager. And depending on its answer, I will need to to check all the tools. So once you when you need to, to adapt the entire workflow of the process, that's where agents shine because they are able to adapt and choose to decide which tool to use depending on the result of previous tools. So that's that's what they all do. And in terms of deliverable of this kind of demo, it integrates well with the traditional way we deliver results in the in the DSS, project, which is dashboard. So you get a generic dashboard, where the first the first tab is simply, is simply what you would get in a traditional project. So you can you can translate all the result of your statistical analysis in terms of KPIs or charts, with filters and anything like that to have, at the first glance, an estimate of all the the KPIs of your maintenance. And then you can integrate your your agents, in several ways. The first one and the classical one is simply, a chatbot. So you can, you can ask question. You can you can have a dialogue with it, and it will adapt depending on your question. So I can show you an example of a discussion I I had with with him. So you can just ask him what should I focus my attention on in terms of maintenance, and the agent will will just, confer with all the other, all the other agents. They will discuss in between them, so it can take a while and then, go back to to an answer. So here, what happened is there is a specific equipment that came as a priority for several agents. So the he has only nine days of remaining useful life. So that's the maintenance manager speaking to the scheduling assistant and telling him, okay. This one should be a high priority, but it's also and that comes from the reliability engineer. It's a critical piece of equipment, with no substitute for one specific product. So we have two factors at the same times coming from two different aspect of the, you know, the reflection. And here in this architecture, it comes from two different agents. The maintenance manager telling him that the statistical and machine learning model tells us that, this, piece of equipment will fade with a 80% probability nine days, and then you get the reliability engineer that tells you, okay. This machine, if it's off, we have no substitute to, to produce a particular product. So this, the the the scheduling agent come on top of that, aggregate this information, and is able to prioritize and to summarize it for you to make a decision. So it just tells you that needs your immediate attention. That's the most urgent. If you have one thing to to know about it, it's that this machine needs, immediate amendments. And then you can dip more into what you could do later, which is, you know, secondary focus. We have a couple of other equipments, etcetera. And what is really interesting with those kind of agent is that they are able to, to confer with, you you can organize a multi agent architecture pretty easily. So you can all they can all specialize in a specific aspect, you know, criticality of a component, maintenance, remaining useful life, those kind of things. But it can also, uncover deeply nested information that, you know, a human could just go over because there are there is there is so many information to, to take into account. Right? So that's this example, for instance, the the a 87 is showing performance regulation and mean of scoring. So, if you have a look at the actual flow, you can see that in one in the spreadsheet where you had the operator's observation, if you have a look at the, at this line, the a 87, you can see that one operator noted mean of scoring on this machine. And that came up because, obviously, the agents is able to scan an entire, spreadsheet and is able to, to, to to to uncover deep nested information, that a human would have trouble doing. You know? If this the spreadsheet is necessarily large, that's something you can, you know, go over and and and not see. And then another ask good aspect, that might answer, Carl question as well, is that it's able to adapt to what you it's able to interact with you and to adapt to what you what's your main focus. So I just asked him, what if I want to focus on OS updates or operating system? So on some equipment, you can have an operating system just the way you you have on a on a on your computer, and you need updates and things like that. And, what we have here is the agent just, you know, going back to the scheduling agent, going back to the specified agent and manage to uncover that you have you have something you you should be focusing on, which is, you have one, you have one OS version, which is, which is pretty problematic, which is called the version 2.1 Vanguard operating system. And how did it manage to to find that? Because in the the document I have so in this folder, I have all the context documents that could be useful. And, I have, you know, the manual you could get about the machine, which is simply a PDF. And what you get with those, what you get with those let me zoom a bit so you can see maybe a bit more. And maybe I can remove that as well. No. I cannot. So you have to trust me on this. But, what you you could have on those type of manual is the software version history where you have, you know, the reason why you we decided to to create a new version on a given software. And for the version 2.1 Vanguard, what you would get is, non bug significant issue with calibration. So you you get some bugs. And as long, along with this particular element in a PDF file somewhere, you get into the machine learning results. You get the fact that the the result have shown the, the the the the the statistical analysis have shown that this specific element, the version 2.1 Vanguard, is at risk of, needing higher maintenance. So by asking by by focusing on that, the the agent is able to go into the the documents, looking for information about operating system. It's going into the statistical result, that show that this version in particular, shows some some issues. And so it answers you that if you're interested in OS version, then you should definitely focus your attention on this specific version. And then it updates your, your priority list by telling you, be careful. Those machines are running version 2.1, OS. So now that I know that operating system, updates are a priority for you, you should focus your attention, as long as the first one, you know, that, we have already identified. You should focus your attention as well on those two items because they are running a problematic operating system version. So that's another good aspect of agents that, with all this information that you'd you can have in the spreadsheet and the datasets and the statistical result, but, also, as I just shown you, the folder with PDF, with any types of text document. Yeah. If you, if you have a specific priority or something you have in mind, you can interact with the agent and the schedule, the the maintenance schedule will adapt live, depending on all this information and with your priorities. So that's the traditional version the traditional way of interacting with those agents using the chatbot, and I'm sure by now we are all used to interact with language model through chat through chatbot. But what we wanted to, to show as well is another possibility in DSS to integrate agent, which is, through web application. So web application are simply, simply one possibility in, in DSS. You have two kinds of web application. So in this project, we have two. The first one is the one I I'm just showing you, which is the a chatbot. So that's what we call visual web application. So our chatbot is one visual web application, meaning that it's out of the box, it's running. You you can just click on new web app. And on visual web app, you will have access to a wall, wall catalog of visual application, but don't need any coding to to run. So you can create a new chatbot and decide to focus on a specific aspect. For instance, if I wanted to have something dedicated to the reliability engineer, all I would have to do is click on on, on, answers, which is the name of our chatbot, and to connect it to our reliability agent. In a couple of clicks, you could have a dedicated, visual application, which is a chatbot with your reliability engineer. I could add a specific prompt to to, to give him some specific instruction, and I could share that with my reliability team. Right? But another aspect, if you want to go a bit further and if you happen to have development team, development teams within your organization, I'm sure that's the case for a lot of, lot of you today. You have development software development teams that work on specific application, and you you, you want to integrate them into DSS. That's a possibility. So you can just have an actual application. So here, that's one I've built myself, which is simply a calendar application. So that's a calendar application where you can just click and write manually that you want to run maintenance on things, like, instance maintenance upon the, 97. I'm just writing random thing just to show you. So that's that's a calendar application. So let's see. Let's imagine I have a a software I have an internal project who has built, this kind of application. What I can do now is I can have it in DSS, connect it to my actual data, and more than that, I can connect it to the agent I've just built in order to, to to improve the the quality of my application and to put some AI in it. So what we, just did this one. So what we have here is simply one button where I can just interact with my agent, which, obviously, it's visually different, but it's exactly the same thing as the chatbot we have seen. And I can just tell him specifically, please schedule, for instance, the rest of the week, you know, based on the the priority that you you, you estimate are the the the highest priority in terms of maintenance. And what you would get as a result is exactly that. So you that's the actual result of the agent. So you would get automatically, automatically, you will get something scheduled automatically with the actual explanation. So if we find the the the same result we had before, so the b zero two seventy two is critical because there is, a remaining useful life of nine days and no available substitute substitute. And then as I ask him to do the rest of the week in terms of priority, now we are at the next one, which is also nine days remaining useful life, etcetera. So that's the second the second big way of integrating those agents that it's really easy to integrate it into your web application. From a code perspective, I can show you just a bit of code really quickly just to show you it's really simple. We have made a lot of effort to to make it as easy as possible. The level of abstraction of those agents is really simple. So if you have already a classical web application, integrating those is just a matter of, of, several line of code. Let me just show you. So it's in the code, the component, which is this symbol of your DSS project. Then you have your libraries. And I've put my my project here with the the code of the back end, which is which is doing all the heavy lifting here. And then just to show you, it's just simply this function. So chat with my, my my chatbot. And it's basically a dozen of line of code to just pass what the the user is writing to the actual model. So it's not rocket science. It's working pretty easily because DSS is doing much of the heavy lift heavy lifting in terms of, agents. So that's basically it for the the demo of the management scheduling agent. Let me just go back to the to the to the question I've been, asking in the in the flow. If I remember correctly, I have, Hiroshi who asked about, an agent identifying trends. So it's not really the subject of, of this demo. It's it's it's a it's a great question. It's a fascinating subject. It's more into the domain of, anomaly detection, and we are actually working on it. So we have another solution, which is called manufacturing event tracker, who really, try to to tackle this issue. But you're right on this error sheet. It's a bit complex for an agent to to, by itself, detect thresholds, and trends. So what we decide to do here at the at DSS is, is, to to pro to provide traditional the anomaly detection. And then the agent part is not really leveraged for those kind of solution. We are more into into the traditional, artificial intelligence solutions such as you label by yourself what is, not normal on on your trends, and then after that, you do a machine learning, lab. But, yeah, even with the visual capabilities of new agents, detecting automatic retrant and, and anomalies remain a challenge for for agents. Let me just go back to the the questions. So it's going to be a bit, so the question I maybe I'm going I'm going to stop sharing. It's a bit easier, and we can take maybe you can take, we can start taking question or do you think, Henry? Yep. Sounds good. I'm also, furiously answering questions in the chat. But Right. Yeah, I would say we just Okay. Yeah. Start at the top and go down. I remember that. We had a great question by Deborah about hallucinations. So that's remained a big challenge for for agent for language model today. We have seen some great improvement. I'm sure you you do also on Chargebee or mainstream language model. Now there are some, lot of loops in those, highly advanced products to make sure they don't hallucinate. In this solution, at at first, it could be an issue, like scheduling maintenance on a on a machine that does not exist. So here you you have two solutions. The the first one is, we have a recipe. So building block that, that is called evaluate your language model. So what you would do is you would create a dataset such as the one we have seen. So you create it manually with, input and what you expect. For instance, what's, in terms of maintenance, that would be what's the remaining useful life of. What's the the machine with the lowest remaining useful life. And you know it's machine a 9025%. You know for sure the answer. So you you write exactly the answer you expect. And this data set is like a golden data set that we evaluate your your agent on. So once you you build your multi agent architecture, you make sure there is no hallucination by testing it on the, on the evaluate URL and dataset. And as soon as something is off, like, answer something that doesn't exist, you get a warning signal. Or, if it works properly and there is no hallucination of this entire dataset, you know that you have lower the probability of having hallucination with this given architecture. Obviously, you cannot eliminate it, but you can lower it by testing the the the LMM. And another solution is what we call a reflection pattern, meaning that you can let me show you basically what it looks like. I'm going to show to share my screen again. So you can have, you know, you can have a lot of architecture that you that you can build. This one is a project not dedicated to to manufacturing. It's more general. But you can see that for instance, this one, you could have, like, for instance, three or four agents checking the results of the previous agent to make sure it exists. Because what happens usually is that one agent can hallucinate, but not not three in a row. Right? Because they don't hallucinate in the same manner. So if you ask a first agent to schedule the maintenance and then if you ask a second agent to check if those machine exist, then you basically eliminate the risk of any hallucination at the cost, obviously, of having to run the the the the thing twice. So in terms of, API, cost with your language model provider, it's it's a bit more expensive, but you basically can eliminate the probability of having hallucinations. Right. So I I, sorted the questions by upvotes, which is a nice functionality I'm just seeing. The top voted question, is from Hiroshi Ochi, asking how do the AI AI agents get the information? Number one, the input from the equipment, so for example, noise readings or or machine speeds and stuff. And then number two, know, what the threshold is to set a trigger action Yeah. Then. And maybe the yep. I guess the quick answer would be that for our specific case, correct me if I'm wrong, Nicolas, it's we're not doing, like, anomaly detection, which is, of course, the the classical predictive maintenance, use case. In this case, we're mainly focusing on, analyzing the historical maintenance data. So we're not just doing simple triggers. There was also one or two other questions on this. It's not just simply looking at certain, thresholds and then setting up a trigger instead. It's looking at the historical performance and then, giving a recommendation, when is it likely gonna fail based on how how often maintenance has been needed in the past. So it's a bit of a different, maintenance use case as, you're describing Hiroshi. Yeah. And in terms of information, you have a lot of possibilities. We have seen that asset can be a a possibility, and you can plug an agent on it. We have seen also that you can have contact documents. So you have a folder, and you drag and drop your document, your PDF files in it, and that can be an input for that asset. And then we have seen also a third way, which is another agent can bring information to this specific agent. And all of those all of those can be scheduled, using, the scenario aspect of DSS. So we have a section called scenario where you decide what triggers what. So you can say, for instance, as soon as someone add a document into my my folder, I want to trigger one agent to actually read this document and answer a specific question and output the result there. So there are a lot of possibilities, but, basically, that's, you know, the main the main possibilities are are there. Exactly. And, you could even probably implement, a streaming approach that data that could have a has a streaming plug in for Kafka. Exactly. And we have a do the live anomaly monitoring. That's what you could use for for what you're describing. But but in this case, we're we're not doing that. But great question. So another question here at the second most top voted is, kind of a challenge of, do you actually need agents like this? So, Arnab Das is arguing that, it's basically what we're describing as a the complex way of replacing fifteen minute, stand up at the beginning of a shift. That's that's exactly yeah. But the idea is maybe fifteen minutes of a stand up. You you can you can compute basically the cost of a fifteen minutes. Yeah. Yeah. I could argue that for this specific maintenance, aspect with all the documents, it might be more than fifteen minutes. But let's imagine it's fifteen minutes. You can compute what fifteen minutes in terms of preparation for the people to come to the stand up, to to be ready for those things, reading the documents before. And the total time, cumulative times, the salary of all those brilliant people that I'm sure are are well compensated. In terms of cost, I'm not sure that would be the actual cost we have for this specific solution, which is like a couple of cents each time there is one request on the on the language model. So the idea is that in terms of cost, it's pretty hard to beat, you know, even if it's fifteen minutes plus the the preparation. And the idea is, as Anna mentioned, it's not to replace people or it's not the idea of replacing the standard with that. It's more idea the idea of having the computer there and helping people at the standard make the best decision possible. So you still have your startup, but this time you have the result of the DSS flow and the suggestion by the agent, and we can discuss it. Maybe that's not the perfect approach. Maybe we can we can leverage what the agent has has told us. But, you know, yeah, the idea is to make this fifteen minute stand up more valuable than than it was before than before. Yeah. Man, I mean, I spent, multiple years going from company to company doing lead transformations and then joining these stand ups like this. And, yeah, having one that efficiently runs in fifteen minutes and identifies the ideal maintenance prioritization, without, any arguing or over overseeing some some details, that's actually quite hard to achieve. So, I don't think not all companies are necessarily possible. I'd have the the ability to do that that effectively. But, yeah, in the end, it's it's more of a tool to support this. Okay. But, it's definitely very important when you're thinking about the topic of agents to to remain critical and really ask yourself, okay, is this a really a meaningful use case, where it actually adds value or is this something that, might not be that, have not that tangible value? And then you should, of course, think about, maybe in a different, use case, especially if you're kind of in the a POC phase at your company and you're trying to still prove the valuable, the value to your management. It's important to to remain critical. So it's definitely a good thing to to ask questions like that. Scrolling down. So, I've answered some of these questions. I'm sorry, Annari. I'll let you manage the question. Yeah. No problem. So I think in terms of what we had another question. What's the business model, the value proposition of the tool? I think you you you kinda somewhat this nicely, Nicolas, in terms of Yeah. No idea. Yeah. I wanted to manage leverage, gain some time. There is an economic value to it because you can you can replace task that exists before will be done quicker. You can, you can, you can increase the performance of doing the same task by having those assistant along with you. And, also, something that we don't think about is it automates thing that can be a bit, you know, boring for for your colleagues such as, you know, reading your report or consulting the statistical analysis. So the idea is to make their job more interesting by automating the boring task, right, which is what it is about. So that's the the world value proposition, you know, cost, performance, but also, you know, improve the quality of life of of your your colleagues. Yeah. Okay. Then we had, two questions, most definitely to to Dataiku. One from, Shivaram Yamuna, asking what kind of factories or production shops are using Dataiku. So definitely check out our customer stories on the link that, Marisha shared. But in general, we're across all industries that you can imagine, basically. Basically, yeah, energy, automotive, aviation, semiconductor were completely basically industry agnostic and these specific applications that Nicolas and myself develop may have more relevance. Some may have more relevance to certain industries but in general, we kinda try to really target, business needs that are, found across any industry such as, for example, maintenance. Then we had a question around the timeline. So, someone's asking what's the estimated project timeline developing such a system end to end. I mean, the answer is as always it depends, but, maybe, Nicolas, you can, give kind of a explanation of what the factors could be to to determine that. Yeah. In terms of your technical aspect, I can tell you how long it took me. Took me, about a month, half time because I have other project to to to do at the same time. So for this initiative, yeah, I had maybe a third or half of my time on it, and I I've spent a a month on it. Obviously, I'm used to the platform, but I'm no particular I'm not particularly quick or anything like that. Any data scientist could do the same, and even people are not that sensitive because most of the things I use were not coding aspect. Obviously, the web app that you have seen with the calendar, there is a bit of code involved. But for the rest of the solution, there is no code involved, so that would be way faster than than a month to actually do it. Most of my work was, like, yeah, try to make the the calendar, you know, look nice and and and interact, well with the AI, but all the other aspect are pretty quick to use. It's more I mean, it's it's more about knowing, which, data you have and connecting to your existing data source, that's the bulk of the work. But knowing your data, you already know it. You already have the expertise. And connecting it to it, that's, you know, the the primary the the the primary work you have to do that that takes a lot of time. Actually, using DSS once the data is there is pretty, quick and pretty straightforward, and that's our value proposition, which is to make to to be able to to build such a solution quickly. Why quickly? Because you will need to iterate before finding the the architecture that works for you and the problem that works for you. So we we try to make this as smooth as possible, and that's pretty simple to to build really. So yeah. Perfect. And then, of course, the big determinant is, like, is your data already you know where your data is? Do you know what data you need? The the nice thing about these, preconfigured, templates we make, which we which we call business solutions, is that they're, like, the the flow and the the back end is more or less, already there and you just need to adapt it to to your own needs. So, yeah. In the end, it comes comes kinda down to to how much adaption you need to do if you're asking about the specific use case. But in general, data like who I mean, across our customers, we definitely typically see an acceleration of, yeah, more than than 50%. Some customers improving their their speed to for ML projects, threefold. So it's, it definitely, results in a lot of acceleration. And and one thing to mention is, both Henry and I, we work in a division called business solutions, and we build those kind of solution ready to use. So if you are if you use the the I can show you. If you use the DSS, you get an option. So let's imagine I'm in DSS on the home page. I can either create a new project or create a Dataiku solution. And there you have a whole range of, of, of domains such as manufacturing and a whole range of solution that you can implement quickly to accelerate, your project at the at the beginning. So to make it even faster to leverage our own expertise because, I'm specialized in, in manufacturing, and we have also data scientist and, subject matter experts such as and I specialize in manufacturing, but also in a lot of different subjects. So that could accelerate your projects as well. And after that, we also have a service solution that can help you, either to build to build with you or entirely for you, solutions, as a service. Exactly. Okay. Then we had a bit more we stopped some time, so we have a bit more detailed questions kind of on, how the agents are working. So, for example, one question is, once the updates are done as pointed out by the agent, where is the update log store? Is the PDF updated or some other location? So so in this case, there's not really like a if you're talking about the documentation of how the agent, reached its decision, That's, on a separate area on Dataiku where you can kind of we call it traces where you can see kind of what's the steps that the agent took. Otherwise, here, you have the chat history of of any questions you ask the agent. The chat history is saved. So you can just scroll up and see, what the answers were from the agent as well as what its reasoning was. I hope that answers that question. Then we had a, a question. Someone noticed that there's a folder located throughout the flows, and they're asking if this is to specify information to be reviewed at the specific step or if it's a rag style retrieval at each, step. Yeah. That's a rag style, yeah. And you can you can do it you can do whatever you want with it. So I I can show you more in detail. So that's right. Yeah. What we have so let me just go back to the flow. So that's a folder with all my documents. I can drag and drop or add a file from my computer. You know? So that's my computer. I won't add anything, but you get the idea. With several types of documents, you can also put in, Excel, spreadsheet at the CSV folder. At the CSV format, it works well. And then you can either you can either have, you you can you have a recipe to create, what we call a knowledge bank, but that's that's right, basically. And then you can have, language, you can have, language model recipes, or you can, so language model recipe is simply one shot prompt that you want to run. So I have one, single shot prompt that I I will be able to build anytime I want or I can schedule, but that's that's one shot. Right? Or I can just, you know, create a tool. So I have a section with tools, which are, there, agent tools. And, you know, search maintenance documents, so that's that's my title. You know, that's how I decided to call it. But, basically, it's connecting it to the knowledge bank I just created. So I create the folder with the embed recipe. I create the knowledge bank. And now after that, I have a tools that is basically do some rag on those documents. So search this knowledge bank. And then in the agents, the maintenance manager has access to these specific tools. So that's the one, search maintenance document. So you can decide which agent can do some rag and which, agent cannot. And you can you can have several folders with you know, you can you can do it pretty easily. It's simply a matter of clicking to create the recipe and then drag and dropping the documents, creating the tools, and giving the tools to a specific agent. Perfect. Looking at some more recent questions. What's the difference between Dataiku platform and Microsoft Copilot Studio? Yeah. Good question. I'll I'll try to keep it short. And, also, of course, depends on, exactly which which functionalities you're prioritizing. But, essentially, basically, Dataiku, as mentioned at the beginning, we see ourselves as the universal AI platform. So if you need, like, a unified collaborative platform for the full data science workflow from beginning to end, from data prep through to deployment, especially, like, in an enterprise or team setting, if you want the no code functionalities, that's really the strength of Dataiku. Copilot is specifically designed for a coding assistant. So essentially, Dataiku has a wider scope. What would be the the shortest answer to that in terms of the the main differences. The application or project built using data agents is scalable and supports concurrency. So now that question is a great question going towards the actual deployment. Yes. No. Yeah. The entire platform is, is around this paradigm of creating a job, having the job that runs. I I could show it. I I am I'm going to show it. The entire platform is built around this idea of when I'm in the flow. So let's say, for instance, I want to to do something such as prepare SAP. I can click on it and click on build. I'm not going to do it now, but, if I click on run on this one, I have a section called jobs. So this one, and those are all the jobs. So they either fail if I am doing something wrong or they are validated. And all those jobs are scheduled and are are running concurrently on your instance. So it it manage concurrency the same way, other platform are doing, you know, such as Databricks or Snowflake or anything like that. You can create jobs, and they just run concurrently without an issue. That's the entire platform. And that was built way before the agent came, and so the the agents are just, you know, piggybacking this architecture. So those calls are done concurrently. And it's scalable. Yeah. And it's on on your instance. But, you know, API calls, because those in the end are API calls to a language model provider without it is external or internal because we can also host our own language model. They are pretty light most of the time. So in terms of scalability, it's it's entirely fine, and, it entirely depends on the size of your instance. If you have, would if it's if it's big instance, you will be able to do more jobs, and we support, obviously, orchestration tools such as Kubernetes. So any project can be either on Kubernetes or just on the on your instance. And depending on the size of of the the the computer, the virtual machine, that you install the SS on, that will scale. Okay. We have, twenty seconds left. Maybe one more question. So commenting on which agents you chose for reasoning models and for why. So I think, to to keep it short here, so Dataiku is compatible with all the the typical LLMs that are available as Nicolas is showing. Essentially, yeah, it's it's typically a matter of, testing a few different ones. It depends on, of course, which ones you have you have access to at your company. But, Dataiku makes it very easy, to compare, to see how the results are for for each one. Yeah. You you you have your you can have a separate team, established connections. So it's it's hidden a bit from the, from the traditional user. In administration, you have connections. And then you have a lot of connection to if I click on new connection, we support, obviously, all the basic connection, but as well as the agent connection. So in terms of LN, we have almost everything. And then in your in your agent, I as I was showing you, it's just a matter of selecting the connection that has been created by your your ops team, for instance, or your IT team. And so after that, it's just a matter of, of the tip of tasks and the expertise of the data sciences. For instance, for for certain task, you can observe that the cloud formantropic is a bit better. For some others, you want something quick and a bit less expensive. So for o mini, from OpenAI should be better. Things like things like that. So I made some choice in this project based on on, you you know, expertise and experience. But those things can change really quickly. You can have new models coming up really quickly. So one strength of the platform is it's completely agnostic. You just need to create a new connection, and you can change the LMM model with one click. For instance, this one is all three mini. Let's imagine 3.7 just redefine what's, you know, edge performance is like. I just have to click on this one, and that's it. I click on save, and the project does change, and I adapted myself to a new landscape without being, you know, closed into a specific, vendor. Great. I'm being told it's, okay. So let's wrap up after after this question. I think we we were able to answer that one nicely. The last one, I guess, we could answer is, can the entire solution or workflow be exported as code? The short answer is no. You can do a lot of, a lot of exporting in terms of, of computation and things like that because, we are built on Python and on SQL. So you get some option to actually export things, on that. But once you get a little bit more advanced such as we manage, you know, knowledge banks and language model, anything like that, you cannot export everything as code just as is. But for each one of the step, we are pretty transparent about the tools that are used. You know, the the pack any package is used. So you can you could reproduce it, but the answer is about exporting in one click and having the entire code. No. The answer is no. Yep. And there are there's ways we offer the the ability to basically access the outputs via an API. So so some of the analysis results, you could you could possibly pull that way. Yeah. Regarding the question on MCP servers and our ability to connect, so we we can via code. We can we offer the ability, to connect, and I think and there's an in progress initiative on offering, like, a no code option for connecting to to MCP. Exactly. But it's currently already possible to answer that question. Yeah. It'd be I think we have a Python generic recipe that you can put you can, you can put everywhere. So if you can do it in Python code, then you can put a Python recipe in it somewhere. That's what I did. You can you have seen probably the logo of Python in somewhere in my flow. So you can if there isn't any no code solution out of the box, you can still write a couple of lines of code. So if it's possible in Python, it's also possible in in DSS. Great. So then I think we'll wrap up for today. First of all, thank you everyone for the the big engagement, all the other great questions that were asked. A lot of fun to answer these, and also for the, the big turnout. And it was a pleasure, from our side to to share our point of view today with you guys. We'll be sharing a recording of the session, where you can also then see the slides. And we'd be, of course, more than happy if you reach out to us, individually, to to have further discussions. At the end of the presentation, you can find, my contact info. So you'll be able to reach out to us, reach out specifically to me via that or otherwise through our general, contact options. So thank you again for everyone for joining, and hope to see you soon again. Thank you.