AI Orchestration for Smart Cities and the Enterprise with Robin Braun and Luke Norris - #755

I'd like to thank our friends at HPE for their support of the podcast and their sponsorship of today's episode. HPE helps organizations bring AI into production faster with its AI Factory portfolio of engineered solutions, delivered in partnership with NVIDIA. These turnkey private cloud systems give AI and IT teams the tools they need to innovate while simplifying operations and keeping data secure and under control. Through its Unleash AI program, HPE also connects customers with trusted AI partners and validated solutions to speed up adoption and results. To learn more, visit hpe.com slash AI. It is just my pet peeve, but I think the chatbot was almost the worst thing to happen to AI, like period. What we're actually finding in mass adoption, whether it's the town of Bale or it's Fortune 500, is back office automation. The normal enterprise looks so similar at the finance area, looks so similar at the HR area, looks so similar at procurement. And Genitive AI can just knock all of that down literally almost overnight. You get those ROIs and then you go on to the sexy new use cases. All right, everyone, welcome to another episode of the TwiML AI podcast. I am your host, Sam Charrington. Today, I'm joined by Robin Braun, VP of AI Business Development for Hybrid Cloud at HPE, and Luke Norris, co-founder and CEO at Kamiwaza. Robin's been leading HPE efforts to build out its Unleash AI ecosystem with the goal of enabling partners to take advantage of the company's infrastructure to deliver enterprise AI solutions. Luke's company is one of those partners. Kamiwaza offers an AI orchestration platform that connects enterprise data and systems with LLMs and agents. Before we get going, be sure to take a moment to hit that subscribe button wherever you're listening to today's show. Robin and Luke, welcome to the podcast. Thanks for having us. I'm really looking forward to digging into our conversation. You've both been working on some really interesting projects together, including a new Agentex Smart City deployment in Vail, Colorado, that we'll be discussing a little bit later on. Before we dive into the details, though, it feels like we're at a little bit of a moment where every company is trying to figure out how to connect AI more deeply into their operations. And the two of you have been right in the middle of this. I'm curious how you're both thinking about this wave of enterprise AI adoption right now. What's the mood and the conversations you're having? Luke, we'll let you jump in first. Well, thanks. And once again, excited to be here. So I think the beginning of the year, it was an AI mandate, AI washer, AI do anything. I think it was coming down from boards. It was coming down from external pressures. and now I think it's turned very sharply to AI ROI. You have to have this sort of return on investment. The investment dollars are there and everyone's excited to put them in, but you have to be able to show some actual tangible reason for it. You have to take that baby step and then you can take larger and larger and larger steps faster and faster. I think we're right at that inflection where people are getting that first baby step done. They're starting to actually see maybe that first ROI and they're willing to take the next large step. And Robin, how does that resonate with what you're seeing? I completely agree. I think that there were mandates. However, people also were looking at what can they do that actually had value, not just to do AI for AI's sake, but to be able to create something, you know, to be able to find kind of what is that right first use case that can help them propel forward, not only in that ROI portion, but also in process people learning. for the organization and how AI can potentially impact them. So I think to Luke's point, we're at an inflection point where some of that learning now has been done over the past year. And people are looking at how to hit the acceleration button, which I guess for AI is not actually meant to be a joke, but I guess is. It really is looking at how they can get started. and there's like all of these like fearful, you know, numbers out there of like 90% fail and all of this. But I actually don't think that's a problem. It's people need to be able to experiment. People need to be able to try out different things to find that right first use case. And the technology is changing so quickly that things that we may have failed out at two years ago or a year ago now, we can solve in completely different ways. And that's one of the things I think with AI, that the speed is surpassing what kind of the analysis is. Yeah, and digging into that failure, I'm curious what you're both seeing in terms of the key challenges that enterprises are facing when they're trying to bring these early use cases to bear. I think there's been, and I think it's changing, by the way, but I think there's been this rapid test cycle in the cloud. and what's great about that is low barrier to entry you can test a lot of tools and you can sort of get this idea of a concept but taking that from the cloud to actual production data production systems production scale has been a very big leap uh nearly impossible actually for some of these organizations once they realize uh where their data is at the state of the data etc i think um that has changed dramatically though in the next let's call it last six months models that progress to the point of building out ontology schemes and entity abstraction automagically they're also able to now visual models be able to look at literally handwritten documents and legacy files and interpret them and actually say okay this means that and I'm going to change it into this format and what that done is that's changed the paradigm from you don't really need to now worry about cleansing your data and changing your data and making your data ready for AI, now you just got to get AI to the data and they can actually start to do amazing things out of the gate. And are the key use cases that you're seeing really gain momentum early on? It's been chatbots and then adding enterprise data through RAG to chatbots and a lot of organizations are still working on those use cases. But what else are you seeing? it is just my pet peeve but i think the chatbot was almost the worst thing to happen to ai like period you you you have this amazing artificial intelligence that that literally has you know inherent capability of a phd in almost every vertical and you have this ability to process at thousands and ten thousands of tokens a second where the human can only do about 20 tokens a second and then we said all right the best way to interact with it is the slowest possible way that we could actually send data back and forth. And I'm sorry, it's just horrible. So off of my horse there, what we're actually finding mass adoption, whether it's the town of Bale or it's Fortune 500s, 100 plus year old companies, is back office automation. The normal enterprise looks so similar at the finance area, looks so similar at the HR area, looks so similar at procurement. And there are so many tasks that are just implemented for implementation sake, for compliance sake, and for process. And Genitive AI can just knock all of that down literally almost overnight. You get those ROIs and then you go on to the sexy and new use cases right from there. Yeah, it's really interesting that you put it like that. I published late last year, maybe early this year, a white paper called Rag Beyond the Chatbot. And the core argument is exactly what you're describing, that we're funneling everything through this chatbot, but the real win is integrating these tools into enterprise workflows more deeply. And we are definitely starting to see more of that. Robin, what are you seeing from a challenges and use cases perspective? I think very, very heavily aligned to everything we were just talking about. I think some of the challenges are that sometimes it's the technology, sometimes it's the data. Almost always there's something about the data, but it's also the process because once you've actually been able to do something back office and impactful in that way, you actually have to change the process that goes with it. And there's a fear of that change, that even though, you know, people are looking for that efficiency, like they didn't necessarily want to change to get that efficiency. They just wanted to somehow or another be magically more efficient. But I do think that like one of the things we started with, and I love like the town of Vail was like, well, this is kind of the unsexy stuff, but it was sexy to them, that they could, something like where working with Kamiwaza around 508 compliance, helping to make websites accessible for people who have disabilities by being able to introduce agentic remediation, essentially review and remediation of the websites, although we still keep the human in the middle, that something like that is incredibly impactful because it is something that's very manually intensive, fraught with with error because of the itty-bitty little regulations and all of their knock-on effects, which is something actually AI can do incredibly well and incredibly quickly versus what somebody manually doing it would be. And that's where that efficiency, that approach is new, but is something that is incredibly beneficial, whether you're a municipality or you're a Fortune 500 company, that type of thinking is really important. I do think one of the things that is also important is that, particularly with agentic and multi-agent agentic, that that is changing the game and what we can do from that back office automation incredibly rapidly so that people who were doing these assessments like a year ago and getting ready for a board meeting really would want to completely revisit what they had looked at because those assessments are out a date now. Luke, why don't you introduce us more deeply into the Vail project and what you're doing with them? Only about three, four months ago, Robin and I actually went up to Vail and we pitched this idea of the sort of smart agentic city to them. And the reason I love emphasizing that is just how far we've come, how quickly we've come. So almost immediate, there was skepticism for sure. But we asked them what their primary pain points were. And the one that was just right at the top, as Robin just mentioned, was that 508 one. And the planning that they were going to have to put into it, the hiring of people and the need to get compliant. And what was great about that one. And just to be clear, 508 regulations are about accessibility with regards to websites and other government systems? Is that the core problem there? Yeah, it's 508 is the federal level compliance. And then there are multiple stages of state level compliance and multiple state groupings at the state level compliance. And they've actually evolved as well. So there's quite an onerous piece onto it. And yes, it requires any institution that at the federal level that's getting federal funds, grants, et cetera, has to make their websites and their overall access accessible to people with everything from visual disabilities to motor neural disabilities to even auditory disabilities if your website's really auditory. The long story on that is someplace like Vail, if you can't access the state and the local city services, as a constituent, you're literally, you know, behind the sort of eight ball. You can't, you know, figure out when the largest free bus system in the country is going to be able to come by and pick you up. You can't figure out what the deed restrictions are on the property. These are all things we did for them. And you can't figure out what's sort of going on in the city, voting, just everything it takes to be a sort of citizen of the city. So they're all having to bring their websites up to compliance. And if you think of the nuance of the technology piece of this, it's every picture every pdf uh has to have an auto metadata tags put into it so it actually describes what's going on in there every graph in the website the contrast of the actual website the html to make sure that the background of the foreground are contrasted enough that a the machines can pick it up but the humans with disabilities can pick it up so it's a vast level of regulations and that was sitting sort of on top of them and that was the first one We were able to quickly turn one of our computer use agents right onto their website ingest in all of the compliance requirements and do a quick pass remediation of it where it actually said here all the problems that the website has with the current code Then we were able to use visual models large language models in conjunction again with a computer use model to actually rip open the PDFs interpret what those pictures are and then re-put all the metadata in there and then allow them to be uploaded. And that was the first pass. Then we did HTML code, remediation, et cetera, et cetera. When it's all said and done, it's a pretty significant level of remediation. on the technical remediation capabilities, we get a website pretty much into that 90% in the first pass. There's a lot of tests that go on that are more, you know, is the picture properly annotated? So you could annotate too much or you could not have enough in it. And that's still subjective. There's not like a hard guideline on those. So we just make recommendations around that because it's hard to, you know, effectively know. But on the technical ones, we can get very high levels in that remediation. And this is an agent chunking through, you know, millions and billions of tokens, literally depending on the size of the website. You can imagine, you know, a university system, say in California, might literally need tens of billions of tokens just to go through their webpages. Or where the city of Vail, I think, just from the demo in my mind's eye, I think was that like, you know, 1.5 or 1.6 million tokens for the first like 10 pages we did. So that's the long story on that. Once we were able to get them going on that, then the use cases just expanded. I'll let Robin talk about the one that looped in the other partners, but the one that I was viscerally excited about was the deed restriction. And I know how exciting that sounds, right? Woohoo! Let's talk housing. Exactly. But it's a lot of people doing a lot of lookup to answer simple questions and to just keep compliance with the large deed restrictions. the deed restrictions i live in a mountain town are very important to a mountain town because we have to open up a lot of the rental properties for long-term people to actually be able to rent and work in our great communities and these deed restrictions are really the way to facilitate that and these deed restrictions go back to the 60s and even 50s they're micro fish literally in a lot of cases so it's literally you know scanned pictures that are horrible there's barcodes maybe on some of them but there's no longer a system that actually reads the barcode and because it's been scanned, literally in a lot of cases, OCR doesn't even work because there's just, you know, words that are missing and they're smudged out, et cetera. This is where the power of visual language models and LLNs can come into play. They can actually infer the missing words. They can infer what the actual barcode means by running it through a check of standard barcode services. And we're able to take all of that data as it sits, put it into an ontology system, then reinterpret it all and actually make a live system for all the people that work in the city in the county to actually be able to respond very quickly to all the deed structures and restrictions. More importantly, we're going to be able to open that up. So now it can be a self-service. So people, when they're buying a net new home, could see if their house is eligible for a deed restriction. They could see the deed restrictions around them for property value and next use and so many other things. And this is that unsexy, sexy part where you take something that's incredibly complicated, incredibly hard, incredibly legacy, and now move that forward to minutes and seconds of answers, automated answers, and now a much better idea of what you're also looking for. That's the kind of stuff that just really gets me and my team going. I just love that kind of thing. And Rob and Luke mentioned that there are some additional projects with other partners. Can you talk a little bit about those? Absolutely. And the great thing is to get to Luke back to the, to call back to the beginning of the conversation. Once you've done all of that with that information, then you could then put a chatbot on top of it to make it easy to interface with. So take that, Luke. Yes, exactly. We'll call back there. But the great thing is working with the Unleash AI ecosystem. Obviously, we're excited about what we're doing with Kamiwaza, but we're also working with several other of our Unleash AI partners at Vail, in particular around vision use cases that around something like fire detection. It was one of the first ones that we've worked on. When you think about the risk in a mountain town, you know, talking with Russ Forrest, the town manager at Vail, you know, it's not a question of if or whether a fire is going to happen. It's a question of when and what is the environment and where is it happening. And that's just the reality of where we live and, you know, the dry nature, the humidity, the climate, all those things coming together. And we have fires. But the impact of the fire is greatly, it depends on how fast you can respond to it. How fast can it be detected? How fast can it be responded to? But obviously, fires don't all, you know, like the lightning strikes, but it's been snowing for five days versus lightning strikes. And it's been like a red flag day and high winds for five days, like two really different scenarios of that lightning strike. And so we're working with partners ProHawk and Vidio. ProHawk AI allows us in real time to be able to restore. We don't say enhanced because they don't use Gen AI, but it's to restore pixels so that you have a much better input image coming in from video. It happens in real time. It essentially happens in the stream of the video coming back for analysis, that they're able to essentially be able to get rid of snowflakes and fog and all of these things where they can really get rid of that particulate matter to be able to see much better at what your even night they're able to see through and really have a much clearer image that can then go into the video analytics of Vidya. And Vidya is able to help us with an immense array of different video analytics that can be run against those images to detect those key motions like, you know, and differentiate between, is that a fog bank? Is that a somebody's backyard barbecue? Or is that actually like a lightning strike? And there's now some trees on fire. And those have all very different responses. But what I love about what we've done is that we have the vision aspect to it, and that's typically what people think of with Smart City. Like when you think about, you know, Smart City, you think about these knocks, you think about a lot of IoT devices, you think about a lot of vision. But what I love by bringing that agentic characteristic to it is not only are we able to get that detection with the ProHawk, Vitio, kind of one-two punch from Vision AI, but then also being able to have her interface with Kamiwaza, where now Kamiwaza can provide additional context. You know, what are the conditions out right now? What is unique about the location that it's happened? You know, something that's downtown may be very different than, you know, something off in a separate culvert. Or was it actually from an accident and the truck that I've returned has a hazardous material sticker on it, you know, and what is that? So I think that you have not only the vision AI, but now paired with that agentic intelligence to give it that contextual understanding that allows them to make a much more intelligent response much more quickly. It's all things that people think about, but how can it all be presented in an easy-to-consume way immediately for the fire department and for the town management to be able to make those decisions. Like, it was really interesting. I was talking with Russell, and he was talking about, like, over the summer, they had had, you know, somebody called in about a potential fire detection or there was one potentially detected. They had to dispatch fire, you know, firefighters out into kind of the middle of the woods at like two in the morning, which is not a particularly safe time to be rummaging around in the woods, you know, trying to find a fire. And thankfully, there wasn't one. But it's something that is very much top of their minds as they're thinking about it. Because when we started this working together, they had a fairly large fire just 30 miles away. You know, this is such a reality on a day-to-day basis that anything that we can do to improve that improves their citizen lives and safety. And so they have cameras deployed in these forests already, or is that part of the effort here? Actually, that's one of the things that I love about all of this technology, is we're reusing the infrastructure that they have. So they do have some cameras deployed on the slopes already for various and sundry reasons. And as one might in Vail have cameras on slopes. The, but being able to reuse their infrastructure so that they haven't had to invest in other cameras or other from that part of the infrastructure, but looking at it and saying, you have the ability to see, but how can we help you see better? How can we add technology to what you're already doing that can allow it to be more impactful and to give you better data? So like Luke mentioned, the free bus system, largest in the U.S., amazing, really great work that they do, that that's available, but they all have cameras on them. So now talking through, rather than thinking about it as a bus, how do you think about it as a moving camera? And now what are all the things it can report back? Can it report back that pothole needs to be filled or that a light pole, you know, that a light's out or those type of things? because all of a sudden, when you start to think about the technology differently, then you start to think about different ways you can leverage the investment they've already made. Is that suggesting a, as opposed to kind of a top-down use case view of looking for opportunities, more of a data or data sources up perspective at looking at opportunities? Do you think that's something that folks should pursue more or just kind of another tool in the chest as you're building out an AI strategy? I think it's a little bit that we started with the use case of the fire detection. And then we started talking about what was the infrastructure of cameras that they had. And that it kind of, what I love about the conversation when you start to make it practical and tactical and start looking at what you're going to do is that, and then it goes to, but what if, but what next? You know, as we've done kind of phase one, now what can phase two look like? So I think it's a little organic, if that makes sense. I think you have to start top down with a business case that once again drives that ROI that we had that initial discussion in. But even back to the deed use case, the second we were able to vectorize all the deeds, actually make it so that they can now start to manipulate them and look at what that data starts to present use case after use case after use case just started presenting itself because now the business experts can be like wow could you also do this can you do this could you group those together can you tell us all the ones in this particular time period or this particular location can you show that on a map i mean it just keeps going and it's all of the business problems that were typically presented to them that they couldn't do because time or effort just isn't available it literally is what Kami Waza means it's that superhuman capability how do you now you get to that 10, 20, 30% you would just never be able to do and now agents can do it just at a blink of an eye yeah I really appreciate that take I feel like you know one of the things that we saw early on was and I think one of you referenced that earlier kind of this boardroom read the airplane magazine you know we're going to do AI now do this do that and that you know being you know in some cases disconnected from the reality in the enterprise. And that kind of is representative of like too high a top-down view. And then, you know, there's another challenge of, you know, you've got a hammer, just start looking for nails. And that's a little bit of a problem with bottom-up. But we're talking about now kind of this relationship between, you know, identifying legitimate use cases and pursuing that. But then in the process of doing that, you're identifying these data assets that you can build on top of and also using that to drive further ideation and iteration around use cases. you know so much of you know what we're doing around what we're doing with ai is based on the the data really it's all based on the the data and so that has to be a key part in how you think about use cases and opportunities and prioritization right i don't mean to sort of hit this hammer but it's incredibly important um data wasn't actionable to the average person until about a year ago you You needed a machine learning, you need a machine scientist, you needed all machine engineering, data engineering, you needed all these people to cleanse it, get in a particular format, and then write very robust code to actually execute and do something with it. Where now literally with generative AI you can get that data make it actionable to an actual business person And then all of the effort all of the problems all of the insights are literally there in front of them now to actually sort of unleash in the case of HPE here but actually start to open up those workflows. And that is what we see. That first use case is the hardest. The second one's easy. And then you're at 40 before you know it, just because the business now can get so much value and just keep moving it. That's such a faster pace than was ever available before in IT. Yeah, I think that's something that Vale kept mentioning over and over again over the last few weeks is just how fast this has been able to come together from our initial two-day workshop in August to announcing it at the end of October at GTCDC, that that was particularly for like a public sector project, that that was an incredibly fast timeline, but that we're now able to do things that quickly because of this different technology. And that's why I think when people are looking back at something they may have wanted to do or wanted to try to do, and if they went, oh, well, I tried to think about it last year and I couldn't do it, like it's always worth a new view because with the agentic technology, with the different ways we're approaching things, it's really allowing for some different solutions, such as like that 508 compliance. If we tried to do that two years ago, we never would have solved it this way, and it wouldn't have been nearly as elegant or nearly as cost affordable as it can be now. Luke, can you help us dig into the 508 compliance, and deed restriction use cases and talk a little bit about the role that Kamiwaza played and the technology that you're applying and helping bring those use cases to life? I mean, I'll go a little high level and we can always double click. But on the 508 one, if you think about it, we take a collection of agents and those agents are typically the execution code of LLMs and a VLM, so a visual language. We use a mix of models some that are small language models, the visual model, the most recent visual model from Quinn, the 3.5 VLM. It's just amazing. And then an actual, the OpenAI 120 model for tool use. And the reason we call it tool use is we're then able to give that OpenAI model the computer use agent so that you actually, in a simple GUI, and we actually call it an agent's head. You're able to put in the web interface you want and a bunch of criteria. And that then kicks off the OpenAI model to then use tool use, then use computer use to actually take over a browser and start executing the website as if a user was actually visually seeing it. And it's then using the visual model to also represent, does this picture actually have the metadata in it? And if not, it downloads it, then opens it back up, will then reinterpret what that picture was supposed to be in the context of that webpage, and then put the metadata back into it. and set it aside. By the time it's gone through that entire web page, looking at every PDF, every picture, every graph, et cetera, it has this corpus of data that it's put into a nice repository with a great summary. And what that does is then give a web admin, here's the summary of the compliance structure, here's what's auto-compliant, here's what was auto-remediated by the agent, and then you upload the files one by one. And that's the sort of human in the loop last step. It's the human sort of taking assertion that, yes, this was needed to actually achieve compliance this is the problem and then it's a real quick remediation in some cases just a couple clicks or a simple file upload so that's pretty much in a nutshell for the 508 but think about this it's doing it over hundreds and thousands of pages there's also internal pages it's doing it from the amount of token processing and like i was saying earlier is exponential this isn't something you typically plug a plug a public api key into this is something you're going to want nvidia cards We really lean heavily on the RTX 6000. It's sort of a workhorse for air-cooled servers. So these servers work great. The town of Vail, which actually has, I believe, a 100% renewable energy source in their data center. Is that right, Robert? Yeah. And they were able to get this server up and going and it just, you know, able to execute flawlessly. Secondarily, if you think about Kami Uwaza's particular piece, it's that orchestration of all of those agents, actually cataloging the data, keeping that data into context cache and allowing those agents to process with multiple models running simultaneously all on that particular server. And then of course the deed one is one of my favorites. What we focus on is connecting data and making data actionable to those agents. So then you can actually start to work business workflows and use cases. When we read all that data from the various data sources, like I said, it was everything from SharePoint to actually plug it into their microfiche capabilities so we could auto scan and auto retrieve all the microfiche to build the vectors, to interpret them, to also know where to go back to retrieve them when somebody needed it. We're also able to build a full ontology scheme on top of it. And the ontology scheme is really beautiful because it actually broke down all of the deeds, the deed restrictions, and all the business and workflows and use cases of them because we're able to use all of the great training material and processes that Vale had already had for their humans on board and to take that on. Once all of that came together in a collection of agents, like I said, the use cases are continuing to flow to today because they only take about 15 to 20 minutes and we can do another workflow or another use case because it's really just running an agent over the ontology that could retrieve the data and it's literally like magic. Can you dig a little bit deeper into the role of the ontology? Are you normalizing data across a variety of sources or are you identifying structure within a particular source of unstructured data? So it's the, I guess, latter. So we take all of the data. So all of the unstructured data, all the structured data, build context-aware vectors and build a metadata structure of it. and then build an ontology scheme on top of all of that. And the ontology scheme will truly understand all the entities. So each person, user, action, and then what the actual verbs and meaning from that data would actually be. And then we can take simple business processes, typically one-pagers to a paragraph of how you would execute. The system actually reads all of those, reapplies it to all of the data, builds all the inner workings and the webs of that ontology. And now you can actually start to get business flow right over it. I call it auto magic because it literally is done just by the LLMs processing all of the data, processing all the workflows and tying that together. And once you've done that, it's no longer an agent that just knows the meaning of the words. It's not an agent that knows the workflow and the actual data. And like I said, when those two things come together, I don't think people truly understand how powerful that is. It's literally the lingua franca of an entire town, of an entire industry is now known and being able to be executed against. You mentioned vectors. Are you building out a vector embedding system that's kind of in parallel to the ontology? Yeah, so the full sort of agentic rag would start with a recursive lookup of the metadata that's required to answer the request. Then that would flow through an ontology system to make sure that it understands what data then go pull. That data then gets represented from vectors. The vectors then tell it where to actually grab the data, then the data is inferred, and you get the action from an LLM. That whole process, which you can just link that seamlessly in code, is really what allows you to do very recursive feedback of editing of agents and getting outcomes. You mentioned that you're doing LLM-based entity recognition entity extraction. That capability in LLMs has been evolving very quickly. Can you talk a little bit about generally the success you're having there? Are there kind of hacks that you're having to apply to get that to work consistently? How consistently are you able to do it? How have you seen that evolve over time? I'd say it's gone from about 75% efficacy to about 90% to 99% efficacy just this year. Yeah, no, it's been pretty amazing. So first off, we have to make sure that the models are either pre-trained on the business logic or workflow, or we have to do reinforcement learning of the model to create the ontology of the business workflow and agency. Also, that's also where I came up with Lingua Franca a lot. If you tried to build an ontology of, say, a pharmaceutical company, they would use verbiage and words that the model wouldn't have enough of an understanding, and you would have to do post or training or reinforcement learning to get it going off of that. Sometimes known as fine-tuning, but fine tuning is its own other black magic and art, which we sort of push to the side. So once you have a model that understands the business workflow and language, then you can work on the individual business workflows. We typically tell business leaders, give us a paragraph of what you do for this action. Give us a whole page on what you do to actually have an outcome. We can link those then into the model, get all those into the CV cache. These models then run 24 hours a day, seven days a week across that data. That's critical because we can watch how it's being used and all ads, moves, and changes of the data. And for the first time in all of technology and humanity, you have this updated ontology system in real time. Now, once again, that takes millions and billions and trillions of tokens. That's something that once again, you're going to want your own hardware. We call it being untoken bound to actually process and to keep up with. But now you have an AI system that understands your data, understands every change of your data, understands how your data is being used, every way it's being used. And now you can continue that, You could even prompt it literally. What's the next best use of this data? What's the next best workflow I can attack? And it's actually smart enough to now start telling us that. And that was just a huge breakthrough when we saw that about three months ago. And so is that approach, the approach that you came up with for deed restrictions in particular? Or is this a general approach that you apply broadly as like a foundation? And then, you know, whenever you're tackling a problem, you're using this approach. or is it like evolving in that way? It's definitely evolving, but no, we pioneered this approach about 18 months ago with Fortune 500 companies. And that's the beauty of what I sort of kicked us off with. The deed restriction looks just like a contract or invoice processing for Fortune 500. They're literally kept in Iron Mountain boxes and they need to be scanned in and people need to understand their impacts. I mean, literally. And there's really no difference in that sort of scenario. And once you can get a business to understand that now that data is actionable, it's amazing how fast they want to just keep flooding the AI with net new data, with net new systems, because you just get the next output, the next value increase. It's so fun to be part of. Yeah, I hear from people all the time, like, why are we seeing new PDF extraction models coming out like every week, every three days? And the answer is like, there are a lot of PDFs. there. There's a whole bunch of stuff. Between Excel and PDF, you have about 90% of enterprise data. No way around it. Yeah. And throw in PowerPoint and you're good. Yeah. Yeah. There you are. And like the entity extraction problem that we've talked about, capabilities with regard to extracting the nuances of PDF structure and tables and charts and the like has been evolving very rapidly again especially this year yeah i mean just last week a new model got up to about 99 efficacy on multi-table pictures within the pdf so non-tag multi-table pictures within pdfs and it's now getting about a 99 efficacy of sort of interpreting those back into metadata structures that can be activated off of and that took it up from 94 but that little three to four percent difference is gigantic the other thing that i always love to point out especially many of the big HPE servers with the big NVIDIA cards, you can run these models over and over again on a process. You don't infer something once, you infer it 50 times. Then you can have an oversight model actually say, look, these three deviated don't make sense. These 50, you know, others were perfectly go with that. And you start to get these very, very high levels of processes now that, you know, frankly, far exceed even humans capabilities of reading hundreds of pages of documents and try to infer them and manipulate them back together. And that's the other thing is it's no longer this like one-shot prompt into a public cloud. It's your dedicated processes running as hard as it needs to to get you the answer at the level you need it at. It's an interesting idea to think about if you're not paying on a per token basis, how that changes the way you think about solving these problems. Literally being unbound, like I said. And you could have long running ones that run in the background at night and just take forever. But you know, you wake up the next morning and you have your entire day's work effectively done for you or now you just click through and say it's out. And there's a whole different way of sort of reinventing workflows, as Robin said, once you sort of have this ability to have true dedicated intelligence processing for you. And before we jump over to you, Robin, I had one more question for Luke, and that is the computer use agent that you use in the 508 use case. Was that like an open AI operator kind of model or your own model or what? Tell us a little bit about that model. So when I said OpenAI, that's the 120 OSS private model. So the open source one effectively or open weight one, you run that. What it's really great at is it's good at tool calling. So then it calls into our computer use model. Computer use model, I believe, was a post one that we had done, which I also believe was based off of a Quen backend as well. Between Gemma and Quen we get a lot of use out of those models And then like I said OpenAI 120 model has just been great for managing different MCP calls and tool calls which we colloquially all put into our tool shed And that just gives access to the individual models that make the calls. And it's a great little linkage that we have. Robin, I think hearing Luke talk about these use cases more deeply and in particular the way they're taking advantage of the fact that, you know, Vale's got its own infrastructure, its own servers, its own GPUs, probably sets the stage for kind of why your program exists and the vision for enabling partners like this. Can you talk a little bit more about that? Absolutely. And I think one of the struggles that people have had in being able to realize true success with AI is when they started in the cloud and they start to hit that token boundaries. and or they get the first bill and they're like, oh, maybe let's not do that. And that's really where we've leaned down from an HPE perspective, working with Luke and others to say, we want to bring that last mile. We want to bring these use cases. We want to bring these solutions. We want to help people truly understand what they can do and what they can do once to a really important point that they become kind of unbound by that token cost, you know, and that they can instead explore how would they use the AI? How would they repeatedly use the AI? You know, like in the case of 508 Compliant, you don't just run that once. You don't just get a website in this pristine condition and then leave it for time immemorial. You know, the next day, somebody from marketing or comms is gonna wanna put a new PDF on there and you're gonna need to remediate that again. You know, how do you start to truly have AI work for you? And I think by being able to have, like in the case of like private cloud for AI, have that private cloud experience, but have it within your data center, within your compute boundary, allows people to be able to start to take advantage of not just the promise of AI, but what we can deliver. because they're no longer going, ooh, I need to stop computing today or I need to not have that run over it again. And they're making choices that are appropriate for their pocketbook, but maybe not for the implementation of AI. So that's why we get really excited creating this ecosystem, building this out together and helping our customers really take advantage and deliver the answers to their challenges. Everything we're talking about, they weren't kind of esoteric use cases. They were genuinely some of the top issues that Vail had that we could immediately help bring benefit and value to through the technology and then continue to iterate and continue to build on it from there. And how much of what you're seeing when it comes to enterprises choosing to host their own infrastructure, their own accelerators, How much of that is cost and TCO driven, as we've been discussing, versus privacy, versus compliance, versus other concerns? Are those drivers as well? Yes. Obviously, cost is going to come into it. When your return on investment is somewhere like six to nine months, that becomes an incredibly compelling argument. If you're doing heavyweight processing like Luke was talking about, maybe you're in like three to four months, you know, from a TCO perspective, like that's incredibly compelling. So I don't want to downplay that. However, the sovereign demands, privacy regulations, regulated environments, privacy security, all of those different components around, you know, really what we're talking about is in the end the data. And, you know, not just worrying about the compute, but really worrying about that access to that core information and data that that entity has. And how do they do that most effectively within the guidelines, regulations, compliance that they need to operate under and also that they want to operate under to protect, you know, in the case of like Vale, their citizenry. But, you know, if you have another entity, you know, whether it's their customers or citizens or all of the above. And so we see people, and that's one of the reasons we have choice that scales out from like AI Factory with thousands of GPUs that can be installed in a sovereign implementation. Private cloud for AI is really one of our flagship offerings where we're bringing that private cloud experience, but again, into that on-prem environment where you can manage it behind your own firewall or, of course, have it hosted. You know, there's all different ideas of what data center looks like these days. And then, of course, we have compute as well, if you want to start out. One of the reasons that we leaned in with Private Cloud for AI is that we also allows you to lifecycle manage the AI stack in a comprehensive manner, which is so important when you think about something like a smart city where we can deliver it all as a bundle, all as a stack that they can lifecycle manage over the lifespan of that asset so that they're not having to worry about this version of Kubernetes with this version of that, with this version of that. Because as nice and neat as the AI stack looks on the slide, when you go into reality, you can go into the wild pretty quickly. So it's really looking at that customer choice. But I think to your point, yes, cost is obviously part of it, But people are making choices also based on where they set, what are they trying to do, and what is the data really that they're doing it with. So, Sam, we at Kamehwaza pioneered the whole concept of distributed inferencing. And the reason for that is in our large federal government and Fortune 500 customers, the main issue for getting an actual high efficacy in an agent workflow is access to data. And really it comes down to data gravity. we're talking organizations that have petabytes and exabytes of data across multiple locations across multiple clouds and you have to have the agent actually be able to traverse that and keep in mind a typical agentic workflow might have a thousand to ten thousand data pull requests between individual files pages of files changes in the files even and that's before you even start looking at structured and SQL queries where it has to like learn the schema and actually pull that in and then rework a SQL call and do it again and again and again until it actually gets the data it needs that fills sort of the request bucket. So it's how do you do all of this in most enterprises for the last 65 years have tried to build that pristine data lake and have the perfect data lake with the perfect compute in one location but not a single organization has done in the last 65 years So now you can finally bring AI to all of the data sets and process it. And that is really the game changer. And that's what we see as sort of a big reason for not just being untoken bound, but having to run this in private data centers and in clouds. You need to bring the compute to the data. That's a real paradigm change that Genitive AI has brought. In addition to that, taking a step back from the particular use cases that we've discussed and your particular technology approach, what would you say are broader insights and lessons that you've gleaned from your work with, you know, Fortune 500s, large governments and the like that, you know, builders who are listening to this, technology leaders who are listening to this can take and apply to their own projects? So, I mean, a couple of just sort of axioms we have is old data is worse than no data. Because if you have outdated data, you're going to get a very good answer back that's going to be wrong. And it's going to be nearly impossible to tell that that's the case. The sightings will be correct and everything will be correct. But yet that data is just simply out of date. So once again, the need to always be scanning any data change, any file change, any add to a file, any delete of a file is incredibly important. And even in the structured scenario, especially in the unstructured scenario, and re-vectorizing that, reprocessing that, people don't realize how much compute, how much of a GPU cycle, that's just going to take up in an enterprise to keep fresh. Second, security is different in the agentic world. Once I get a little tooth of our horn, we had to build a concept of relationship-based authentication because you have a user that talks to an agent. This agent now could talk to this much data. He'll talk to an agent that could talk to this much data and so on and so forth. and if those agents infer from a larger data set to return an answer you've effectively now risen the escalation level or risen the the capability what that user should have been able to see because you can have two totally uncorrelated uh sets of data that a larger agent now pulls a reasoning together and gives an answer back and that user shouldn't know that whatsoever i think there's a bunch of vendors trying to solve that with like agent to agent protocol and and others along those lines have you looked at those things uh i'd say good luck to all of them yes uh the problem set is actually you have to understand the data. And to understand the data, you have to understand the metadata and understand the metadata. You probably even need an ontology off of it. Once again, it looks all uncorrelated in what it's sort of doing. And then you have to tie all the security and access controls to all of that metadata to the actual data retrieval. Like I said, it's quite a complicated thing. It typically takes one system to manage it. Now you can have an agent within that sphere talk to another agent. That's great what agent to agent or even some of the agent systems like AG2 and Crew AI do a great job on that. But that's more talking about external systems. I'm talking them within the sphere of control of an enterprise. Last but not least, once again, I can't stress it enough. Your data is ready to start today. Like it's just like my number one pitch. Accenture and all these great companies are still making all of their money doing the FOMO speech about saying you got to tag your data. You got to move your data to the right place. It's got to be cleansed in the right format. And then by then, go back to my first one. It's already old and out of date. It's even worse than not having data again. So you just got to get started today. Those would be our big axioms we're seeing out there right now. Awesome. How about you, Robin? You know, I think that we can't stress enough that just getting started perspective, people are scared of their data, sometimes quite rightfully so. But I think being able to leverage the new agentic capabilities, you can start to actually understand what you could do as opposed to having it prevent you and instead leaning into what you can't do. And I think that to me, like the fear factor is still rampant, whether it's a fear of failure. I'm like, the AI is not going to giggle at you. Like, just try something. Well, maybe it will, but that's a different issue for a different podcast. Yeah, like that's a totally different podcast. Depends on that system prompt. Right. But I think that that to me is one of the, you don't have to be a data scientist. You don't have to be brilliant at coding. You can work with partners such as Kami Waza and other to help you achieve those end results that are keeping you up at night. And to start to break it down into something like I refer to as going mud puddle by mud puddle. Because a lot of times people, they're like, AI can do all of these things and you have this grandiose vision. But grandiose visions are super hard to get right out of the gate. But if you start with a mud puddle and then you go to the next mud puddle and the next pud puddle, suddenly you're getting like a bit more of a pond. And then you can have, you know, like you can actually start to bring this together and you can have much more of that grandiose vision. But if you start with a very, again, my favorite practical and tactical, very focused approach to what is the first thing we're going to do. Something like 508 is on compliance is certainly not the only thing as Kamawaza is demonstrating they can do. But it was something that was very meaningful and very focused that we could go execute and that we could help address one of their major pain points. Then we went to the housing one. But that was a mud puddle by mud puddle. We didn't say we're going to go fix all of your backend processes day one, or we would have kind of fallen over under the weight of our lofty and well-intentioned, but somewhat misguided intentions. Awesome. Awesome. Well, Robin and Luke, thanks so much for jumping on and sharing a little bit about what you're up to and this very cool use case or set of use cases. Yeah, thanks. Glad to be here. Thanks so much.