SHI Fall Summit 2025 | Luke Norris, Kamiwaza.ai,

Clips
More from SHI Fall Summit 2025

Luke Norris

Co-Founder & CEO

KamiwazaAI

play_circle_outline Luke Norris shares insights on agentic AI and real-world implementations.

play_circle_outline Overview of Kamiwaza, a software startup focusing on Fortune 500 and government clients.

play_circle_outline Enhancing Accessibility: AI Agent ARIA and Section 508 Compliance for Government Websites Supporting Individuals with Disabilities

Info
Transcript

Luke Norris, Kamiwaza.ai,

Luke Norris

Co-Founder & CEO KamiwazaAI

In this SHI Fall Summit interview, theCUBE’s Paul Gillin sits down with Luke Norris, co-founder and chief executive officer of Kamiwaza, to unpack how agentic AI is moving from theory to production inside large enterprises and government. Norris outlines Kamiwaza’s horizontal AI orchestration engine – built on an “inference mesh,” a distributed data engine with an enterprise-wide metadata catalog and a relationship-based authentication control that preserves user permissions across chained agents. The discussion dives into real security implications such as t... Read more

explore Keep Exploring

What is the name of the software company co-founded by Luke Norris, and where is it based? add

What recent developments has the company experienced and what are its key offerings in the AI space? add

What is the significance of section 508 and its compliance in current initiatives? add

bolt Powered by CUBE AI

Luke Norris, Kamiwaza.ai,

search

Paul Gillin

>> This is theCUBE. I'm Paul Gillin. We're here in Somerset, New Jersey for the SHI Fall Summit 2025 in partnership with HPE. Joining me today is Luke Norris, who is the co-founder and CEO of Kamiwaza. I mispronounced that name the first time I used it. Kamiwaza, a software company based in Colorado. And I have to admit, Luke, I wasn't familiar with your company before, just a short time ago. You are a startup, so how did the company come to be? What are you good at?

>> Yeah, so we really broke on the scene in the last year. We focus on the Fortune 500 global 2000 and large Fed space, where we have significant contracts. The company itself has been formed to be a horizontal AI orchestration engine and we really differentiate ourselves through connecting all the data of an enterprise, all the security to that data of the enterprise and then layering in an ontology scheme. And last but not least, we process the data where it resides. So you don't have to move it, you don't have to change it, you don't have to do anything with it. So this really enables AI agents to work anywhere across the data state of an enterprise and get immediate results.

Paul Gillin

>> So that's a pretty sophisticated technical mesh you've put together. What are some of the core technologies you use?

>> So from a core technology base on the orchestration engine, we really focus on three factors. First, we have what we call an inference mesh, so the ability to have multiple stacks of our software and multiple locations. An inference request can come into one of them and it'll route it dynamically to the right stack that's next to the data. Routing based off of data through our distributed data engine. Since we have our own metadata catalog of all the data in the enterprise and the access to that data, we can actually tell the agent where to actually move the inference so it processes that data locally using the same security. And that's the third part. We had to build a relationship-based authentication control. So when a user accesses as an agent and that agent might access a larger agent, that same relationship of that user is maintained all the way through. So the agent's only inference data that the user could have. When you combine all of that, we can do anything within the enterprise.

Paul Gillin

>> So essentially the agent is a proxy for the user and has the same permissions and privileges as a user would have?

>> That's the simple way to say it, but think about it, most users are going to talk to a smaller agent who talks to a world agent, who might talk to another smaller agent at another site. So it's maintaining that relationship is very, very complicated and that's what keeps the controls in place.

Paul Gillin

>> So you're way out there on the agent learning curve obviously. Have there been any surprises along the way?

>> So ultimately I think the surprise was the level of complexity that the enterprise had to figure out from a security standpoint. Because the fact that now you have this limitless intelligence that now knows about all your data and what it can actually do with that is incredibly complicated. So keeping the security confines and it's been around for about 25 years, this concept of the mosaic effect. And you can have two uncorrelated copies of data with two totally different sort of representations. But if an AI can correlate a larger meaning between those two datas, it actually raises a security threat or the security vulnerability of that data. AI opens up all of these new things that were only theoretical, which now can actually be concrete implemented in the enterprise. And once again, securities at the forefront of thinking through all of that.

Paul Gillin

>> So maybe relationships that wouldn't be evident to a human can be evident to an AI and that can create an unanticipated security threat.

>> Yeah, if you think about it even from a simple form, if a user could access an agent that says the typical tier of an employee makes this much money, and then another user can ask the agents what are the options? Are RSU vesting for employees that started it at this particular time and maybe within that tier. Well now you can correlate those two things and actually understand what your boss or somebody within the company actually makes. Now it's a simple, simple answer, a simple example, but now you can take us to the next level and the next level and really start to get some cross correlations of very sensitive data.

Paul Gillin

>> So I can see why you're targeting government agencies as one of your core constituents. You're here at this conference talking about section 508, which I'm sure many people are unaware of, but for government agencies is a very real issue they have to grapple with. Why is 508 something that you're focusing on now?

>> So section 508 is the federal guidelines for ADA compliancy At the state level though, every state has adopted their own stringent compliancy terms that fall even further than 508. And then you have consortiums of states, et cetera. Just now and all of these states have been working it out along with those federal guidelines. Next year, the fines and the actual limitation of state funds, government grants and just government contracts are going to be tied to people achieving 508 compliancy. At Kamiwaza, we always try to do an AI for good initiative every quarter. We're doing these very complex use cases for the federal government and large Fortune 500s where we try to get something out there that, A, is good and, B, actually shows the power of agentic AI. We worked with our friends over at HPE to really wrap in this 508 compliance agent, we call it ARIA. That allows us to now actually change the lives, literally it does, for people that have disabilities, whether it's motor our visual on how they actually access government services and functional websites and we're really proud of it.

Paul Gillin

>> So what does your agent look for when it goes in and begins crawling a website?

>> So it's multiple agents. If you think about it, you have a computer use agent, which actually can take over a browser, take over a machine, and actually start to utilize that to access the website. You have a standard large LLM which then understands the actual website. And then you have to do a visual model that actually visually represents what's going on inside of the PDFs, the actual metadata structures, even the graphs that might be in the HTML. You have to orchestrate those three things together to get a full understanding of what's being represented on that particular webpage. Further, the agent then seizes that compliance, not only at the federal regulation, 508, but at the state regulations as well. Then it actually goes through and remediates that, and that's the incredibly cool thing. It applies the metadata structure to the individual images. It actually changes the HTML to represent what's in the graphs so that when a browser that has EDA compliance for the reading capabilities or once again even motor generative capabilities to move through, it now knows what to do and accomplish off of that.

Paul Gillin

>> What are some of the stickiest problems that you encounter when doing this remediation?

>> If you think about the millions of ways individual pictures could be represented, it's quite sticky. So we actually have to have a tight interplay of the visual models and the large language models to get a good description and a tight, accurate description of what's in that picture. You can over-describe things in the picture. You could talk about individual blades of grass. You could talk about not just hair color, but maybe the type of hair design that's there. And you don't want somebody with a E-reader having to go over that picture and spend 20 minutes just to get a description on that picture. But at the same time, you need enough description that's in there so they understand what's going on there so that they can mentally navigate through that as well. And that cross-play of making agents that are rather robust actually work with humans semantically is a very interesting and fun challenge that we're able to take on.

Paul Gillin

>> How specific are the regulations about how you do something like you're describing, creating the alt text for a picture?

>> It's extremely specific, to be honest. All the way down to the amount of metadata, the type of metadata, even the HTML code, what's the contrast of the actual background to foreground? All of that stuff is laid out in extensive regulations. And if you're a human, having to manage each one of those regulations and actually adhering to it is quite complex. And that's why it takes so long to do this. You have to go through each one of the regulations and check off each page, each line of code. And with this, it just runs right through and remediates it instantly.

Paul Gillin

>> What kind of verification steps do you have in place to make sure that the agent is doing what it's supposed to do?

>> So that's the fun part as well. This is a living policy, even at the state level, they're adding to it and they're changing it on nearly a quarterly basis in some of these states. So keeping up with that is very, very complex. So we run that multi-agent schema just to do that first remediation. And then there's a second one which actually checks all the latest regulations, rebuilds that out, and then reapplies it to any of the remediated pieces to actually see if it's been fully kept up with law and implemented.

Paul Gillin

>> What's been your experience in the field so far? How quickly are you able to fix these problems at a mid-sized government agency?

>> The amazing part, and we do call this sort of a wedge solution, instead of trying to bite off these mid-sized agencies that do everything in the agent world, we say, "This one particular agent do this one particular job. It's bundled with our friends at HPE. It's actually installed and serviced from SHI. And it can get in there." And we're talking within days. It starts its process. Within a week, the website's typically at a full remediated state.

Paul Gillin

>> And how would this compare to a manual process?

>> So our very first Lighthouse use case was a decent sized city, world renowned, and they were thinking it was going to take years, literally years, to not only get the full website fully remediated, but then be kept keeping up with it because of all the document changes, the processing changes, and just everything that goes on from these live websites. They're not simple information giving websites. They're fully interactive websites for you to access your municipality and services that are needed. So it's rather dynamic. So they were thinking it was going to be two to three people that were going to work anywhere between two to three years to get a fully remediated and be actually caught up from a reasonable time, about a month behind on any ad libs and changes. The fact that this can come in and remediate this and be active within a week is not only a time savings, but it's a infinite CapEx to OPEX conversion. You're moving that idea of having multiple people over multiple years have to maintain and manage this, to a simple CapEx purchase and unleashing an agent.

Paul Gillin

>> So you're taking four to six person years and condensing that down to one week?

>> That's correct.

Paul Gillin

>> And this is what you called an earlier wedge solution. What does that mean?

>> So once again, the concept of buying into agentic AI, it's quite complicated. And since it's so broad, it's scary to where to start with this. With this it's a very tight solution that has a very tight ROI that's bundled nearly like an appliance that can be installed and turned on and get that immediate ROI. The beauty is now that this agent understands the entire website of these agencies, it also understands the link of that website, it understands what they're doing. That data now can be used for other agents to do other things within the enterprise. So it's a great part to start, very simple to start, but now it's an expansive capability that you can add on and add on.

Paul Gillin

>> Do you build your own LLM for this purpose or are you using a variety of commercial models?

>> We use a variety of commercial and what's called open weight models for these particular solutions, and that's the other great part of this. Those models are simply added onto our orchestration platform. So next week, next month, next year, it's actually going to get better and better as the state of the technology comes up and our customers get that just as a standard support and service from us.

Paul Gillin

>> How did this partnership come together with HPE and SHI?

>> So that's the beauty of somebody as forward-thinking as HPE when it comes to AI. They wanted to get into actual solving problems and ROIs and solutions, and from SHI's perspective, we needed somebody that could do the services and the integrations to make this wedge solution actually implementable. So when we all got together in a room and talked about what were the ways to get the easiest ROI, what are the ways to make the biggest impact? The 508 solution came to market. And it was cemented when we were talking to Lighthouse customer, and they said that was the number one problem that they were having to deal with from a budget overrun next year and just an outright effort that they weren't planning or they didn't know what to do. We tackled it, the three of us together packaged it up and we're really excited to bring it to the market.

Paul Gillin

>> Finally, Luke, Kamiwaza. When I heard I was going to be speaking to the CEO of Kamiwaza, I thought it'd be a Japanese gentleman. What does the word mean? How did you choose it?

>> So like a lot of amazing words in the Japanese language, there's not a direct translation here in our English language. Effectively, it means divine skill or divine capability. It really means something that can go to that next level in human capability and processing. So we're really excited that the word was available on top of it, a fun little history. My favorite video game growing up on the PlayStation 2 was named Kamiwaza: Art of the Thief. And Seth Godin actually had a great paragraph on Kamiwaza in his book, the Icarus Deception. So it just mapped perfectly to the timing of us starting this company and bringing this capability to market.

Paul Gillin

>> I'm a huge Seth Godin fan. Luke Norris, thanks so much for joining us.

>> Thank you.

Paul Gillin

>> Telling us about a real live implementation of a agentic AI that's out there and working today. I'm Paul Gillin. This is theCube, the leader in tech news and analysis.