DM Radio · 2026-07-01 · 45 min
Key moments - from our scoring
Substance score
38 / 100
Five dimensions, 20 points each
Chad Kinney, VP of Products at Everpure (formerly Pure Storage), and Sean Rosemarin discuss why context is the critical missing piece for enterprise AI systems. The conversation centers on data primacy - the architectural shift from siloed, application-centric data to unified, context-aware systems that enable AI agents to make informed decisions at machine speed. Kinney walks through a practical example: an agent approving profitable orders can only do so intelligently when it has shared context across Salesforce, supplier data, product costs, and COGS - something siloed applications cannot provide. Everpure's approach unifies the data plane with an intelligent control plane, uses agents for machine-level decisions (not humans), and solves both the 'data up' problem (unifying fragmented data sources via their One Touch acquisition, which builds knowledge graphs of data interconnections) and the 'data down' problem (classifying, understanding, and enforcing policies on data). The company is also applying this to predictive infrastructure management - their support team now relies on machine-driven fingerprint detection to prevent outages before they occur, shifting IT practitioners from reactive firefighting to proactive strategy. This represents a fundamental architectural shift from traditional OLTP/OLAP separation toward a unified runtime powered by AI agents, underpinned by Pure's key-value store, direct Flash pairing with Purity software, and high-concurrency design.
Siloed applications lack shared context about data interdependencies. For example, a Salesforce agent approving orders based only on price won't understand COGS, supplier costs, or product economics - so it makes uninformed decisions. When context is unified across all data sources, agents can make intelligent, profitable decisions.
One Touch scans your CMDB and data endpoints, interrogates data sources with read-level access, classifies what exists, identifies PII and policy violations, and builds a knowledge graph showing how data interconnects across systems - a process that once took weeks but is becoming increasingly machine-driven.
Pure's support infrastructure identifies event fingerprints that cause disruptions, then uses predictive systems to forecast issues before they occur. This predictive approach now generates 80+ percent of their support tickets automatically, so customers can address capacity, performance, and potential failures proactively rather than reactively.
Data primacy places unified, shared data at the center rather than applications. Instead of ETL processes running overnight and humans approving orders the next day, agents and analytics run on real-time shared context via a common system of record - enabling machine-speed decisions while maintaining compliance with defined intent and policies.
Everpure builds both scale-up and scale-out products on a key-value store architecture with direct Flash pairing via their Purity software, enabling very random transactional access with high concurrency. This contrasts with competitors using off-the-shelf SSDs, giving Everpure better efficiency and throughput for hundreds of agent calls per minute.
Our reviewer’s read on each dimension, with quotes from the episode.
There are a handful of genuinely useful, non-obvious points - particularly the CSV export stripping native security permissions and the intent-aware policy enforcement concept - but these are buried in heavy vendor marketing cadence, repetitive 'right?' exchanges, and re-statements of the same core ideas (context, context, context) with little depth added across the three segments.
by downloading it in a flat file in a CSV or an Excel format, you've essentially stripped all the security that was in that file
if you had A policy, for example, that said any copy of data, let's say for test dev or analytics or whatever, should never exist beyond 30 days. Then you have at least a window in time that, you know, I need to monitor data copies and track them for 30 days
The AI-factory / raw-materials analogy and the 'intent-aware' framing for data governance are moderately fresh, but knowledge graphs for enterprise context, data primacy, and AI-ready data are all well-circulated concepts in 2024 B2B tech discourse; nothing genuinely contrarian or first-principles is offered.
I think everybody jumped to machines first. Uh, the raw material was just whatever data I had in whatever uh, data lake data swamp
We don't bring heavy crude and light crude and oil sands into a, into a company making plastics. We refine it
Both guests are Everpure (Pure Storage) employees speaking at their own company conference, making this essentially a vendor product pitch rather than independent practitioner insight; Chad Kinney as VP of Products has legitimate product depth, but neither guest shares hard-won operator learnings from outside their own company's narrative.
That was my first job out of college, really. Implementing SAP
80 plus percent of our tickets are opened up automatically by this predictive system
The episode is almost entirely abstract; aside from one internal metric about automated ticket opening and a handful of named technology references (ServiceNow, Salesforce, Snowflake, Databricks), there are no customer names, no ROI figures, no comparative benchmarks, and no concrete deployment outcomes cited.
80 plus percent of our tickets are opened up automatically by this predictive system
Think ServiceNow. Here's all my data endpoints
The host demonstrates genuine historical knowledge (OLTP/OLAP, SAP HANA, data warehousing evolution) and lands one sharp follow-up about agent-driven access patterns at scale, but there is zero pushback on vendor claims, an indulgent sense-of-smell tangent, and a relentless affirmation loop ('right?', 'wow', 'that's wild') that lets ambitious marketing assertions pass entirely unchallenged.
how are we going to keep up with the kinds of access patterns that are going to be needed of agents asking thousands of questions a second instead of people asking 10 seconds and 10 questions in a minute
have you sent the notice out to all the data management vendors that uh, their services may not be required anymore
Computed from the transcript - who did the talking, and the words that came up most.
Recorded live at the Accelerate conference, this episode of DM Radio explores why data - not applications - is becoming the foundation of enterprise AI. Join host Eric Kavanagh as he speaks with Chadd Kenney of Everpure about the shift toward data-centric architectures and the importance of shared context for powering intelligent AI agents and real-time decision-making. Next, Shawn Rosemarin of Pure Storage explains why context is to AI what memory is to humans, illustrating how connected, meaningful data enables more accurate, trustworthy outcomes. Together, these conversations reveal why contextual intelligence is emerging as the key ingredient for the next generation of enterprise systems.
Transcribed and scored by The B2B Podcast Index.
Speaker A: You've heard am, um, you've heard fm. Now tune in to dm um Radio, the world's longest running show about data. Each week, host Eric Cavanaugh interviews the brightest minds in the world of information management. Want to be on a show? Send an email to infomradio Biz. Now, here's your host, Eric Cavanaugh.
Speaker B: All right, ladies and gentlemen, hello and welcome back once again to DM M Radio. Yours truly, Eric Cavanaugh is here on the road at Everpure, Accelerate. Everpure. Pure storage, then pure. Now it's Everpure. It's going to always be pure from now on. I think the name change is going to stick. And I'm here with Chad Kinney, VP of products. And Chad, uh, you and I have talked before on the show and, uh, some things have changed, some things haven't. I think that I've been looking at the press releases and listening to the presentations. And this concept of data primacy I think is pretty interesting. And it really challenges a lot of widely held architectural beliefs around information systems. Right.
Speaker C: So, you know what's interesting is, um, as you look at data just holistically, I think we all believe that data should be in the center of everything. But unfortunately, if you think about it, it's very much siloed in applications. And the context and workflows of those applications are also centralized and siloed.
Speaker B: Right?
Speaker C: So as you're starting to, like, rationalize data, you have the problem of not being able to have shared context in order to do things. Like, I want to create an agent to go create a workflow that has contextual understanding of multiple different applications.
Speaker B: Right?
Speaker C: So let me give you a fun example because I think it articulates it even better. Imagine if you were going to create an agent and that agent was going to approve profitable orders.
Speaker B: Okay, Okay.
Speaker C: I only want you to approve profit orders. Well, in a world where applications are siloed, it may make a decision just based on the look of it. So let's look at Salesforce. I see an order and, well, it seems 50 cheaper than the last one, so maybe that's more profitable. So I'm going to approve that one versus this one.
Speaker B: All right?
Speaker C: It has no context of cogs, price of parts that go into the system whatsoever.
Speaker B: Right?
Speaker C: So it makes decisions based on the context that it has.
Speaker B: Right.
Speaker C: But when you have context on all of your suppliers, how that's embedded into products, what the cost of those products are, all of a sudden they can make informed decisions to say, now That I have shared context of all of your data. I can go off and say I know the price of the product, so I know when it's profitable and so I can improve profitable order.
Speaker B: Right. Well, you know, like the workflows are similar, they're just tremendously expedited, I think is one way to look at it. Right. Because the agents are going to go find little pieces of information in a very short period of time and then make a decision as opposed to the traditional way like think ETL into the system overnight and then we'll approve manually this whole set of orders. And we're now changing that by orders of magnitude in terms of speed for sure.
Speaker C: Right, let's break it down. So today you've got applications that are fragmented out of all this information.
Speaker B: Right.
Speaker C: Our belief is that there needs to be shared context and that's what we're calling our universal, uh, data intelligence. That shared context now lets you understand what exists out there. Below that you've got a set of data that is also siloed to its understanding. And data analytics, we integrated a lot of that in order to build dashboards to analyze it.
Speaker B: Right.
Speaker C: But we also believe that with data primacy, that data should be analyzed and processed on real time information. The last thing you want to do is wait for the ETL process to complete and then have an agent be able to approve the order. Well, the order may have already missed that time at that point.
Speaker B: Right.
Speaker C: And so you want to be able to run agents, analytics and your apps on a common system of record. And to do so you've got to do a bunch of kind of, you know, pulling that application information and consolidating out to create that system. Right. And then all that's really left are these workflows. And agents are killer at doing workflows.
Speaker B: Right.
Speaker C: Uh, and so it allows you to be able to then process that data in a more holistic fashion. So shared, shared, uh, context, systems of record, and then you can build these agents, apps and analytics on top.
Speaker B: That's wild. And you were really bringing it all back together again. That's the thing I find so interesting if you think about the history of data warehousing. Well, 30, 40, 50 years ago, 45 years ago, people figured out it's really hard to query an erp. That's not what it was designed to do. It's designed for transaction processing, um, which is a very different thing. That's why you had OLTP and olap, Right. Olab, basically. But now we're trying to collapse all that into a unified runtime that uses agents to bang through these processes, like I say, in a fraction of the time that we used to. So that's why you got to be real careful. Right.
Speaker C: I think there's going to be a bit of an evolution too. Right. There's certain applications that uh, you can pull this off and there's certain ones that will probably stay still in the transactional layer where it's an app. Uh, but the context needs to be shared and that system of record needs to be consolidated up so that other things can process that data.
Speaker B: Right.
Speaker C: Um, but this will be a bit of an evolution, I think. There's a couple things that we're announcing at this show that I think would be interesting to articulate. We're solving a lot of the infrastructure up to data problems. That is unifying the data plane, building an intelligent control plane that allows you for operations autonomy, using agents to manage machine level decisions, not humans. Humans don't scale enough, uh, to be able to manage all the decisions that need to be made anymore.
Speaker B: Right.
Speaker C: You need a, a machine to help you with those recommendations and if anything, just go off and do it.
Speaker B: Right.
Speaker C: But we're also solving the data down problem, which is I need to understand and classify data so that then I can make sure the policies that my infrastructure runs on align with those things.
Speaker A: Right.
Speaker C: And there's a multitude of other things you may understand. Like certain, certain data. Huh, is hotter at certain periods of time. And so you may want to place it in different locations based upon the way that the data is accessed.
Speaker D: Right.
Speaker C: And so allowing them to have a data down view gives you a bunch of key ways of thinking about things. But architecturally the thing that we're doing very different here is we are, um, while we work, while, you know, our data intelligence works great with pure. It actually goes across all enterprise data, whether it's on pure or whether it's a third party storage solution, whether it's SaaS or infrastructure as a service, whatever, it can interrogate it, uh, and give you that one kind of context view of everything.
Speaker A: Wow.
Speaker B: And so you've also got the One Touch acquisition and now One Touch is coming in and basically doing this scan of your information architecture, finding all your different data sources, classifying them, warning you where this pii, warning you where there are things that are wrong. And then you've obviously got to go through and have human beings or first agents maybe validate all that stuff, say yes, no, yes, no, all the way through that. And they say it takes Like a month. I'm like a month. Wow.
Speaker C: So what's pretty cool is it goes off and looks at your CMDB and says like, what data sources are out there? Think ServiceNow. Here's all my data endpoints, right? Then you give it access to interrogate the data. Uh, you give it read level access and it goes through and interrogates all of the data, starts to classify it so you understand what it is. Then what is the coolest part of this is it builds a knowledge graph m that tells you how data interconnects with each other. These are areas of integrations that used to be very human oriented, which now, when you classify all of it together, is becoming more machine driven. And the capabilities of that type of stuff is going to be, uh, pretty awesome to see in the future.
Speaker B: Well, it is. And it seems to me that again, it's like this AI is a forcing function that's making us pay attention to the stuff we should have been paying attention to for all these years, right? Yes.
Speaker C: It's putting a light on the fact that there are certain aspects that have gone beyond human capabilities. Let me give you a fun one. Most customers, when they think about policy as an example, I say that I'm going to um, you know, make sure I have this policy for retention of snapshots. Just as an example, right? Well, humans inevitably give different results all over the time because, you know, you're in a rush to do something, maybe you forget to do some stuff or maybe you didn't know about the policy and created something completely different with no controls and no tracking. These just kind of are out there in the wild. And the only time you find out that something's wrong is usually during an incident like there's a ransomware attack and you realize, oh, we don't have snapshots for this volume. Uh, thankfully at pure we default safe mode and make snapshots be available for recovery. Um, but maybe you don't have a long enough retention. Maybe your policy was 30 days and you only set a seven. But having a machine actually go look at the telemetry data and tell you when you have violations of your policy, well, it's great. So you've got to define the policy. You've got to then go build a system to go make sure it's looked at and then you got to go find one to remediate it. And so a lot of the announcements we're making uh, at the show here today is helping people build data controls, helping them build reporting facilities of violations and Then go off and remediate it and to make them compliant.
Speaker B: Right. Well, it's interesting because these storage vendors are now very, very interesting and very exciting because we have to find some way to be able to feed the agents. And that's the clarion call.
Speaker D: Right?
Speaker B: It's like we've got to be able to get the data to the right place at the right time, which we always talked about, but it was at human speed. Now agents are the consum. Consumers and even though, you know, I think we're still a ways off from them running fluidly, all cylinders, that kind of scenario. Yeah, it's happening already.
Speaker C: Right.
Speaker B: And that's putting tremendous pressure on it teams on data teams to come together, coalesce and make sure that everyone has their right view of what's happening. Right.
Speaker C: Yeah. Uh, I think a lot of it too is that you now can run at machine speed, but you need to be able to articulate intent.
Speaker B: Right.
Speaker C: That's one thing that's been kind of missing because without intent, the agents don't exactly know what your intention was originally.
Speaker B: Right.
Speaker C: So it can't tell you if it's right or wrong in the way it's actually deployed. And so as you define intent, then the agents can very quickly go say where it's wrong. But, uh, part of that control mechanism is defining that first and then you can run at machine scale to validate. Um, so a very different world we're trying to build in.
Speaker B: Yeah. So a good buddy of mine has just launched a company and they do agentic data pipelines. And one of the things he talks about is being intent aware. And the immediate thing I picked up on was, you know, that's actually very interesting because GDPR tells you that you need to define your intention, uh, for this data when you collect the data, when you use the data. So now if you have this intention, well, guess what? An AI, an LLM infused AI basically can look at that, read it, parse it, go, okay, this is supposed to keep it for X amount of time. So by just documenting the intent, you can now actually leverage that intention with
Speaker C: these new workflows and validate that it's actually being delivered upon that. That's the cool part of it. Like I mentioned this for a quick, quick example. Like I hear consistently from many organizations that like exfiltration events occurred some cybersecurity issue.
Speaker B: Right.
Speaker C: Um, because it was a copy of data that someone created.
Speaker B: Right.
Speaker C: Forgot that existed, no one was monitoring it, and so the data was exfiltrated. But if you had A policy, for example, that said any copy of data, let's say for test dev or analytics or whatever, should never exist beyond 30 days. Then you have at least a window in time that, you know, I need to monitor data copies and track them for 30 days. After that I know they're gone.
Speaker A: Right.
Speaker C: And if there's ever an instance where they are still around, it's a violation to my intent. They should, they should get rid of that.
Speaker B: That's just wild. I mean, how is it changing the job of the day to day practitioner to adapt to all this?
Speaker C: I think they're loving it because I would say it would be tough for me to sleep at night knowing that I don't actually know if my policies are being enacted upon or if I configure things correctly. It would be so great to just know that the intention that I have has been delivered upon and I feel safe because of it. Um, today it's an unimaginable pain of just getting lots of alerts and trying to rationalize and make decisions that in many cases you're not informed well enough to actually go and act on.
Speaker B: Well, it's, you know, the speed has increased so much, but in many ways, like I said, we're getting back to basics of what we were supposed to be doing. And what's kind of exciting and scary is that you can now scan an entire data environment you can release. And here's what I'm a little bit nervous about is suddenly realizing how many policy violations there were. Oh yeah. And then going, oh my goodness, how are we going to fix this?
Speaker C: What's cool about a, uh, policy driven approach is you can say, here's the policy or intention and I want you to go fix all of them.
Speaker B: Right.
Speaker C: It just goes an upd. It's all going to show it on stage to on main stage of, you know, changing a policy and actually seeing it fix violations.
Speaker B: Right.
Speaker C: Um, but yeah, I think that true power is an interesting one and one of the reasons why this is so cool for us. Like when we built our support infrastructure, we realized that there was a series of events that occurred. We call these fingerprints that caused, let's say, uh, some sort of disruption, outage, whatever. And we started to now predict issues before they occurred.
Speaker D: Wow.
Speaker C: So that they would never occur. And the reason why our support's so stellar is because, you know, 80 plus percent of our tickets are opened up automatically by this predictive system.
Speaker B: Right. Interesting.
Speaker C: What we're doing now is we're taking that same methodology towards, you know, uh, Performance and capacity and forecasting that. Hey, by the way, you're going to hit a performance issue in about 30 days. I want you to take this action. Well, what's great about Pure is we have this repository of telemetry information for every single customer. They can go use Copilot, they can ask it questions. It knows everything about everything.
Speaker B: Wow.
Speaker C: Things that would take you hours to figure out. Well, with that backing this thing, now, these systems can give you recommendations. And I think at some point people are just going to say, just go off and do it because you know better than me. So I don't need to even be part of the loop, um, because you're going to make a better decision than I would. I don't know any of this information.
Speaker A: Right.
Speaker C: But things like, hey, I should move a workload before that guy calls me and says that the application is running slow is such a big win.
Speaker B: Wow.
Speaker C: Most of their job is firefighting. Trying to fix these issues and taking that off their plate is like the biggest win ever. So when you say, what do practitioners think about it? I think they absolutely love this. This is like the best thing that could have ever happened to them. So they can focus on things that are much more strategic.
Speaker B: Right. Well, you make a good point that by being proactive and knowing when something is going to happen, that's confidence building because you know you're working on something valuable. In the old days, if you're just kind of scanning a system looking for any of the 87,000 problem can be out there. It's very demoralizing and you wind up cherry picking. Right now this is, it's more directed based upon the data that's in the system which needs to be addressed. And so at least, you know, uh, you're climbing up the right hill.
Speaker C: We used to always say to customers, um, Pure Storage are, you know, and now Everpure allows you to get your weekends back.
Speaker B: Oh, nice.
Speaker C: The reason why is because most of their weekends were spent firefighting with sev ones and outages on the old products. And our products were so reliable and they predicted the issues before they occurred that they actually got to spend time with their families on the weekends.
Speaker B: Wow.
Speaker C: Well, for this model, if you think about it, no one's really solved the problem of an application owner calling up and saying, my application slow. Go figure out what the problem is. And everyone runs around firefighting it.
Speaker B: Right.
Speaker C: If you can get ahead of that, the life of the practitioner is pretty awesome.
Speaker B: Yeah, that's pretty cool. You know, I remember, um, Pure talking About a few years ago, the data operating system. And that was the mechanism of action, basically. Is that still the mindset and is that why you're able to deliver on this? Because you think about the fact that the access patterns have now changed dramatically for accessing databases. Agents can do hundreds of calls in the time it took a human, you know, five minutes to do that sort of thing. So tell me about the OS and how it is that pure is able to bear all these requests in this different access pattern.
Speaker C: Yeah, so there's two big key areas. I said that. Well, more than two, but I'll just cover two. The first one is everything's built on a key value store.
Speaker B: Okay.
Speaker C: Both systems, one's a scale up product, one's a scale out product. Um, but that key value store allows us for very, very random transactional access with high concurrency in our scale out products. So we are incredibly well built for this new world. And I've always been built for, with just Flash as being the core media.
Speaker B: Wow.
Speaker C: The second big area is that we built a software and hardware pairing with direct Flash and purity. And what this is is we built our own SSDs. It allows us to manage densities and um, globally represent Flash. So now we're a software driven approach to handle the way that Flash is both weird leveled, um, how it's accessed, how we do GC, where everybody else is using off the shelf SSDs which act like little tiny storage arrays.
Speaker B: Interesting.
Speaker C: And so this is, you know, this is why we have the best efficiency in the market. Our key value store gives us the best concurrency and um, we're just amazingly well placed for this market.
Speaker B: That is very interesting and it's exciting to see the foundation get so much attention because if you get the foundation right, everything else works fine. It's a fundamental shift, folks. This is a really seriously big change. This is not some minor, uh, evolutionary alteration to the norm. This is a whole new way of doing things. Well folks, will be right back. You're listening to DM Radio.
Speaker A: Welcome back to dm radio. Here's your host, eric kavanaugh.
Speaker B: All right folks, back here at uh, Accelerate in Las Vegas, yours truly, Eric Kavanaugh here on DM Radio. Next up we've got Sean Rosemarin. Very, very intelligent, knowledgeable person. He thinks about things. He's got some great ideas to share with you today. And let's maybe start with that magic word, context. Everyone's talking about, context. Well, for a large language model, first of all, it's just a String of tokens that goes into the. That monster and a string of tokens that comes out. So the context is pretty important, especially how it's shaped and where it comes from. And Everpeer has a lot to say about that these days. What are your thoughts on context and how to get it where it needs to be?
Speaker D: Yeah, so I mean, one of the things I love about talking about new technology is there's always something you can relate it to.
Speaker B: Right.
Speaker D: Uh, when you think into history. And so I like to think about humans since I think the human race come quite a long way and has a lot of really interesting things going forward, I think so.
Speaker B: Pretty cool.
Speaker D: I'm totally biased. Uh, so if I, if I was to show you a picture of you in a canoe, uh, fishing when you were a kid.
Speaker B: Mhm.
Speaker D: The picture only shows the canoe and probably maybe you with fishing rod. But in your mind if you thought about that photo, you would say, oh yeah, that was, uh, that was when I went fishing with my dad. And uh, we were in this, you know, pretty cool place. And I remember what lure we were using and I remember what we were fishing for. Remember what we caught.
Speaker B: Right.
Speaker D: Remember what the weather was. Remember we got caught in that big storm and we had to race back and.
Speaker B: Right.
Speaker D: You know, we lost the fish because we were so, you know, worried about getting home that we actually didn't get our gear properly put away.
Speaker B: Right.
Speaker D: When you think about context like our. As humans, we're incredibly good at this. We know how everything relates. If there was a song at our wedding, we can replay, not just, uh, every time we hear it, we replay in our mind everything that that song, everything that that music means to us. So when we think about how we get AI to intelligently understand our data, it is really the grasp of that context. It is how do we get it to know the data and what is related to that data? Um, in the same way that a human would. As we look at photographs, we listen to music or we even hear a voice or talk to someone we haven't talked to for a long period of time.
Speaker B: Right? Yeah, that's fascinating. And a quick note about memories. Just a side note, maybe there is some connection here on the hardware level. But do you know the sense of smell triggers the most memories because there's no interneuron with, with touch, taste, hearing and vision. There's a sensory neuron, an interneuron and a motor neuron in your brain with olfactory, with your nose, there's no interneuron. So it's immediate. It goes from sensory neuron right to motor neuron. That's why the sense of smell is so powerful and really makes you go, I remember things in that wild.
Speaker D: Probably has to do with fight or flight. Probably has to do.
Speaker B: Yeah, right.
Speaker D: Making sure that we get out of the way.
Speaker B: That's right.
Speaker D: Really bad thing happens.
Speaker B: The original sensory perception concept like hey, I need to know what's happening. Business have that these days too. And since we're talking context, One Touch now, part of this ever pure world, uh, this is a really big deal. And granted these things always take time to fan out. There's always a long tail to enterprise software, to really any kind of software. But really to be able to take this data primacy angle, that's a foundational change in information architecture and one that is dramatic and significant, right?
Speaker D: It is. Because I mean you really have to think so when we talk about One Touch, we're talking about Everpure Data Intelligence.
Speaker B: Right?
Speaker D: Um, that's the, that's the branding that we launched for that product this week.
Speaker B: Right.
Speaker D: It's, it's not so much about find me the data.
Speaker B: Right.
Speaker D: I mean we've been able to ask Google to find us pictures of whatever. We've been able to use legal discovery technologies to find, uh, certain documents or certain strings of numbers or anything that might have been pii or anything that might have represented a Social Security number with One Touch. It's way more than that. It's really what is all the information that is contextually part of that knowledge graph? Right. We call this. It's sort of a canonical knowledge graph that says. I'll give you an example. If I'm looking at an invoice, finding the invoice is not hard. But where was the purchase requisition that's tied to that invoice?
Speaker B: Right.
Speaker D: Who was the person that originally asked for that invoice to be created? Um, what was the check that was actually cut to pay that invoice? What was the delivery and bill of lading that was attached to the product actually hitting uh, the floor. All of that contextual data has been the challenge because everything's been a static individual document.
Speaker B: But. Right.
Speaker D: How do I actually, how can I see everything that's connected to that document?
Speaker B: Right.
Speaker D: And that's what, that's what everp your data intelligence or One Touch, uh, fundamentally unlocks for us.
Speaker B: Yeah, that's a big deal. And I was very pleased to hear the word knowledge graph. We're talking all about graphs these days because graphs connect the dots. Basically, the dots are the entities. You have nodes and edges in a graph and the edges tend to be the characteristics, the nodes tend to be the things which can be people, products, services, whatever the case may be. But once you have a robust knowledge graph underpinning your information systems, you can hop all the way to the answers and you can connect those dots and that gives you the complete picture. You know what that gives you? Context. Mhm, right.
Speaker D: Yeah. And then context allows AI to see things with the same context as humans would have seen them, as opposed to kind of what we've been playing with over the last couple years, which is it'll read whatever you give it, it'll try to connect the dots. But ultimately what is the map? What is the graph?
Speaker B: Right.
Speaker D: I like what you said. I mean, graphs are cool again.
Speaker B: Right? Right, right. Well, they're cool. And because they are, they're the connective tissue that allows you to make sense of things. The, I guess the, the concern around agents, AI agents is that if they don't have a connection to the data they need, they're just going to make some stuff up. And that' language models do in low distribution areas of the training is they just throw something out there. Like for example, when they first came out, they said I wrote three books. I was like, oh, that's great. What were they called? I don't remember that. I blackout so long that I forgot writing a book. No, it was wrong. It was a hallucination. Unless it was seeing the future, which is possible.
Speaker D: Well, that's what we're finding more and more is that hallucinations really are tied to a lack of AI, uh, ready data. And you know, the other sort of formulation, uh, I'll share with you and your audience is, is when we kind of break down where we're really at.
Speaker B: Right.
Speaker D: We're building AI factories. The good news is, once again, as humans, we built a lot of factories in our time.
Speaker B: Right.
Speaker D: Think about the industrial revolution if we're building a factory. The first question we'd ask ourselves when we're building a factory is what are we building?
Speaker B: Right.
Speaker D: Shoes, cereal, clothing, socks. Uh, and then we would ask ourselves what raw materials would we need in order to build those goods. And then we would ask ourselves what machines we need to put in the factory in order to convert the raw material into the output.
Speaker B: Right.
Speaker D: So now fast forward to where we are with AI. I think everybody jumped to machines first. Mhm. Uh, the raw material was just whatever data I had in whatever uh, data lake data swamp Right. It happened to sit in and the output is, I don't know, take whatever I had, put it through the machines I bought, and let's see what we get.
Speaker B: Right.
Speaker D: And by the way, for prototyping and getting to know technology and getting, you know, people, uh, in tune with what's possible, this is all good, right? But now we're at the point where someone's going to ask, what is the return of investment on what we're building in that factory? And is the output and the cost of that output actually worth it versus what we had before?
Speaker B: That's right.
Speaker D: And so we have to get smarter in terms of thinking about this as a factory. And the number one thing we're focused on this week is how do I get the raw material in the case of data to be AI ready, data to be refined data like refined petroleum. Right, right. We don't bring heavy crude and light crude and oil sands into a, into a company making plastics. We refine it.
Speaker A: Right.
Speaker D: We then take the refined material.
Speaker B: Right.
Speaker D: And then we use that to create the factory. So that refining of the data is what drives context and what drives AI ready data that allows us to then get the output we want.
Speaker B: Right now, it makes a lot of sense. And, you know, this whole data primacy thing also makes a tremendous amount of sense. You think just the copies of data that are everywhere, knowing which copy is the latest, you always make copies to do some analysis, for example, or download a bunch of files to load into Excel to do your little machinations, basically. Well, every time you do that, you create a fork in the data trail, essentially, and that obscures the audit trail, that obscures giving that strategic view. So to take this data primacy approach, what you're saying is, look, get it right the first time, and that's the main thing. Billy Joel song for those who like Billy Joel, um, Get it right the First Time. It is complex, though. I was asking, how long does it take from the time you set this software up to the time where you're actually getting somewhere? At least a month, I'm like, I would tend to think. But there is so much value in having the golden record. We talked about the Golden Record for 25 years. Master data.
Speaker D: Single source of truth.
Speaker B: Single source of truth. Right. Single version of the truth, basically. But this at least puts us on a path to get there in a meaningful way. Right.
Speaker D: Well, and you've brought up a really good concern, which is if I'm taking a copy of the data, let's assume I've exported that copy of the data from the application in which it lives. Let's assume that application is Salesforce. That application is, you know, financial system. Well, you used your credentials to log in. You then ran a report, you then exported that report. You now have a flat file that you're going to load into your agent of choice. But by downloading it in a flat file in a CSV or an Excel format, you've essentially stripped all the security that was in that file. The only reason you got that report, because you have access.
Speaker B: Interesting. Yeah.
Speaker D: But now as that agent evolves and starts to be shared with other people, well, that data essentially has been broken of all its native security considerations.
Speaker A: Right.
Speaker D: Getting to your point of data primacy and the point we're making this week, when I'm actually looking at the data and the single source of data, that data is only accessible with its security intact.
Speaker B: Interesting.
Speaker D: And when you think about what are my agents allowed to see, what are my employees going to see, what are my suppliers going to see, what others going to see as they embrace some of these applications. Do you want to be thinking about, am I working on a copy? Am I working on a copy from the original security?
Speaker B: Right.
Speaker D: Not only has the copy changed, you brought that up earlier, but is that data, is that data consumable by that audience?
Speaker A: Right.
Speaker D: Yeah. These are big questions answered.
Speaker B: Yeah. So, so here's my biggest question in this new paradigm is data access patterns. So we've spent decades optimizing the performance of databases which have been the foundation of all information systems.
Speaker D: Right.
Speaker B: Everyone accepts that the database is the foundation. Large language models come along, we slap that on the end, and it's almost like adding a noise filter to a 40 year pristine process of deterministic answers. It's like, well, why would you do that? And the answer is because they're so amazing and they do such amazing things, but you have to use them at certain points in the workflow.
Speaker D: That's right.
Speaker B: You have to be careful about that. But my biggest question back to you now is data access patterns. If someone really embraces this data primacy vision that you have and makes that their go to plan going forward, how are we going to keep up with the kinds of access patterns that are going to be needed of agents asking thousands of questions a second instead of people asking 10 seconds and 10 questions in a minute.
Speaker D: So you're bringing up an incredibly, incredibly, uh, valid point, which is we are all thinking about this from the context of a human asking an agent a question.
Speaker B: Right.
Speaker D: We talk about latency but as humans, actually we're not that concerned with latency. If you take 5 seconds to respond to me in a conversation, I'm okay with that.
Speaker B: Right.
Speaker D: But when I have agents working with other agents, right. On multifaceted problems, right. I'm actually looking for sub milliseconds or even microsecond latency because I'm waiting to complete a much bigger workflow.
Speaker B: Right.
Speaker D: It's no longer an LLM where I'm asking a question, getting an answer. Thank you. Or Google Search back in the day. This is a constant two way communication. And by the way, that model might be hit by multiple agents at the same time.
Speaker B: Right.
Speaker D: So now we start to talk about, okay, what kind of hardware solutions, what kind of storage solutions? We talk about iops for training, inference, uh, for um, or latency, uh, for inference. That, and these are the concerns that come with what is your data platform, how is your data estate now set up?
Speaker B: Right.
Speaker D: And what I would tell you is the systems that we bought to do these things back in the days of a big Oracle database, and that's pretty much what we ran, or AP system, right. They're not going to scale and they're not going to deliver what we need for the next generation.
Speaker A: Right.
Speaker B: Well, so that's a very good question. So now we have to take a hard look at the substrate, if you will, the infrastructure that's going to be delivering these things. And I think it's going to be piecemeal. You know, I'll say we got about three minutes in this segment, but I remember the first time I talked to someone from Pure Storage and I was really impressed, this is probably seven, eight years ago, about how consultative the approach was, especially for a hardware vendor. Typically hardware vendors are like, oh, here's your stuff. Thank you, have a good one. But no, these, the people that I had on my show were very thoughtful about using the hardware as a stepping stone to get to the next level, to get to the next architecture. And that's still baked into your DNA, Right?
Speaker D: Right, it is, absolutely. So if you kind of follow our journey from back with the days we were talking about it, the first problem we solved was Flash. Bring flash to the masses, make Flash consumable by everyone.
Speaker B: Right.
Speaker D: And doing that we delivered purity, which essentially is the most efficient operating system for all. Flash everywhere. And in doing that we built that data storage layer on top of that. We then said, let's go after data set management. So you see fusion, you see pure one, you see a lot of the operational tooling that comes within Pure's platform. And you think, what did we do? We made it easier to manage. Day two, day three, day four day non destructive upgrades, uh, self service upgrades, uh, that whole piece of it. Now you're seeing us add on top of that M with the rebranding to Everpure to say, okay, not only are we storage management, not only are we data set management, we're now data intelligence as well. So the same place where your data has already moved to Flash running efficiently, operationally efficient. Now I'm going to put the context layer on the top.
Speaker A: Wow.
Speaker D: And that essentially delivers against the enterprise data cloud platform vision that we laid out the solution one. Right.
Speaker B: Well and that's a big deal for lots of different reasons. You know, I Remember learning about SAP S4 HANA and the vision that Hassel Platner had years ago.
Speaker D: That was my first job out of college, really.
Speaker B: Okay.
Speaker D: Implementing SAP.
Speaker B: He's amazing. I mean he's such a, he's such a neat guy. He's such a cool person too to be such a visionary. But he sat there and said, you know what, there's going to be this intersection of cost for solid state versus spinning disc. We're going to go ahead and bank on that and start building for it now. And they sure did. They built this thing up. Now I don't think he saw Kubernetes. Betty's in the uh, in this crystal ball which is a bit of a curveball. But nonetheless I, I see that same kind of vision now coming out of Everpure. Realizing that there is today we have to solve all these problems today we have to stay on top of things. This is one of the challenges of investing in new technology. But to my point, a minute ago, Everpure and Pure Storage, back in the day, the approach is very consultative. So it's, it's cognizant of the past, past, the present and the future. And that's a big deal folks. We're have one more segment coming up in a second uh, with our friend here. But yeah, send me an email if you want to be on the show sometime. Info Mradio biz that does come right to me. We're here at Accelerate. Very, very fun stuff. Very. It's game changing folks. Really, really amazing stuff. We'll be right back. You're listening to DM M Radio. Foreign.
Speaker A: Welcome back to dm m radio. Here's your host, eric kavanaugh.
Speaker B: All right folks, back here on DM Radio talking all things Everpure. We've got Sean Rosemarin with us from Everpure, formerly Pure Storage, then Pure, now it's Everpure. So you're kind of extending it out there, there. And you know, we had a very interesting talk this morning with your CEO who uh, made some, some bold statements and I asked them, hey, so have you sent the notice out to all the data management vendors that uh, their services may not be required anymore? That's a bit hyperbolic. Obviously there's a long tail to these things. But you know, when you think about the complexity of the so called modern data stack and you think about how you've now collapsed that down to one layer with automation, that's a huge deal. I mean, yeah, you have to burn some tokens, you have to, you know, spend some money classifying and organizing. But once that's there, all of a sudden I remember just to put it in context, 18 odd years ago in the early days of Diem radio, asking ETL vendors like, can you strategically see which data sets are moving where and when? Because I'll bet you're overlapping things and they can probably be more efficient. They're like, no, no, we can't see that. You've collapsed that whole architecture.
Speaker A: Right, Right.
Speaker D: Well, so, uh, yes, we're thinking, we're, we're asking, we're asking the world to think a little bit differently about how they are viewing their data holistically. M so in order to connect and drive context and drive intelligence and be able to answer more advanced questions, right. I need to be able to get outside the confines of an application controlling that particular piece of data.
Speaker B: Right?
Speaker D: But the most logical place to connect all of that, the most logical place to drive the intelligence and context and you know, uh, interrelationships across all that data is in the data platform.
Speaker B: Right?
Speaker D: Is in the storage platform.
Speaker B: Right?
Speaker D: And if you think about, you know, our unique position in that through purity, we essentially own and run and engineer the entire IO path. Right? And at this point in time, we're spending more money in R and D on an annual basis in storage and data platforms than any of our traditional competitors. Um, we're all in on this. I mean, this is, right? This is all about how do we build the innovation. In 2010s, it was all about getting flash to the mainstream, right? In the late 2010s, into the early 2020s, it was all how do we drive the operational efficiency for storage and data. Now it's all about how do we connect the data and how do we um, prepare it and ignite it. And get it at a place where it can essentially feed AI and drive the next round of innovation for businesses every.
Speaker C: Right.
Speaker B: That's very interesting. And I think it's becoming more clear now to the audience and certainly to me how it is that you are following through on Hassel Platinum's vision. What he saw is like, look, it can all be in memory. And when you have your ERP in memory, you expedite all kinds of things. You could do Monte Carlo assessments, very easy. You could do planning, you could do all kinds of things that used to take a lot of time and have big chunks of latency as you throw things over the wall and wait for that team to come back to you. All that stuff is not now real time. And if you're, if you're delivering on the entire data plane, you're serving CRM, supply chain management, erp, uh, alt, ticketing systems, everything is in there. So you can be able to problem solve and just, and save so much time and effort and focus on getting the business right. Isn't that about right?
Speaker A: It is.
Speaker D: And like, let's be honest, right, Everything here happens in iterations. So the original iteration of what we're just discussing was, oh, just take all the data and just dump it into Duke. Right, Right.
Speaker B: I remember those days.
Speaker D: And because that was the answer, it was like, okay, we got to get it out of the application.
Speaker B: Right. We knew it then.
Speaker D: So we'll just dump it out of the application, we'll just put it into a, uh, data warehouse.
Speaker A: Right.
Speaker D: Then we got a little bit smarter and said, okay, we basically created a data swamp. None of this has any context, none of this isn't connected, but it is one big bucket. Okay, maybe that wasn't the best idea. Right now we're seeing the snowflake and databricks. This concept of, okay, I'm moving my data into an intelligent platform where I have some ability to connect all of this data and I can run, um, uh, I can run workflows to ensure that those data sets are kept up to date. But ultimately, ah, this essentially is the next level that says how do I discover what data I have across any storage platform, across any cloud, across any as a service platform? And how do I natively build the intelligence into the data so that when I want to serve up a very specific set that. Right. I'd like to say certain AI agents, we have to feed them a high protein diet. We got to watch what goes in their mouth, make sure that we're very careful about what we give them sure. So how do I protect and govern that data? Send the right data upwards, allow it to be processed as a data stream.
Speaker A: Right.
Speaker D: Uh, and essentially get us to this holy grail that we started on 10 years ago.
Speaker B: Right. Well, and there's the stream component too, which is pretty important. So maybe let's just dive in real quick to that. How you. The execution path. Right. Because orders matter, you know, which some things are done matter. You need some orchestration there. Can you explain that a bit?
Speaker D: It yeah. So the easiest way to kind of think about it is the first thing I got to do is find the data. So let's assume I'm looking for. We'll get back to, um, I don't know. Let's say we're doing that invoice project we talked about earlier. So I got to find all the data relative to the invoices and all the data that is intercorrelated to all those invoices. So I can use Everpure Data Intelligence or One Touch to go out and find all that data, essentially tag it. But tagging it's not enough because there's pii and confidential information. All that document and I may not be able to export that, I may not be able to use it. So I have to actually mask it or redact any of the fields or any of the elements that are not considered, uh, or able to be brought forward.
Speaker B: Right.
Speaker D: Once I've identified all that and redacted it and essentially tagged it as part of that data set, I can then take that data set and I can index and vectorize it into a data stream where it essentially becomes a collection of what has been found. That collection that's now vectorized and indexed is in a perfect format.
Speaker B: Interesting.
Speaker D: And now feed A.I. uh, inquiries because essentially I'm going to use the metadata at the index to answer questions about that data set as opposed to having to go and open each individual document, m read what it is, try to connect it. All that intelligence is already in the vectorized data set.
Speaker B: Right. That's very interesting. So the workflow is end to end here. You do need to obviously deploy data intelligence and do the classification and the organization and kind of understand where everything is. But once that, that, once the bulk of that work is done, the rest of the information environment is running on all cylinders.
Speaker D: That's right. That's right. And remember, data stream is not one thing. You don't just create one stream. You're going to create different data streams for different use cases. We talked about an invoice project. But let's say we had an automobile project. Or if you happen to be a court, you might have a data stream for traffic court cases, might have a data stream for other kinds of cases. And by isolating the data that's most important and protecting, governing that data, you're going to get more accurate insights from the agent rather than giving it everything.
Speaker A: Right.
Speaker D: And to your point earlier, because then it's going to go make a bunch of assumptions to connect it all together and you're going to spend the next, you know, months to years troubleshooting all those assumptions, as opposed to find the data that's most relevant to the conversation, give it only what it needs.
Speaker A: Right.
Speaker D: And make it smarter from day one.
Speaker B: That's very interesting. And the stream itself, is that like Apache Kafka? Is it Apache Kafka under the hood? What, what is that and how does it relate to something like Kafka?
Speaker D: It's a vectorized database. Okay, okay. So for folks who've lived in a database world for a long time, uh, you know, we always like to joke that, you know, databases are the unsung hero. Right. We all rely on them. The most critical applications in the world run on Oracle, run on Microsoft, uh, SQL. And databases are really, really, really cool again.
Speaker B: Right.
Speaker D: Uh, for the general public because you really have to think, and you said it earlier, when I build all this context in this connection and what's related to what and what's in that photo and what is that about? I have to store that in a database, but I have to store it in a vectorized database because essentially I'm going to now have AI reading that vectorized database, answering its questions. And you can imagine that me going across metadata tables to answer.
Speaker B: Right.
Speaker D: It's way faster than me going across an entire repository of thousands of files and opening them up in different formats, trying to figure out what they are.
Speaker B: Right.
Speaker D: So this is all about accelerating.
Speaker B: I see.
Speaker D: The time to outcome.
Speaker B: I see. Okay, I understand.
Speaker D: Or what we call time to first token. Right. I mean the latency between the question.
Speaker B: Sure.
Speaker D: Uh, and when I start writing the answer.
Speaker B: Sure. And then when you get again, compare it to the complexity of the old way of doing things. Kicking off SQL, uh, queries and sparkle queries or whatever. That was a, a long complicated process. And you begin, kind of collapsed all that into one stream. Is that about.
Speaker D: Yeah, I mean you're uh, you're giving me, uh, giving me nightmares back to my early days in Microsoft, taking various data sets and trying to connect what's the primary key and what's the primary key? And these two connect and is that the right way to do it? And how am I going to bring this other table in?
Speaker B: Right, right, yes.
Speaker D: So thankfully, vectorization and indexing.
Speaker B: Right.
Speaker D: Has given us tremendous power.
Speaker B: Yeah.
Speaker D: Uh, to take on the projects that are, are, uh, are in front of us and uh, we're pretty excited about what the future holds.
Speaker B: Well, and so we got about a minute and a half left here. If you do this right, you should be optimizing the, the data layer and that should allow for cost containment and even cost reduction in storage and in other kinds of, of, let's call them, arcade processes.
Speaker C: Right.
Speaker B: Is that also part of the vision? It's like, guys, we're going to clean up the data house. The data house is going to be in order finally. We've talked about that for 40 years. That has never been done. Okay.
Speaker C: Until now.
Speaker B: It's never been done, but it's finally going to get there.
Speaker D: What do you think? Well, so remember, we've been improving, moving.
Speaker B: That's not done.
Speaker D: Um, but we've every little step right in a direction gets us closer to the end.
Speaker B: Right.
Speaker D: I do feel we're at an incredibly, uh, important time because when we look at every company clamoring their board, um, or being clamored by their board, by their executives, get the AI projects out, right? The number thing, number one thing holding them back is AI ready data. So when the board hears, yeah, ah, we got to clean up our data state in order to be able to really embrace AI.
Speaker B: Right.
Speaker D: Well, now you finally have that impetus of, okay, we'll go get the data state estate.
Speaker B: Right.
Speaker D: And when you think about, you know, what it's going to take to pull this forward, what we do know is I think I touched, touched this on in a conversation with you earlier today. We're in experimentation, right? As we get into production, it's not going to be just about what open source model did you use and what did you tag, what piece of hardware did you bring in? It's going to be like, can we support this at scale?
Speaker B: Right?
Speaker D: Is this something that our existing IT team can take on?
Speaker B: Right.
Speaker D: Is this a service that we can, can effectively afford on a monthly basis and is the return better than the cost?
Speaker B: Right.
Speaker D: We are going to get there. And when we look at, you know, everpure in our history and driving efficiency and operational efficiency, right. That is going to continue to be the resounding, uh, differentiator is that we will deliver the platform that is operationally efficient and sustainable at scale. And, you know, I think that's going to make a big difference on the success of these projects. Uh, I love it.
Speaker B: Folks, you've been listening to DM Radio.
Other episodes covering the same guests and topics, from across The B2B Podcast Index.
Episode 426: Claude Cowork vs Microsoft 365 Copilot CoworkMicrosoft Cloud IT Pro Podcast · on GDPR compliance83 / 100
Episode 105: From AI Idea to Production RealityValue Driven Data Science · on Data classification82 / 100
Crown Jewels In, Crown Jewels Out - The Hidden Risk of AI with Devan Shah (IBM)ShipTalk · on Data classification82 / 100
Data, AI, and Knowing When to Let Go - with Tommy CotterDefinitely, Maybe Agile · on GDPR compliance81 / 100
Unlocking Agentic AI: Building Enterprise AI Agents with Salesforce’s Patrick HeinenFintech Files · on GDPR compliance81 / 100
The Checklist Your Deck Is Missing ft. Jeff McMillanInvisible Machines podcast by UX Magazine · on Knowledge graphs77 / 100DMRadio Special: Conversations from We Make FutureThe 'Death' of Analytics VendorsAI-Native Innovation: Synthetic Market Research and the Future of SaaSTechnical Debt: The Hand You Hold...Training the Machines: The Race for High-Value AI Data