Amperity Reimagines Data and Developer Workflows with AI - Ep. 271

0:10Hello, and welcome to the NVIDIA AI podcast. I'm your host, Noah Kravitz. From agentic AI to vibe coding, our guest has been at the forefront of showing how AI-powered systems can empower engineers while delivering measurable business value. Derek Slager is co-founder and chief technology officer of Amparity, a company that's redefining how enterprises use data to better understand and serve their customers. Derek's here to talk about his journey founding Amparity, the tools his team is building, like Chuck Data, an AI agent for data engineers, and his perspective on how AI is reshaping not just the enterprise landscape, but the developer experience itself.

0:50So let's get to it. Welcome, Derek, and thanks so much for joining the NVIDIA AI podcast. Thanks, Noah. Great to be here. Great to have you. So let's start at the beginning. Tell us a little bit about M-Parity. Yeah, so M-Parity is an AI-powered customer data cloud, and we help brands unify, understand, and activate customer data at scale, which is a lot of things. And, you know, we're really particularly focused on data quality, right? Because we're big believers that better data equals better results, right? You're funneling that through agentic use cases or otherwise. And so we have a lot of great capabilities to help people all over a consumer business take advantage of that data, but it's all about getting the data right.

1:34So what inspired you to co-found the company and how does your background, your own journey as an engineer kind of shape the direction of Imperity? Yeah, for sure. So we started in 2016, right? Which, you know, I sometimes refer to as like the false start AI era. There was a lot of, you know, a lot of excitement about kind of, you know, deep neural networks and a lot of other kind of innovation happening in the AI space. But certainly it was, you know, nothing in comparison to the current AI wave. But nonetheless, right, like that was kind of, you know, it was on a lot of people's minds. And, you know, what we observed in kind of researching Amparity was almost every consumer brand had a project to unify all their customer data.

2:19And yet we couldn't find a single one. And we tried pretty hard. We couldn't find a single one that said, yeah, yeah, we solved it. Right. Despite all that effort. And we just found that ridiculous. And so, you know, what inspired us to start Amparity was really the opportunity to help all these people who were trying and failing to solve a problem that was really important to their business actually succeed in doing that. And of course, like the key point there was, you know, all these people were smart. They were trying really hard to solve it. Right. So clearly the existing tools were insufficient because if they were sufficient, people would be achieving more success.

2:56And so, you know, the big idea of Amperity was really how do we bring AI to this problem? And we felt like, again, this is through a 2016 lens. We felt like AI was sort of a perfect fit because a lot of the reasons that people were struggling with this problem was that they were trying to make the data perfect. Right. And, you know, AI kind of lives in that space where, you know, maybe it doesn't give you a perfect answer, but it gives you a really high quality answer. Right. And when we applied it to that problem, it worked really well. And so, you know, we've kind of carried that forward, you know, into the modern AI era.

3:28And, you know, that's allowed us to kind of, you know, bring acceleration to a whole bunch of other things, which I'm sure we'll talk a bunch about today. Maybe before we dive in a little deeper, can you give us kind of a high-level overview of some of Amparity's offerings, product, services, how you work with customers? Yeah. So we have, you know, kind of a core capability around helping users organize and unify data. So a big component of that and kind of our initial R &D innovation was around stitching. And that's about kind of finding all the Noah Kravitz's, you know, all over these different, you know, data sources.

4:01A typical Amparity customer will have 25, maybe 30 different, you know, data inputs that contain information about customers. And so, you know, we're looking for all the Noah Kravitz needles across those various haystacks. So it's a, you know, it's a challenging at scale problem. And so on the other side of that, then we can sort of utilize the outputs of Stitch to create a really high-quality customer 360. And that allows people to kind of ask questions and understand their data better. And so we have a bunch of tools that allow people to kind of introspect and slice and dice. And many of those are powered in 2025 by modern conversational interfaces, which is really exciting because it empowers many different people from the organization.

4:46And then we have a lot of capabilities to take the data and take the kind of different segments of that data and then get it out into the ecosystem. Right. There's thousands of marketing tools and thousands of customer support tools and thousands of all these, you know, kind of tools that that all are hungry for better data. So Imperity is really focused on getting that data right and then activating that ecosystem. And we have a number of capabilities, you know, from kind of campaigns with A-B testing to journeys and many other kind of, you know, pieces of capability to sort of allow that to kind of integrate with the ecosystem.

5:22Right. Fantastic. One of the things, one of the, I don't know if you'd call it a tool per se, but one of the things that's getting a lot of attention currently in the world of AI is AI agents, agentic systems. We've talked a lot about it on the podcast, talked about it from the NVIDIA perspective of building the blocks that allow folks like you and Imperity to build the tools to serve your customers. What does agentic AI look like from your point of view when you're out working in real life business situations and developing applications? What does agentic AI mean right now? Yeah. That puts you right on the spot.

5:56Yeah. How many definitions of agentic AI? And I'll admit sometimes, you know, I'm a little flexible with my definition. Sure, sure. I can rephrase and say, how do you approach it? What do you think about what's the lens? Yeah, sure. So, you know, I'll start with my own definition. And I've heard many definitions that are very complex. You know, my own definition is very simple and I think pretty expansive to the definition of agentic, which is it's just a program where the LLM can impact control flow. And so, you know, impacting control flow might be retrying, might be calling a tool, it might be interacting with another agent.

6:30And so, you know, it's pretty open-ended in terms of what could fit that definition of agentic. The reason I use that as the definition is because, you know, when the LLM is dictating control flow, that's very different from a traditional program, right? A traditional program that makes a call to an LLM is still a traditional program and your eval metrics and other things look awfully familiar, you know, relative to, you know, some other things that you would do. But once you kind of, you know, in essence, let the algorithm take the wheel, things change a lot. Right. And it changes kind of how you build systems, how you monitor systems, how you evaluate systems.

7:01And more importantly, it changes what those systems can do, you know, for your business. And so, you know, obviously we're working a lot with that technology, but also I have the opportunity to talk to a lot of customers who are, you know, working with that, you know, technology in their own organizations in various ways, oftentimes using, you know, kind of imperative data assets. But we see use cases, I think customer support is one of the biggest use cases. And I think, you know, one of the interesting things is, you know, a lot of people think, yeah, of course, like, you know, agentic for customer support, that makes sense as a cost-cutting exercise, but that's terrible customer experience.

7:37I think what I've found talking to customers about it is their customers like it. And so even when they're kind of fully eyes wide open, hey, I'm interacting with an agent, there's just, it actually yields a good customer experience. So, you know, I think that's one I'm seeing a lot, you know, analytics, introspection, kind of understanding the business. There's a lot of agentic use cases that I see there. And many customers are experimenting with bringing agentic use cases to the end consumer. And I don't see quite as many of those. But ultimately, I think that's, you know, what we're going to start to see a lot more of as time goes on.

8:11I think, you know, I'm seeing a lot of those kind of up close and personal in prototype stage. Sure. A lot of them are a little bit stuck there at the moment as they kind of work through that kind of, you know, long tail of challenges. But I think over the next year, we'll see a big increase and that becoming a prominent component of customer experience for consumer brands. Right. How do your clients, the users you work with, how do they feel about letting the LLM take the wheel? Is there, I'm sure there's a mix of things, but what's the sense you're getting and how are companies adapting? It's a really interesting question.

8:45I would have given you a boring answer six months ago, you know, but like it's amazing how much it's changed in six months. And I'll start by saying like, there's huge variants, right? And I don't always know, right? It just kind of depends. Probably the major variable there is kind of, you know, how much AI experimentation the leadership team has done. You know, because I think I think once they get it, it flows down really, really quick. Yeah, yeah, yeah. But but yeah, I'll have conversations with people who are like, yeah, we want we want everything in our organization to work this way yesterday.

9:16How fast can we get there? And I work with other people who are like, yeah, we're still kicking the tires. Right. And so, you know, it's it's incredibly, you know, diverse, the attitude. But boy, the trend line is clear. Right. And in the in the last six months, it's it's amazing how many people have gone from kicking the tires to all in. And I think after another six months passes, I think we're going to see probably, you know, 80 % of people are going to be in the all-in mode. It feels like time has sped up over the past few years. Yeah, like extremely. Yeah. Yeah. Can you share a story of delivering an agentic-based, an agent-based experience application for a large client, a Fortune 500 client, something like that?

9:57And kind of tell us about how it went, if there was a challenge or kind of a surprise to overcome. and, you know, how the customer responded. Yeah, absolutely. So, you know, it's interesting. I'll maybe give an imparity-centric story since, you know, those are the ones I'm closest to. We started, like a lot of companies, are, you know, LLM in the product journey with text to SQL, you know, and that was kind of, you know, in a part of the product surface that was, you know, for people who already knew SQL and it was, you know, kind of a nice augment of the experience. But then we challenged ourselves.

10:32We said, hey, what if we built something that was for people who don't know SQL, right? People who want to ask questions of their data and better understand and make decisions about their data, but don't have that skill set. Like, what could we build for them? And so, you know, we built a capability called AMP AI. And to the point about a curveball, I'm going to admit I was a little skeptical because what I thought was we're going to build the surface. People would try it a little bit and then be like, well, I'm not sure if the data is right. So I'm just going to go back to what I did before, which is ask one of these SQL people to give me the answer.

11:04And I suppose I could say I was pleasantly surprised because what we found is when we gave people AMP AI, there was a lot more energy for people to kind of really get in and introspect the data than I had thought. And it was surprising to me how much that kind of SQL interface between the user's question and the answer was a barrier. And so, you know, when we created the experience that was for them, I was amazed at how many people used it and not just used it a couple of times to try it. But like, you know, we did a cohort analysis and we looked at usage over time. And, you know, people who use the product, you know, a little bit once they started kind of using the AI interface went up to using it a lot and stayed there.

11:46And so it was really fascinating to me to see, you know, a whole bunch of people without a very technical skill set all of a sudden become more data informed. Right. And all we had to do was put this, you know, little surface on top of the data. It was really, you know, kind of a great surprise and an exciting surprise and certainly has given us the confidence to, you know, lean more into these AI use cases. Yeah, no, that's that's what you want to see. Right. You get that adoption and it sticks. For sure. Yeah. Along those lines of opening up more technical capabilities, possibilities for the less technical people, let's talk about vibe coding.

12:23Yeah. Maybe nine months, 12 months ago, whenever it was, I saw a video, and forgive me, I can't remember the name of the company, where somebody, I think they were doing it on a mobile phone, said to the phone, create an app that does XYZ. And we watched as on the screen, it spit out the code and ran the app. And it was, oh my goodness, like what's going to happen to the world? Now we fast forward. And the other day I asked a coding tool like this to recreate one of the arcade games I grew up with. And, you know, as you alluded to before, the results might not have been perfect, but it sure wrote a whole lot of code that I couldn't have written myself.

13:00This is amazing. It opens up possibilities. And this is more from what I hear from people who are more versed in coding and security and things like that than myself. it also opens up some potential issues. Talk to us about vibe coding. What does it mean to you? What does it mean for, you know, engineers, people who are actually steeped in what they're doing? And how is this, you know, changing the landscape now and going forward? Yeah. So, you know, vibe coding, I love that it has a name now. Right. And I didn't even explain what it meant. I just assumed at this point. I know. Yeah. I think if they're listening to the NVIDIA AI podcast, they're up on things.

13:40Yeah, I think that makes sense. Yeah, I think vibe coding is has, you know, impacted everybody. And you give the examples of kind of these, you know, increasingly impressive one shot examples where you give a single prompt and something pretty amazing comes out the other side. Right. That's maybe one category of vibe coding. And then I think like vibe coding for the, you know, professional software developer set looks very, very different. Right. And I think part of that is, you know, people are working with very large code bases and and, you know, just the properties of the system and what you get out are very different than the big one-shot thing where it's generating lots of code and things need to integrate with an existing workflow.

14:19I think what's very interesting is Vibe Coding changes the workflow of programming. And I think this is something that, you know, I've really seen the people who in particular get tons of value out of Vibe Coding are not just doing what they've always done, but doing a little bit faster because they're involving in LLM, they've kind of re-architected how they write code. And specifically, a lot of times they're taking advantage of asynchronicity. So, you know, some engineers will launch 10 different, you know, LLM-based processes at the same time. And then you may go off and do something that, frankly, if you were sitting there watching them code, would look exactly like what they've always done.

14:59And then, you know, at the end of that session, they're going to go back in and check on the results of those agents. And, you know, they might take three or four of those and just say, ugh, garbage, delete it, you know, or just start from scratch, right? And then they might take a couple more of them, make a few refinements, you know, check in on the code. And then there's a couple others, maybe it just nailed it, right? And those will kind of go straight through following a quick review. But if you think about it, like, you know, they've essentially, you know, put kind of, you know, 10 additional engineers on their desk.

15:29And the engineers have some interesting properties, right? Sometimes they are really impressive and sometimes they're astonishingly terrible, but they're, you know, they're effectively free, you know, the cost of some of some token generation. And so I think, you know, there's very different workflow that comes with Vibe Coding and, you know, obviously that has a, you know, kind of profound impact on what a team can accomplish. I think the other interesting thing about Vibe Coding is we're increasingly starting to see people with different skillset profiles contribute directly to the product, right?

15:59So, you know, rewind a year. If you're a designer or you're a product manager, you're probably not impacting production code except by influence, right? But that's changing, right? You can say, I think this should be this way and you can go and try it. And, you know, we've built a lot of infrastructure here to enable these vibe coding use cases to be accessible to a lot of people. So you can try something and, you know, it'll put it in an environment where you can test it so you don't have to spend, you know, two days setting up your local machine. And then you can say, yeah, this is good. And then have an engineer review the code and put it straight into production.

16:32When it comes to things like quality assurance, security, other concerns around putting code into production, is the answer, and I'm speaking from my own experience kind of as a content creator and using AI systems to help in a very kind of similar way to what you're talking about. Just my words are content, not instructions to a machine. Is the answer as simple as having the right human in the loop to review the output? And like you said, know how to incorporate this into the workflow, launch 10 processes, review them, discard, use, adapt. Is it that simple and sort of at a high level? And what are some of the things kind of maybe more nitty gritty that you think about when it comes to vibe coding, quality assurance, security, you know, and working with these large code bases?

17:22Yeah. You know, it's interesting. I think there's a lot of discussion about this. The way I see it, it doesn't change much, right? And what I mean by that is, you know, human beings are typically the, you know, the weak link in a system, right? I think building systems that are robust from an infrastructure perspective, from a scale perspective, from a security perspective requires you to be resilient to error. And when designed intentionally, you expect error. And so, you know, I do think that, you know, good system design looks the same, you know, when when vibe coding is involved versus not.

17:56And arguably, it's more important, right? Because, you know, the impact is larger. So, yes, I think you still need accountability mechanisms, right? Like the, you know, at the end of the day, you know, we were very clear, you know, engineers are responsible for what comes out. And, you know, just like rewind 10 years, somebody might have copied and pasted, you know, some code from Stack Overflow. They're accountable, you know, for that, whether they use that as a shortcut or not. This is exactly the same in Vibe coding, at least in terms of kind of creating the surfaces that we have. Now, there's security concerns, obviously, that are that are new and different, you know.

18:31And, you know, when we do, you know, penetration testing in 2025, you know, we're looking at things like prompt injection and other things, you know, in terms of the operational surface. But if we really kind of zoom in on the vibe coding era, it's like, yeah, absolutely. The humans are still accountable for sure. Right. I'm speaking with Derek Slager. Derek is co-founder and CTO at Imperity, a company that is doing, I think it's safe to say, a variety of things to help their customers, other companies, users understand and make the most of their data. We've been talking about engineering the developer experience.

19:07I want to ask you about one of Imperity's tools, Chuck Data. Tell us about Chuck Data. Yeah. So Chuck Data is something that we launched. And, you know, it's such an interesting extension of our Vibe Coding conversation because, you know, we were, you know, experimenting and understanding and learning the best ways to, you know, extract value out of the various kind of interaction patterns around Vibe Coding. And then we asked ourselves what seemed like sort of an obvious question, like, why is this only for software engineers? Like, you know, and, you know, we looked at some of the things that, you know, our customers or even some of our own teammates were doing related to customer data engineering, right?

19:45Where, you know, they're pushing around, you know, big, big mountains of SQL or, you know, trying to kind of manually build, you know, workflow coordination patterns or, you know, maintaining, you know, huge kind of DBT code bases. And we thought, whoa, there's an opportunity here. And so the idea behind Chuck Data is, hey, what if we created a Vibe coding tool for data engineers that are specifically focused on customer data use cases? And so, you know, Chuck was kind of packaged up to be a very, very easy way for people to, you know, in essence, have these kind of like, you know, one shot experiences that you described earlier in sort of the classic software engineering sense, but for customer data engineering use cases.

20:26Are there competitors? Are there other, when you were building Chuck Data, were there reference points? And less asking about what those were, but what's different about Chuck Data? What sets it apart? Yeah, it's such an interesting question because the thought I had, you know, as we were kind of forming Chuck and no doubt we took a lot of inspiration from Cloud Code, you know, that's definitely kind of, you know, number one in the Amperity rankings of Vibe Coding Tools. And so we took a little bit of inspiration from that, though Claude, of course, is aimed at more traditional software engineering.

21:00But I expected, you know, there was a point in time where I was like, OK, let me go find what other people are doing. I'm sure other people have thought of this. And I couldn't find them. You know, perhaps they, you know, they did exist and hadn't made their way up the, you know, SEO and GEO rankings. But I literally couldn't find them unless I was trying that. Which is always kind of an exciting and scary feeling. That's exactly what it is. Yeah, yeah. I was like, hooray. And then I was like, uh-oh. What am I missing? Yeah. Are we too early? Yeah, right. And so, yeah, no, we leaned in and I think, you know, in parity, we always really think about things starting from first principles, right?

21:34And it's like, hey, we can help people really accelerate some things that we know they're going to be challenged with. And of course, that's valuable. So let's lean into that. And yeah, I'm sure the competitors will follow eventually. But yeah, I was a little bit surprised to have trouble finding. Where does the name come from? The name, it's inspired by, you know, one of our very early engineers who, you know, kind of did a lot of work on some of our, you know, initial R &D. And so we kind of wanted to, you know, have a little nostalgic throwback. Yeah, awesome. Yeah. Very cool. So talking about vibe coding and AI-powered workflows and the experience of a developer and, you know, a new developer kind of learning the trade and getting their feet wet, an experienced developer, as you said, kind of rethinking their workflows and how they do things.

22:22There's concern in some circles that, you know, again, I'll relate it back to my own experience, similar to how there's concern that students, adults, professionals for that matter, are using AI to do their writing for them. Concern around Vibe Coding, AI-powered tools, kind of enabling programmers, if you will, to skip the process of really learning to understand how the code works, how to structure applications, how to do all those things. What's your take on that? Are there risks of over-reliance? Is it just kind of where the wind's blowing and will adapt? You know, what's your take and what's Amparity doing to make sure that, you know, programming doesn't turn into just hitting tab repeatedly and then hitting ship?

23:07Yeah, the concerns sound familiar to me. You know, I've been around a little while and, you know, my career started in the late 90s. And, you know, you know, a lot of the conversation in the late 90s was like, ah, you know, these kids today with their high level languages, right? They don't even, you know, they don't even know how to do pointer math. And then, you know, and then it sort of moved on to all these kids with their IDEs, with their fancy autocomplete, these kids with their, you know, garbage collection language. Back in my day, you know, we used to manage memory ourselves, et cetera, et cetera, et cetera.

23:40And so, you know, it sounds a little bit like the grumbly old person rant, you know, totally a little bit. But sure, like Vibe Coding has a bunch of properties that allow you to create a bunch of code quickly. And I think, you know, there's many prominent examples that aren't hard to find of people who didn't understand what was happening there. And then really bad things happened, right? And so in some sense, back to maybe the accountability point from before, nothing really changes, right? You've got some tools. It allows you to work differently, to work faster. But like you're ultimately accountable.

24:10And it's not optional, you know, to understand how that code works. It really isn't. And I think even, you know, if we look over the horizon and imagine a step change in the model and, you know, kind of more, you know, agentic sort of verification and validation workflows, it'll get easier, it'll get faster. But at the end of the day, you know, I think society is built around the notion that, you know, humans are going to have accountability for the things they do, right? And, you know, I don't really think that changes. And so I think it's a great new set of tools. I celebrate the great new set of tools.