The enterprise analytics landscape is undergoing a seismic shift as organizations rapidly transition from traditional applications to autonomous AI agents. Leading this transformation is StarTree, whose recent announcements around AI-native analytics capabilities signal a new era for real-time data intelligence.

In an exclusive interview on TFiR’s “An Eye On AI,” Peter Corless, Director of Product Marketing at StarTree, revealed how the company is empowering enterprises with features including Model Context Protocol (MCP) support and automated vector embedding hosting, alongside their innovative Bring Your Own Kubernetes (BYOK) solutions.

The Autonomous Agent Revolution

StarTree‘s real-time analytics platform, powered by Apache Pinot, was originally designed for high-concurrency, user-facing analytics. However, as Corless explains, the emergence of autonomous agents is creating unprecedented scale requirements. “If there are 8 billion people in the world, each of them could have 10, maybe 100, or even 1,000 agents,” he notes, painting a picture of exponential query growth that traditional databases simply cannot handle.

This agent-to-agent communication revolution represents a fundamental shift in how enterprises think about data access and analytics infrastructure. Where traditional applications served human users, the future will see fleets of specialized AI agents making millions of concurrent queries to support everything from customer service to complex business intelligence operations.

MCP: The Missing Link for Enterprise AI

StarTree’s implementation of Model Context Protocol (MCP) addresses a critical challenge in AI infrastructure: connecting large language models to authoritative data sources without compromising accuracy. Unlike traditional RAG (Retrieval-Augmented Generation) implementations, MCP provides a standardized approach that maintains context across extended conversations and even between sessions.

“MCP is going to allow LLMs to maintain state back to a day, a week, a month, a year,” Corless explains. This persistent context capability is crucial for enterprises deploying specialized AI agents that users may interact with irregularly but expect to remember previous conversations and preferences.

The protocol also serves as a safeguard against AI hallucination by ensuring that agents access real-time, authoritative data for critical information like pricing and availability. This is particularly important as enterprises face increasing scrutiny over AI-generated responses, with several high-profile lawsuits highlighting the risks of unchecked AI outputs.

Vector Embeddings at Scale

While many organizations rushed to implement vector databases in 2024, StarTree is addressing the often-overlooked computational challenges of vector embeddings. Traditional approaches require expensive batch processing jobs, often involving Apache Flink and GPU resources, creating bottlenecks that contradict real-time analytics requirements.

StarTree’s solution integrates vector embedding generation directly into their ingestion pipeline, eliminating the need for separate batch jobs and enabling real-time vector indexing. This is particularly valuable for use cases like observability, where teams need to perform vector similarity searches on complex traces and logs as they’re generated.

“We want to build that in from the get-go, so it’s a much more automated and scalable process,” Corless emphasizes, highlighting how this architectural decision positions StarTree ahead of competitors still relying on external processing pipelines.

Kubernetes Flexibility Without Compromise

StarTree’s evolution from SaaS to Bring Your Own Cloud (BYOC) and now to Bring Your Own Kubernetes (BYOK) reflects the enterprise demand for greater control over their analytics infrastructure. This progression addresses multiple enterprise concerns: cloud credit optimization, data governance requirements, and the need for air-gapped deployments.

The BYOK approach is particularly significant for organizations with strict data sovereignty requirements or those pursuing cloud repatriation strategies. By enabling fully managed Apache Pinot deployments on customer-controlled Kubernetes clusters, StarTree delivers enterprise-grade capabilities while maintaining customer control over orchestration and provisioning.

“This gives you a fully managed service, but again, in an air-gapped environment,” Corless notes, expanding the addressable market to include organizations with the most stringent security requirements.

The Observability-AI Symbiosis

Perhaps most intriguingly, Corless highlights the emerging symbiotic relationship between AI and observability systems. While AI-powered analytics can enhance observability platforms, there’s an equally important need for observability of AI systems themselves. This “yin and yang” relationship addresses critical concerns about AI drift, hallucination prevention, and performance monitoring.

“AI making better observability systems, and observability making better AI systems,” Corless explains, suggesting that enterprises need to adopt a holistic approach rather than treating AI as a silver bullet solution.

Preparing for the Real-Time Future

As enterprises navigate this transformation, StarTree’s upcoming Real-Time Analytics Summit promises to deliver practical insights from practitioners across industries. With over 5,000 registered attendees globally, the virtual event will showcase real-world implementations and lessons learned from organizations at the forefront of this revolution.

The summit’s focus on democratizing real-time analytics knowledge reflects a broader industry recognition that successful AI implementation requires more than just technology—it demands a fundamental rethinking of enterprise architecture, data governance, and operational practices.

StarTree’s comprehensive approach to AI-native analytics, combining MCP support, automated vector embeddings, and flexible Kubernetes deployment options, positions the company at the intersection of several critical enterprise technology trends. As organizations continue their digital transformation journeys, the ability to provide real-time, context-aware analytics at unprecedented scale will likely determine which enterprises thrive in the age of autonomous agents.

For organizations evaluating their AI and analytics strategies, StarTree’s recent announcements serve as both a roadmap and a reality check—highlighting the infrastructure investments necessary to support truly intelligent, autonomous systems while maintaining the reliability and governance standards that enterprises demand.

Transcript

Swapnil Bhartiya: StarTree has recently made significant announcements around AI-native analytics capabilities, including support for Model Context Protocol and vector embedding hosting, as well as launching their Bring Your Own Kubernetes solutions. As organizations increasingly shift towards autonomous agents and conversational interfaces, today I have with us once again Peter Corless, Director of Product Marketing at StarTree, a company empowering enterprises with actionable intelligence through real-time data analytics. Peter shares insights on how these innovations are reshaping enterprise architecture and decision-making. Whether you are a data professional, AI enthusiast, or business leader navigating digital transformation, this conversation promises valuable perspective on the future of intelligent systems and real-time data analytics. So let’s go and talk to Peter. Peter, welcome back to TFiR and An Eye On AI.

Peter Corless: It is always good to be back on your show, and thank you for having us today.

Swapnil Bhartiya: It’s my pleasure, and there are a lot of announcements that you folks made, and we will deep dive into all of those. But before that, just quickly refresh memories of our viewers: What is StarTree all about?

Peter Corless: StarTree is the real-time analytics platform powered by Apache Pinot. It is a service that’s designed for fast, highly concurrent queries per second. We’ve been talking in the past about user-facing analytics, but I believe that everybody who’s watching what’s going on with the revolutions in technology today is seeing that we’re not just having applications facing users, but towards agents. And I think the vital thing that people have to consider now is that if we have 8 billion people in the world, every one of those people in the world could have 10, maybe 100, maybe 1,000 agents in times to come. And so if we believe that query concurrency is very high today, in the future, when you have these autonomous agents driving all sorts of applications—purpose-built applications—you could see that there could be this exponential growth of greater orders of magnitude of many agents doing very purpose-built things, then talking with other agents. We see technologies like MCP—Model Context Protocol—to tie applications to systems, but we also know that there’s an agent-to-agent revolution also going to happen, where agents will be talking with agents.

Swapnil Bhartiya: Excellent. Thank you so much. And when I look back, if somebody talked to you even in 2020, or 2021 after COVID, and somebody would say, “How do you think this whole architecture will change?” before ChatGPT was announced, we wouldn’t even be able to envision this world. So sometimes, when I talk to a lot of folks, “How do you see yourself in five years?” I don’t like to ask those questions, because we don’t know what will happen. You know, tomorrow, Kubernetes, all these containers, the way they revolutionized things. So it’s incredible. So what I also want to talk about is that, as you’re talking about the emergence of these agentic AI and as organizations will start—not will, they are already embracing them in production—you folks just also announced support for Model Context Protocol and vector embedding hosting. Now, looking at the way you explain what StarTree does, talk a bit about how do these features that you folks just introduced enable real-time, agentic AI, and why does it matter to enterprise intelligence today?

Peter Corless: We’re not going to be the only organization announcing MCP. Everybody will, and if they haven’t done it, they will be doing so soon, because MCP—for people who aren’t familiar—MCP is this protocol by which you can attach an agentic AI application to a backend system, like a source of truth, an Apache Pinot, but it could be any other kind of database. So MCP is going to be very important for the retrieval-augmented generation systems we’ve been hearing about. RAG has always been a concept, as long as there’s been LLMs, that the LLM would be allowed to do the on-the-fly generation of context based upon its own training models. But it wants to go to a source of truth to get things like pricing information or availability—things that the AI should not be coloring outside the lines of. We’ve seen the lawsuits of companies where the LLM has hallucinated answers to questions. And what we’re trying to do now is we’re trying to say, “No, we have terabytes, we have petabytes of data of actual source-of-truth information, and we need to open that up to the LLMs,” and that’s what MCP is going to allow for in a much more standardized way. Again, you’ve been able to build RAG applications in the past, but there was no standard around it. With MCP, keeping context for a user and what that user cares about is going to be much, much easier.

Swapnil Bhartiya: If you just look at the way organizations are shifting from traditional applications to autonomous agents, what kind of challenges—it could be architectural, it could be also operational—that enterprises face? And how are you folks helping? As you said, you know that you folks are not alone. A lot of folks will join. But just let’s talk about the problem area and how you folks are helping organizations.

Peter Corless: Yeah, now again, because you can hook up any sort of database on the backend of an AI, the issue is, what happens when you’re trying to do that at scale? Where Apache Pinot came from, where StarTree got its founding, was at LinkedIn, when they were trying to answer questions. When LinkedIn at that time was only—only—100 million users, and that’s when Pinot was born, was in that context of any one of those 100 million users at any time can ask a question with an agentic AI system. Now LinkedIn is already 1.2 billion, and they’re still powered by Apache Pinot. So when you have that kind of scale of concurrency, you need to have a system that’s more than just something running like in Docker on a laptop. You need to have an enterprise system that can handle the workloads of potentially a billion-plus people, and so that’s where Pinot is going to have this key differentiator, in the sense that sure, everybody will have MCP, but that doesn’t change the fact that the backend database needs to be scaled to meet that kind of workload, and Pinot began at that scale. So it’s going to be much better positioned to be able to handle the kinds of scales you’re seeing out of MCP-based applications.

Now, another thing you talked about in terms of scalability, and this has to be considered, and you mentioned it a little bit, so let’s go into it, is vector embeddings. Apache Pinot has supported vector indexing—again, that was the last thing that happened, was that everybody was launching into vector indexing, right? That was last year’s whole trend was that everybody became a vector store suddenly. But now, when you start getting into it, you realize that vector embeddings are not cost-free. There’s a lot of compute that you have to throw at it. In fact, not CPUs, but GPUs. There’s a lot of work in getting those vector embeddings. And the problem has been is that there needs to be a lot of infrastructure and infrastructural integration to be able to get vector indexing, getting those vector embeddings into the database. And so what we’re doing here at StarTree is we’re going to make that process a lot more automated and, again, built for scale.

In the past, what a lot of people were doing is that they were getting vector embeddings with batch jobs, and that’s an anti-pattern. When you’re trying to do a use case like, let’s say, real-time observability, maybe you want to do a vector index on a log line or something—you know, some a trace or something. And you want to understand this trace, which is a complex thing, and you want to do a vector embedding against it to get a vector index for how is this—what’s the shape of this trace look like? And how is this trace similar to that trace? So to do those kind of vector embeddings, we’re going to be making it such that Apache Pinot can get those vector embeddings at scale, and that’ll be integrated into our ingestion pipeline. Traditionally, what you’ve had to do is you’ve had to do maybe like an Apache Flink, and you’d have to send it as a Flink job to get that embedding and bring that back into the real-time data pipeline. We want to build that in from the get-go, so that’s a much more automated and much more scalable process. So on the ingestion, you have the vector embeddings going into the vector indexing in Apache Pinot, and then on the outbound side, you have MCP. So we believe that with this, we have a much more end-to-end architectural concept of how we want to integrate with AI.

Swapnil Bhartiya: Now I also want to talk about Bring Your Own Kubernetes. Talk a bit about—of course, when we talk about Kubernetes, and when we talk about managed Kubernetes, sometimes you lose the flexibility, right? Because it kind of becomes opinionated, and sometimes organizations, they do want that level of control. I mean, that’s the whole point of using open source, right? So talk a bit about what was the driver behind this initiative, and how do you enable organizations to have the control and flexibility they want with Kubernetes?

Peter Corless: So let’s go back in history to how we got here as a cloud-based organization, right? Everybody began with SaaS, but with your SaaS, and let’s say, you know, we’re running an open source with managed services around it. It would be Apache Pinot running then with the StarTree applications like Data Manager and ThirdEye and other things that we built around Apache Pinot, and that would all be running in our Kubernetes on our instances. And that’s great for us as a vendor. But then the customers say, “Well, how about our cloud credits? How about our discount structure? We’d like to be able to run this in our cloud so it never leaves our data governance.” And so SaaS became Bring Your Own Cloud, and now it was Apache Pinot and the StarTree services, but still running in our Kubernetes, but it was at least running on your infrastructure, getting your discounts. And that was the first step. And this has enabled—Bring Your Own Cloud has enabled so many businesses to be able to, again, bring data back into their governance and to get the best bang out of their buck for their cloud spend.

The next step of this is to run Apache Pinot and the StarTree applications, but on your Kubernetes and then on your cloud infrastructure, right? So it’s BYOC plus BYOK. So it allows you to run in a much more controlled system. And as you’re saying, people get very opinionated about Kubernetes. They want to have control of that. They want to manage the orchestration, they want to manage the provisioning. And so for them, Bring Your Own Cloud and Bring Your Own Kubernetes are natural fits. But the other thing that Bring Your Own Kubernetes also allows is the fact that you don’t even need to run on the cloud. You could potentially be doing a hybrid cloud or an on-premises installation. And with Bring Your Own Kubernetes, you can be running anywhere. And you can be running those cloud-native kind of services that you get from StarTree Cloud—Apache Pinot, and all the wrappers around it, all the additional advantages you get from our services—but you can be running it on-prem if you so desire. So we believe that Bring Your Own Kubernetes is kind of an unlock for corporations, especially when they have much, much stricter data governance, data sovereignty issues, and again, they’re looking for cloud repatriation to get the most out of their infrastructure spend.

Swapnil Bhartiya: Before I switch to the next question, which is, once again, going back to MCP—anything else that you would like to add to Kubernetes? Or have you touched on some of the key points regarding—

Peter Corless: Yeah, I think that the way we’ve done Kubernetes—again, different companies have done Kubernetes in a lot of different ways—but I think one of the ways that we’ve done Bring Your Own Kubernetes would even allow for air-gapped environments, so to completely run autonomously. And I think that that’s going to be very vital for a lot of organizations that truly need that kind of privacy and security, that kind of ownership. And again, I think that that’s one of the things we’re looking forward to, is that it expands where people and how people can run Apache Pinot. In the past, you’d have to run Apache Pinot on premises, completely self-managed, and this gives you a fully managed service, but again, in an air-gapped environment, again, also deployed on Kubernetes for that ease of orchestration, especially when you’re running at scale.

Swapnil Bhartiya: Now I want to go back to MCP for a bit, right? As we all know that conversational querying is a major focus here, after listening to you. And of course, prompt, right? Prompt, wrong prompt. I mean, it’s a—you know, we used to say, you know, that Google will not give you the right answer if you don’t ask the right question. The question always used to be the most important thing, or in general, also that question is more important than answer. Talk a bit about, how does MCP make it easier for enterprises to deploy natural language interfaces to their data, and what impact do you expect this will have on business users?

Peter Corless: Yeah, I think that the key thing that we’ve seen with models is that they tend to lose context over the course of a conversation. But also think about this in your own personal implementation of how you might integrate with AI—you might want to ask the AI, “Hey, you remember what I was talking about last week? I was doing this thread for this new video I was working on, right, and I had all these questions. Where’s that document?” Right? MCP—Model Context Protocol—is now not just based upon that individual chat, but hopefully it’s going to be able, for the LLM, to keep context over your experience with it. So potentially it could be stateful back to a day, a week, a month, a year. And if you think about how people might have these agents set up for very specific things, and they may come back to it very irregularly, you need that state to be maintained. You need for it to remember where you were, just like, you know, when you watch a show on television, you want to be able to come right back to that episode, right at the place where you last were watching, and without context, none of that makes sense. You need to have that context so you can come right back into the conversation at the point where you last left it, and you don’t lose any cycles in trying to get the LLM primed back to where you last were in the conversation—it could pick up directly. That’s going to be vital, especially again, when it’s not just an agent. You’re going to be running fleets of them, each one dedicated towards different kinds of channels. And I think that these are the kinds of things that MCP is going to allow for, but it does mean that somebody is going to have to maintain state on this.

Swapnil Bhartiya: How prepared are organizations? Because every organization, they do want to take advantage of these technologies, and they all feel that, “Yeah, they are ready.” And, you know, AI does make things very, very easy. Actually, AI can actually walk you through how to use it, but still, what are some of the challenges that you see? Is it about the workforce, their teams already? Or they don’t know where the right investment needs to be made? They don’t know what is the right vendor to go to? I mean, what I want to—as you talk to customers or clients, what kind of apprehension, challenges you’re like, “Yes, they want it, but this is something that kind of becomes a roadblock.”

Peter Corless: I think that the first thing you need to be able to do is pop your own bubbles, right? I mean, there’s a lot of hype around MCP and AI and what it can and can’t do, but I’m taking a look at organizations that have been very realistic about the limitations of where we’re at right now. You take a look in the observability space, and the work at Honeycomb—they’re taking a look at, like, you know, how to actually use this effectively, but also the extents, where does it start falling apart? And I think that, you know, for StarTree, we want to be a trusted advisor for people who are doing this for the first time, and also listen to some of the people that we know, some of our customers. We take a look at Dialpad. They have been an AI-forward company for years now. They use StarTree, but they also have been doing their own work on AI, and we have a lot to learn from them. And so I think that this is a conversation we need to have, frankly and honestly in the industry, about the extents and limitations of what the systems can do. Again, we’ve been trying to beat our heads against RAG before MCP, and MCP is going to make it easier, but, but again, there’s lessons to be learned as we’re going into this revolution. It’s not a friction-free kind of thing, and I think the first thing we need to do is really be honest about limits and capabilities and the learning curve, because there is a significant learning curve in getting this done.

The other thing I want to mention is the fact that there’s also this kind of a yin and a yang, right, with this beautiful symbol of observability and AI, where you need to have AI to watch your observability data. And we’ve talked about AIOps for a long while, but now there’s also a need for observability of AI. How do you watch the AI to prevent things like drift and, you know, make sure it’s not hallucinating? And I think that those are the kinds of things that we’re going to see—is AI making better observability systems, and observability making better AI systems. And I think that these are the kind of conversations people need to have, a much more holistic sort of approach of not just saying AI is going to solve everything, but how do we ensure that AI doesn’t break the things, right? How do we make sure that observability is there to watch AI performing optimally? And so I think that a lot of people need to have that more holistic sort of strategy. And again, there are some companies to look at, and we’re working very closely with our customers that are AI-forward.

Swapnil Bhartiya: And when I look at or when I listen to you, and I talk to a lot of folks, and remember, in the cloud, or, you know, server, we used to have the pet and cattle analogy, so I think, yeah, AI is also more or less like cattle. It’s not an independent worker which is going to do things for you. It’s like cattle. And you still have to monitor, manage, and it will make a lot of mistakes. So yes, you do need to keep an eye on it very closely.

Peter Corless: So that’s—let’s get into that, because that’s a great analogy, pets versus cattle, and we’re not going to go too much into this, but maybe for another conversation, there are some AI that are going to be very much pets. AI that might be making life-or-death decisions in a healthcare service—that’s very much a pet, and there’s going to be other kinds of AI that are summarizing your emails, and, you know, taking care of all the cruft that builds up in your inbox and all the docs that you’ve forgotten about, and organizing them and indexing those—those can be cattle. But I think that we need to treat AI—we’ll need to determine which of our AI are the pets and which are the cattle, and not all are going to be created equal.

Swapnil Bhartiya: Awesome. Thank you. Now, in the beginning, you also mentioned that your schedule was busy and you’re getting ready for the Real-Time Analytics Summit. Talk a bit about this summit. For those who don’t know about it, talk a bit about StarTree’s presence at this event.

Peter Corless: RTA summit.com—you can go and register right now. Over 5,000 people have already registered. It is a global event. Because we’ve gone virtual, we’re seeing attendees sign up from everywhere, and I think that we wanted to be able to reach those audiences. When you have an in-person show at a hotel, only so many people can take those days off or get a visa even to attend it. With a global, virtual, free event, we’ve opened up real-time analytics. We’ve democratized the knowledge that comes from practitioners, our executives from not just StarTree, but also we’re going to have a speaker from Uber, we are going to have practitioners at every level, and our own engineering people are talking about, how do you build these kinds of new services? What do you—what needs to go into a Kubernetes software-as-a-service? What needs to go into doing these AI applications? So I think that this is going to be a fantastic opportunity for people to learn. I know I’ve been talking with a number of my peers, and they’re all excited about it. I’ve been watching the videos as we prepare for the event. These are going to be knock-your-socks-off kinds of talks. We, in fact, this year, we were so impressed by the talks we got that we actually are doing Real-Time Revolutionaries awards, because these are fantastic events. So if you want to see who’s on the driving edge of real-time analytics, this is the place to be. RTA summit.com.

Swapnil Bhartiya: Peter, thank you so much for joining me today. And of course, talk about these new announcements. And of course, Bring Your Own Kubernetes. And of course, the Real-Time Analytics Summit. Thank you so much for great insights. Always a pleasure to talk to you, and I can’t wait to talk to you again.

Peter Corless: I really appreciate it. Thank you for the time. Have a great day.

Swapnil Bhartiya: You too.