Oracle Data Deep Dive NYC | Tirthankar Lahiri, Oracle

Clips
More from Oracle Data Deep Dive NYC

Tirthankar Lahiri

SVP, Mission-Critical Data and AI Engines

Oracle

play_circle_outline Agentic AI era: multi-agent, multi-step agent actions and coordination

play_circle_outline Oracle embeds agentic AI into databases: Data-proximate control plane with reasoning and co-located data-centric agents

play_circle_outline Unified Memory Core: derive short-, medium-, long-term memory constructs from database

play_circle_outline Drag-and-Drop Private Agent Factory Meets Autonomous AI Vector Database for Enterprise Search

play_circle_outline Open standards support: Iceberg, Model Context Protocol, open agent specification interoperability

Info
Transcript

Tirthankar Lahiri, Oracle

Tirthankar Lahiri

SVP, Mission-Critical Data and AI Engines Oracle

In this interview from Oracle Data Deep Dive NYC 2026, Tirthankar Lahiri, senior vice president of mission-critical data and AI engines at Oracle, joins theCUBE's Dave Vellante to discuss how embedding agentic AI capabilities directly into the database is helping enterprises move beyond pilots to production-scale AI deployments. Lahiri explains how the transition from RAG-based chatbots to multi-step agentic workflows is reshaping data architecture requirements. He outlines Oracle's strategy of co-locating agentic processing with data to eliminate fragmentati... Read more

explore Keep Exploring

At a high level, what problems are customers asking you to solve with the shift to agentic (multi‑agent) AI, what new challenges does this era create, and what does it mean for organizations? add

What is the recommended approach to placing agentic processing relative to data so agents can act on current, consistent, real-time data? add

What is the Unified Memory Core, how does it map to short-, medium-, and long-term memory, what is new or different about it, and why does it matter to customers? add

How does your product enable both developers and non-developers to create and deploy AI agents and perform vector-based search over business data? add

Do open data formats (for example, Iceberg) and open-source standards affect how enterprises secure, manage, and share data for LLMs, and how is the open-source community addressing those concerns? add

bolt Powered by CUBE AI

Tirthankar Lahiri, Oracle

search

Dave Vellante

>> We're back here at theCUBE's NYSE Wired Studios at the New York Stock Exchange. I'm Dave Vellante. And we're going to talk about agentic AI innovations. This is Oracle Data Deep Dive NYC. Tirthankar Lahiri is here. He's the senior vice president for Mission Critical Data and AI Engines at Oracle. Thank you for coming in, sir. Appreciate it.

Tirthankar Lahiri

>> Thank you for having me.

Dave Vellante

>> How's the week going for you guys over at Javits? You guys got sent some deep dives. A lot of the engineering teams we're in today, right?

Tirthankar Lahiri

>> Yep. It was amazing. Very, very uplifting. Great.

Dave Vellante

>> Yeah, awesome. Well, anyway, this year we're seeing this shift toward agentic. We all remember the RAG-based chatbot days, make a request, get an answer, and then reasoning came in, which was a whole new era. Now, we're into this multi-agent, agentic era where customers are executing multi-step agent actions and you're coordinating with other agents. This puts a lot of pressure on you guys. You're embedding a lot of these capabilities into the data layer. I wonder if you could just at a high level, just to kick us off, what problems are you trying to solve that your customers are coming to you? What kind of problems does this new era create? And then, we'll get into what it means for organizations.

Tirthankar Lahiri

>> Oh, great. That's a great kickoff question. So, agentic systems are going to become the future application development. They're the present and they will be the future. They're the way people build enterprise applications. Many organizations are still struggling though to realize value from agents, because ultimately, agents are only as good as their data. So, this is why we have announced new agentic AI innovations by architecting agentic AI deep into the database. So, that will enable customers to move beyond pilots and fill POCs to full-scale production workloads. The three main themes we're following with this agentic journey is to allow customers to innovate faster with AI for data by building capabilities for agentic AI, such as Unified Memory Core, deep into Oracle Database. We also want customers to be able to avoid AI data risk. So, provide agentic processing that is safe and secure and trustworthy within the database. And lastly, this world is changing so fast. You can't lock yourself into a certain direction now, it's too early for that. So, our approach is to use open standards, open formats as much as possible to give the customer the full choice and flexibility of the entire landscape offerings that exist today. And these capabilities are all very flexible. They're available on all environments, on-premises, multi-cloud, hybrid cloud, and public cloud environments.

Dave Vellante

>> All right. Thank you for that setup. So, let's double click on that. When I was prepping for this interview, I was watching an interview that you did with John Furrier and he said essentially Oracle's effectively making the data layer the control plane or the control hub for agentic AI. And I said, "Okay, that's interesting." And I said to myself, "All right. What does that mean from an architectural standpoint?" Because there are organizations out there, technology companies that are saying, "Okay, we're going to be the orchestrator of agents." That's their layer. There are others that are saying, "Well, we want to be the knowledge graph," et cetera. You guys are architecting the entire stack from hardware all the way up to the applications and obviously the data in between.

Tirthankar Lahiri

>> Yes.

Dave Vellante

>> So, I wonder if you could elaborate on this idea of the data layer as the control plane and what that means architecturally.

Tirthankar Lahiri

>> Right. That's a great question. So, our approach is to try to bring agentic processing as close to the data as possible so that the agents can act on current, consistent and real-time data. There are basically two types of agents. There's reasoning-centric agents and data-centric agents. And agents are going to be a hierarchy or a graph. The data-centric agents are really best run co-located with data, that's what our approach is. We want to eliminate the need for multiple round trips, multiple database accesses. It's a bit like you want to take a direct flight, not multiple stops on the way to get to destination. And architecting agentic processing along with data access avoids fragmented or fractured AI. You get AI that runs again on clean, real-time, current data without the need to split your data in multiple repositories.

Dave Vellante

>> So, if I can follow up on that, I get that you want to have the proximity because of latency. Does it also reduce the number of hops that you have to take, as you just described, and why?

Tirthankar Lahiri

>> Yes, there are other vendors who've advocated for a diverse database strategy, where you use a graph database for graph interactions and a document for document interactions. We think that's the wrong model because that ends up with, first of all, a giant sprawl of these data stacks. It ends up also with inconsistency across the different data views. So, with Oracle Database, we support all of those data formats and presentations on the same underlying unified data. So, that way you can use whatever data model is appropriate for agentic processing without the need to move data back and forth between different systems. So, it's not just the latency benefits. It's also efficiency, it's real-time access, and it's a fact that I can view that same data in multiple formats depending on the need of the agent.

Dave Vellante

>> So, earlier we were talking to Juan Loiza and we were talking about Unified Memory Core. And the way I understood it is it's memory, it's short-term memory, long-term memory, humans have both. And I wonder if we could get your technical perspective on this. What is that? What's new and different about it? And why does it matter for customers?

Tirthankar Lahiri

>> It's a great question. I love this because memory for agents is just like human memory. We have short-term memory, we have medium-term memory, we have long-term memory or facts. So, agents work exactly the same way and you need different types of representations to encode those memories properly. In databases, data begins life in tables and rows and columns, maybe documents, but they're not the facts the agent actually makes use of. What we have built with Oracle is a unified core, where on that existing database data, you can derive the memory constructs you need for each memory use case. Sometimes you want associations and you want a knowledge graph. Sometimes you just want a factual representation of an event that happened. Sometimes you want to know that so-and-so was the most highly-paid employee in a department at a certain point in time. These are all derived ultimately from actual data. That derivation, if it's done in place with the actual data, is current, it's consistent, and it's fully secure. We call that the unified memory core for that reason, which is much more efficient than using multiple storage systems to represent the different kinds of memory. It's just like a person, the most effective person has everything in their head. They don't go to the library, to the bookstore, to various places to find information. They all have it. The best lawyer knows everything upfront. Oracle Database strives to be that unified memory core to enable the most efficient agentic processing.

Dave Vellante

>> All right. Well, that makes sense. And we always say it's not just the LLM. LLM's obviously super important. It's not just the agent, but it's the data that's going to give you that competitive advantage. The other observation, of course, is that we're now speaking to the data.

Tirthankar Lahiri

>> Yes.

Dave Vellante

>> Even with RAG-based chatbots, it was like, "How do you do this?" Or, "What is that?" Or, "Explain this." Now, it's, "Build me something." So, that's different. Now, as a business user, obviously developers, they get their hands full. They're doing a lot of the work. IT, they're doing their thing. But as a business unit, I'm going to want to interact with my data. How do I do that? Can I do that simply, safely without having to necessarily go through the IT department or the developers that are maybe backlogged, maybe they got other priorities. Can I directly go to the data and interact with it, build things as a citizen builder, if I will? And how do you do that? How do you enable that? How do you make it safe?

Tirthankar Lahiri

>> Right. We love IT and developers, of course.

Dave Vellante

>> Of course we do, yes.

Tirthankar Lahiri

>> We want to empower them. However, we want to make it easy for users to build agents as well. So, one of the offerings in that space, we're calling the AI database private agent factory that basically is exactly what it says it is. It's a factory for creating agents simply where the end user has full control over the agentic workflow of a ride decision are made in a certain way. They control that reasoning process. It's a very simple drag and drop based visual builder for agents that contains custom prompts. You can add your own prompt to it. It allows very quick and rapid agent assembly. We have just rolled out limited availability, an autonomous AI vector database that can enable super fast, very easy to access vector search against your business data, which is, again, endemic agentic processing. You need to find similar data when you're doing agentic reasoning and vector search excels at that. And autonomous vector database lets you consume that very easily. So, these two things combined, the agent factory and AI vector database will enable that seamless harmony between agentic processing and data processing for developers and end users.

Dave Vellante

>> Okay, thanks for that. I want to ask you about mitigating risks. So, when we're using chatbots, if it hallucinates, which it does, gives you the wrong answer. They're easy to manipulate with, "Oh, no, you're wrong," this, that, and the other thing. Okay, that's no big deal, consumer applications, et cetera. But when I have agents actually taking action on my behalf, that becomes very risky. It can make a wrong move. It could leak certain data. I mean, that could not only violate compliance edicts. I mean, if it's egregious enough, I could actually go to jail or experience huge fines. So, there's the compliance, there's that data risk, there's security. All of this changes with agentic and in this AI era. How is Oracle dealing with that?

Tirthankar Lahiri

>> That is a very, very relevant question today. AI data security, especially where we're sitting right now, is super important because this information is business critical. And yes, you can go to jail for divulging the wrong information. The problem is LLMs are statistical machines, even as they are getting better, but they will make mistakes and sometimes they could go horribly wrong and give you the wrong query. The problem we have today is in many systems, security is built in the application tier. And it's a bit like locking your valuables in your house by putting a lock on your front garden gate. It doesn't really secure the house because somebody could bypass the garden gate and going through the back door, back entrance. AI is an example of a user that can go into a system in many different ways. They may not use the application interface. They could just go to the database directly. And if you rely on application security, you'll bypass that entire lock. It's like the garden gate. I didn't even go to the garden. I went through the back entrance and I went into your house, I stole everything. The only way to solve this problem is securing data at the source. You have to secure your data at the data access layer and we call that deep data security, which takes into account not just what the operation is, but who the end user is. The fact that it's me, Tirthankar, accessing data. Even if I'm doing the access through AI, it has to limit the data shown to me as if it were me doing that data access. So, even if the query is a query that's malformed, it can't return data shouldn't show. That's really what deep data security gives you. And I think in this AI world, that's the only way to secure data. Application security is just a bad idea. Deep data security is like putting a lock on your front door, on each room, on your safe, so that if somebody does get into your house, they can't get anything anymore. So, this, I think, is paramount for agentic security.

Dave Vellante

>> So, just to follow up on that, I get the safe. That's a granular sort of, "I'm going to put all my passport, my money, my valuables, my watch, whatever, in that safe." What actually is that technically? Is it like a carve out of the database that has identity knowledge? What is that?

Tirthankar Lahiri

>> It's actually very simple. It is a policy that you define in line with data access. So, all data access in the database has to go through that policy, you cannot avoid it. The database makes sure that all accesses to the data have to go through that policy layer. And in that layer, you define who can see what, that Tirthankar can only see employees in its own organization, nobody else. So, even if the AI comes in and says, "I want to see every employees' salary," it won't show it to me because it's coming in as me and my policy limits my access to only my employees, as an example.

Dave Vellante

>> So, in that example you gave, I mean, a malicious prompt injection, how do you deal with... Let me ask you broadly. How do you deal with prompt injections and getting unwanted outcomes from a prompt that is inappropriate?

Tirthankar Lahiri

>> Great question. Same strategy, deep data security, because prompt injection is like trying to create a situation where the AI generates operations that are ill-intended. It's like saying, "Hey, by the way, that side window is open. Break in through that window." You can come into the window, but all my valuables are locked in my safe. You have to get to my safe to get to my data." As long as that's locked and bolted and secured, you can't get to it. So, prompt injection, same story. Deep data security addresses all data access, whether through AI prompts or whether through just malicious hacker interactions.

Dave Vellante

>> You've been spending a better part of this week with a lot of your customers here in New York City, and I want to probe on what you're hearing from them, particularly around this issue of governance and security. When I talk to organizations, there's this inherent tension between the business wants to move fast and it's the organization, whether it's the security team, the IT team, the audit team, everybody who's worried about compliance, having to protect the organization from themselves. Bad user behavior beats good security every time I often say. Are you hearing that tension and what other concerns are you hearing from customers?

Tirthankar Lahiri

>> Yeah. One of the concerns, especially in a place like New York, where there's so much mission-critical confidential data that cannot be shared with third parties, as customers are saying, "We can't share data with third parties." Even with us, even Oracle, they need to keep all of their data confidential within their own defined data perimeter, so they can't use public embedding services or LLMs. So, for those customers, first of all, we have the Private Agent Factory that guarantees your agentic interactions stay inside your data perimeter. The second thing we have is what we call the Private AI Services Container, we've developed as a sandboxed AI solution that you can use to do LLM processing. You can use it to use do vector embedding generation machine learning without sending data outside of your environment. So, if you're NYSE, you keep your data within NYSE with this container. It's a local large language model. It's a local vector embedding service, local machine learning models. You can access that service just like a cloud service using REST APIs, but no data leaves the perimeter. No data leaves a building. That's been a key requirement for us for many mission-critical customers. One more thing we've heard is LLMs, again, as I said, they're like smart teenagers. They're very smart. They know everything, but they don't really know everything and they make up answers when they don't know. That's the life of LLMs. You can't bet your business on what LLMs say. So, when customers want to run reports, they need to run them reliably in a way that they won't get random answers. So, we've built something called Trusted Answer Search that uses vector search, not LLMs to find the correct answer from the user's question. So, you're guaranteed a response that is within a bounded set of responses that are known to be accurate.

Dave Vellante

>> Okay. So, it's LLM like an experience, but you're not getting the hallucinations?

Tirthankar Lahiri

>> Correct. Exactly right.

Dave Vellante

>> So, you were just talking about how you lock down the data and protect the data, some of the concerns that your customers have mentioned. Does things like open data formats, open table formats, Iceberg, does that change thing? Does it make it harder? How's the open source community doing? What can you tell us about that?

Tirthankar Lahiri

>> I think open standards are compatible with privacy. There's nothing wrong with open standards. And like I said earlier, this world is changing so fast, it's too early to commit to any one standard. So, we try to first work with all the model providers, all the LLMs, all of the vector-embedding services, all the AI agentic development frameworks, which you can run on public cloud or you can run in your private sandbox container. But our goal is to empower customers do what they think is right for their business and to change without changing their application architecture. We work very closely with the leading model providers. Iceberg has become a very important open data format and many organizations now use Iceberg tables as a way to share data across different business units. So, we have a new... I wouldn't say it's a new show in Las Vegas. It sounds like that. It's called Vectors on Ice, where you can actually now vectorize content in Icebergs, access that data in Oracle Database, index the data , and run queries against it. So, your data stays in your corporate enterprise data lake, but you can access that Oracle Database freely and with full flexibility. We also support the Model Context Protocol, MCP, allowing database access. And lastly, we think the world of agents today is highly siloed because every development team does their own thing. We build something called the open agent specification that lets you interchange implementations between different development teams, regardless of agentic platform they're using. So, that, hopefully, will bring us more harmony and homogeneity across different agentic environments.

Dave Vellante

>> Vectors on Ice? I like that name. Better than some of the names I've heard from Oracle in the past, no offense.

Tirthankar Lahiri

>> Okay.

Dave Vellante

>> So, my last question is, so you mentioned things are changing so quickly, and so this is a confusing situation for a lot of customers.

Tirthankar Lahiri

>> Yes.

Dave Vellante

>> Final advice, final thoughts, or give a plug for what you guys are doing. I'll give you the last word.

Tirthankar Lahiri

>> Sure, sure. Well, final thoughts. Well, AI is super important. AI is changing so fast and it is everywhere. So, all of the things I described today, they're available for no extra cost. Everything I said is free. The Agent Factory is free. The Private AI Service Container is free. Trusted Answers is free. So, we want people to start using their databases correctly for agentic AI without paying anything extra. We think that the data is the hub of AI and the AI should come to the data for the best AI experiences. So, for that reason, we want to empower developers to do any AI-style development they want. Does that make sense?

Dave Vellante

>> Yeah, love it. I mean, we've heard this throughout the day, the platinum levels, even the diamond level, if you got the right infrastructure, it just comes with the package. So, congratulations on... I mean, I know you guys spend a lot of time and effort and elbow grease moving fast, so thanks for coming in and sharing some of that with us.

Tirthankar Lahiri

>> Thank you. It was good. Thank you.

Dave Vellante

>> All right. Thank you. And thank you for watching. Keep it right there. This is Oracle Data Deep Dive NYC. You're watching theCUBE's coverage at the NYSE Wired CUBE Studio, right back, right after this short break.

Tirthankar Lahiri

>> Great.