Name: Jeevan Pathuri, Microsoft | AI Luminaries with Neo4j
Uploaded: 2026-07-29T18:20:00.000Z
Duration: 22 min 55 s

Clips
More from theCUBE + NYSE Wired: Data + AI: Turning Data Into Knowledge for Autonomous Systems

Jeevan Pathuri

VP, Software Engineering

Microsoft

play_circle_outline Company Brain: AI Inference and Graph Databases Unify Structured, Unstructured, and Connected Data

play_circle_outline Grounding LLMs with Domain Graphs for Reusable Connected Intelligence and Multi-Agent Reasoning

play_circle_outline Knowledge graphs, ontologies, and semantic layers for contextual grounding

play_circle_outline Data ecosystem: data lakes, Fabric, Neo4j as graph datastore

play_circle_outline Engineering tradeoffs: design, pre-compute vs real-time, scale and latency

Info

Jeevan Pathuri, Microsoft | AI Luminaries with Neo4j

Jeevan Pathuri

VP, Software Engineering Microsoft

(INTRO)

John Furrier

>> I'm John Furrier with theCUBE. Here in theCUBE's NYC studio, of course, we have our Palo Alto studio connecting Silicon Valley to Wall Street as the world is getting aware of AI. They now learn the value. And one of the big things is the inference piece of it and the graph side of it. As companies look to build their competitive advantage, it's all coming down to what is their company's brain and intellect? and how is that being put to work for value? This is our AI Luminaries series. We talk to the leaders who are working in the area. We're moving things from pilot into production, moving the needle, creating value. Jeevan Pathuri here, Vice President of Software Engineering at Microsoft. Knows a little bit about software, but also Microsoft is leading the AI wave with their cloud and Azure, and among other things, their relation with OpenAI. There's so much work there. Jeevan, thanks for coming on this series.

John Furrier

>> Likewise, John. Nice to talk to you.

John Furrier

>> So my little monologue there on the intro, lays out the premise. Every company is looking to inject intelligence into their organization. It's not an IT project. This is a business imperative. And so, okay, how do you do that? Okay, it sounds easy. The board says go, put intelligence into the company. Yeah, IT's involved, but the technology in the company now is everywhere. This is where we're starting to see formation of a playbook. Data lakes, data pipelines, graph databases. this is where we're starting to see the neural pathways, for lack of a better description, you know, context, leveraging the growth of the AI infrastructure. Tell me your perspective on this because, you know, it's all software, whether it's coded by AI or not. A new architecture is emerging that actually can turbocharge the path to putting intelligence into a company.

Jeevan Pathuri

>> Absolutely, John. The way I like to think about it, LLMs have progressed so much in the last few years, almost every three months now, the real intelligence stays within the company with the data that they have. And I like to think of data in three rough buckets, right? We've got the structured data, this is all your ERP and the tables and all of that. The unstructured data, which is the documents or Teams recordings, or even the skill files. Many companies are now just writing down everything they know in the form of these MD files and skills. And then the next one is the connected data. So the structured, unstructured, and connected. I think the real value of intelligence comes from these connections and the relationships between the different distinct data that we've got across the enterprise. I think that's where, like you mentioned, graph in the beginning, as we model things in the form of nodes and edges and the relationships, it becomes really easy for these language models to kind of understand and answer questions in a more sophisticated way than your standard answering questions by a SQL query.

John Furrier

>> Yeah, you've seen a couple of things. You mentioned SQL query, but also search comes up a lot. A lot of people have this search paradigm, get a retrieval engine going. We've seen RAG, retrieval augmented generation do that. But search is just the answer you get. It's the first answer. If we did that in our brain, it would be like, wait a minute, that's not the right answer, but I said it first. Reasoning is a neural thing. And take us through what's needed to build that architecture because you're seeing intelligence be about that. There's general intelligence and you have specialty intelligence. They're not mutually exclusive. I want to know what's going on on the web. I just go to the big models. But if I want to do something in the domain specific, write code, paint a picture, do something specific in the domain, that's different. It's like tapping part of your brain. I have my creative side, my left side, or go back to my memories. All these constructs come together in a computer science way. Could you share your view on this? Because this comes up a lot. We hear recursion, traceability. That's infrastructure and coding. Coding. Graphs are perfect for that. And then explain the model integration because they love graphs too.

Jeevan Pathuri

>> Oh, absolutely. I think you were actually mentioning this. You can make the intelligence really powerful if you give it a lot of context that's relevant for your domain. You take a large language model. You try to apply that on a generic problem. it does some things better because of all the distribution data it looked at in the past, right? But to really make it work for your scenario, your company, your team, and your group, you really need to ground it in the context that's meaningful for you, right? And this is where the graph and the connectivity really help a lot because I'll actually cite an example. In our own world, we have a problem for managing bill of materials for products. I support the supply chain division over here. And when we started modeling our BOM, the bill of materials as a graph construct, it became very easy to answer different kinds of questions using the same data set. For example, we've had this, a graph layer on top of our product data, just leveraging that, we were able to build 10 to 15 agents in production in a matter of a few weeks, right? So without this kind of connected intelligence, we would have had to build these agents individually one after the other, kind of rediscovering the relationships and the metadata along the way, versus having all of this created once and making it available for many scenarios. So this kind of structured way of creating the context and giving that is a pattern that we found very successful.

John Furrier

>> That's a great example. And I think if you go back to the early cloud days, one of the things I loved about cloud was it was horizontally scalable. And so I could build an app and host it in the cloud, but the app was still siloed. You had to have all the stuff for the app, the instrumentation. But now this layer is almost an AI infrastructure layer. I hate to use the word infrastructure, but it kind of is because it's a software layer that sits on top of the pure AI infrastructure, the factories, the clusters, all the gear and the memory. But this infrastructure layer has to be horizontally scalable, connected, but the agents aren't carrying the overhead of all the other stuff that you'd have to do if you build a SaaS app. So the SaaS and previous generations built silos and scattered data everywhere. This has become the preexisting condition. So I'm sure you agree, you're nodding your head. Okay, so what are people doing? Let's suss out. Do I want to leverage the trend? The trend is my friend. How do I manage all this brittle or disparate, disaggregated resource?

Jeevan Pathuri

>> It's no different than approaching, organizing your data, right? In the earlier days, you had this pattern of creating a data warehouse, finding the entities and relationships, right? So that structured data is definitely still needed. But beyond that is we're trying to make intelligent connections between the different entities where some of them you tell the system, you define based on your domain knowledge, based on your working knowledge. But the other is you leverage the intelligence itself to go and discover, right? So in case of us, we are leveraging many of these relationship intelligence capabilities to go and infer relationships. And then that, that's then surfaced to the humans to go and validate, does this make sense? So there's some of that working along with the agents to help you define that layer. So that's an interesting dynamic we're seeing now.

John Furrier

>> Well, I love this AI Luminaries series because you're so on the frontier. So let's talk about the knowledge layer, because this is where we mentioned earlier, the talk was about the ontologies. This is a critical piece. And it's going to be even more important as we go into 2027? Because a lot of people are doing the plumbing. They're plumbing their pipelines. They have data lakes. They got data everywhere. How important is the knowledge graph piece of it?

Jeevan Pathuri

>> Very much. I think if you are trying tojust repeat the kind of automations we used to build in the past, thenhaving the traditionalstructured, unstructured data set is sufficient, right? But if you want to really extract the intelligence and make the system work for you, defining the semantic layer on top is going to be very critical. And I'll tell you, another example is, we've been building these agents in our case over the last9 to 12 months. But as the LLMs are advancing, what became very clear is if we define our intent very clearly and explain the relationships and leverage some of these connected data sets, the time to market to build these experiences is shrinking quite a lot. An agent, we spent three or four weeks to develop in the past. Now we can do that in a matter of a few days or weeks or less than a couple of weeks. And that is because defining all of that, the relationships in the form of this semantic layer is making it really easy for us to write much less code and leverage the intelligence of the LLMs to actually do the work for us.

John Furrier

>> Yeah, and these ontologies and these new knowledge layers create that system of intelligence. Like you said, it's connected. So talk about reliability and production, because you mentioned you put agents in production in a few weeks, which is great. A lot of people are not there yet. So they're in test mode, experimentation, pilots, and they're doing good pilots because they're tackling hard problems. Take us through what does it mean to be trusted AI and make that reliable? Because at the end of the day, that's going to be the proof in the pudding.

Jeevan Pathuri

>> Oh, absolutely. I love this question because even internally, we always say humans are still in charge. So there are certain activities that the agents can do within certain parameters, but ultimately the decision is still owned by the humans. The way we approach reliability is at the most basic level, like everybody else, we spend a lot of time creating good evals. So evaluations for our problems. And we work with the business SME, the subject matter expert, in understanding exactly what are the parameters within which we can kind of let the agent do its thing versus in which situations it needs an approval from the human before taking actions. So I think beyond the evals, we look at the production traces on what's actually happening. The other thing we do slightly different is, as we build the agent, we work with the SME along the way. So that at the end of the day, the solution is not a surprise. There's no, you work for a few weeks and then there's a UAT and an acceptance. No, everything happens real time. Each day, every week, we see a set of capabilities, we get good feedback. So I think that close collaboration, evals and production traces, it takes a collection of things to really get the confidence that it's working in production. I don't think we've discovered a single technique that will give us confidence, but this is what's working for us.

John Furrier

>> Well, I appreciate that view on tracing because knowing your background in software supply chain, you mentioned SBOMs earlier, it's just SBOMs, but software bill of materials, you know the issues on supply chain with software. And we saw that with containers and Kubernetes and cloud native. Now you move to AI, you have a similar contextual issue. You have token context windows coming out of the factories or devices. Now you have context relationships in graphs. Okay, both are going to be managing context. You have now contextual relevance. But so I want to come to the human piece because the first reaction is, whoa, whoa, whoa, my job's going away. So what's actually happening with agents is that you're pointing this out. I want to unpack this. The AI architecture around intelligence, that structure actually creates a collaboration loop between humans and AI. So it's more like an F1 situation. You got a driver driving the super duper car. They got to be talented, domain expert, but there's a whole crew behind it. There's a pit boss. You got people changing the tires. You got people on screens, looking at IoT devices of the car. I see similar AI interactions. What's your vision on this? Because if this plays out, no job goes away. The job shifts.

Jeevan Pathuri

>> Oh, 100 % agree with you. When we approach AI solutions and automation, we think of this framework called CI plus AI. CI is continuous improvement of the lean, so in supply chain, that's a big deal. And then applying AI with continuous improvement gives us CA, which is capability addition, right? It's one of those things that you are able to do now, which you are not able to do before, right? So I think I love your F1 example is being able to do things better, smarter, more efficiently is where I think AI is actually going to help us.

John Furrier

>> Talk about the ROI and the ecosystem, because you mentioned some of that, the CA piece. One thing that's coming out of the past two years of theCUBE interviews and my reporting on SiliconANGLE is that everyone I talk to, that rhymes together, they're taking on the toughest problems with AI. That's different than the IT projects of yesteryear because the IT project playbook was, oh, we don't want to break anything. Go out to the corner of the property and over here, we'll experiment with it. We'll watch it. No, no, no. With AI, they're going after the hardest problems. And when I ask why, they say, because we have new capabilities that we didn't have before because they were difficult because they knew they wouldn't have bandwidth. That's too much heavy lifting or we don't have the horsepower. So you're building on this capability point. So I want to ask you, what kind of capabilities does having a neural company organizational brain, like with graphs connected into the data, what does that enable? What kind of capabilities do you see in reach and maybe down the road?

Jeevan Pathuri

>> Oh, that's a great question. The ROI, we still track the traditional ROI metrics, in case of us, it's what cycle times have we improved? How many hours have we saved? What non-value-added actions are we taking away? Speed of decision making and so on. But the capabilities we're seeing is more on being able to explain what the system is doing. That used to be a very hard task in the past for us, going through the logs or engaging an engineer or spending hours and hours going and talking to people on what's happening in the system. But leveraging the AI to actually collect all of this disparate information, drawing the connections, understanding the relationships, leveraging semantic layer, that has become much, much easier. So people are spending less time in discovering things, trying to analyze things, and then immediately going to the next step. okay, now that I figured out this is what's happening, what action can I take? So the speed of getting to that next step is actually what we're seeing more with these AI solutions than the traditional automation. So the other one is the speed with which we get the solutions to the users as well, right? Now we've got agents that can write code, right? So the engineers can actually direct the agents to get to the solution they want and then get it immediately in the hands of the users to get much faster feedback. So this whole loop is becoming much faster.

John Furrier

>> Yeah, the ontology lets you see things faster that you didn't see before. That's input to the human, again, this human AI loop. But also, it replaces the older mechanisms that you mentioned, getting product out there, A -B testing, all these things that were mechanisms because of the bottleneck. Absolutely. Gone. And now the engineers are driving it. And forward -deployed engineer concept, which I think everyone's a forward -deployed engineer here at this point given some of the trends not to digress but that you need the data so i have to ask you about the data ecosystem because when you deal with plumbing we just talked about the data that's available people love data lakes who doesn't love data lakes we know what relational databases look like there's a ton of unstructured data you got GitHub over there you got Slack over there you got Discord you got everything going on there's a lot of context that's laying around why is the data ecosystem equation so important for bringing in this graph, this ontology, this enterprise brain?

Jeevan Pathuri

>> Yeah. So, for us too, we leverage Fabric quite a lot, like having all the structured data in one place, having a single source of truth. But to add on top of it, for example, we use Neo4j as our graph data store. where the relationship intelligence is what really makes it more exciting for us is because we leverage that framework to understand how these different entities are connected to each other. And as we try to answer the questions that our end users have, having this semantic layer kind of reduces our workload by almost half, right? Like we don't have to spend the time trying to curate these connections. they're available in the form of the semantic layer. So I think that we definitely -

Jeevan Pathuri

>> So cut your workload in half.

Jeevan Pathuri

>> What part of it?

Jeevan Pathuri

>> Like the performance or is it the specific -

Jeevan Pathuri

>> Building this intelligence into the agents, right? So this layer is actually taking care of some of that work for us, right? Whereas traditionally you'll have to go and like construct your queries or build the queries on the fly. And here there's some of that thinking already baked into the layer, right?

John Furrier

>> Got it, okay, cool. Sorry to interrupt.

Jeevan Pathuri

>> No worries, yeah. So I think that's what we're seeing is spending time in building this layer actually pays off in getting the solutions out to users in a much quicker way.

John Furrier

>> So you got to do your homework.You got to lay down the right groundwork, plumbing, pipelines, Fabric. If you get that right, the spoils are for everyone to enjoy.

Jeevan Pathuri

>> Yes, definitely. There are no shortcuts here. It's still hard work. I think the only thing maybe that looks like a shortcut is you have some of these coding agents and GitHub Copilot, all these agents kind of helping you along the way. But creating that clean data layer, the connected data layer is absolutely a must.

John Furrier

>> And that's a real engineering opportunity. I think that's super cool. That's like the new SRE, I call it. It's the new data engineering. What's the consequence for the shortcut? Share your opinion on that because everyone will take a shortcut if they're allowed to. That's human nature. What's the pain? What's the downside for taking a shortcut? cut?

Jeevan Pathuri

>> Well, I think the downside is you'll end up spending more time and more throwaway work. For example, you may get to production much faster, but once you actually see it in action in a few weeks, you get real feedback from the users that, well, it worked well when I was testing it, but in my real usage, it's throwing off errors, right? So I think that is true.

John Furrier

>> The bad data or the lack of data or the misconfiguration creates compounding effect.

John Furrier

>> Absolutely.

John Furrier

>> That blows up downstream.

John Furrier

>> Yes.

John Furrier

>> All right. Talk about scale. You mentioned Neo4j and some of the test cases where you put it out there. A lot of people are playing with graph data. I've heard them say, well, I got open source. Scale is a huge deal and latency. How important is that and what do you do to mitigate any kind of risk there or just good design?

Jeevan Pathuri

>> Definitely good design, but it's a lot of experimentation to get the, at what layer do you do the analysis? How much do you do real time versus how much you do pre -compute based on the scenarios? I think the traditional engineering best practices still don't go away, right? It's more that you've got these multiple kinds of blocks that how do you combine and make them work better? So good design is still, there's no substitute for that.

John Furrier

>> Well, Jeevan, I love this conversation. What's your favorite thing that you've done with agents, graphs, that's been cool and relevant? It could be anything. What's the coolest thing you've done?

Jeevan Pathuri

>> Well, I think more recently, I've been leveraging our Copilots, which is just a Microsoft thing. It's kind of like NotebookLM, but for enterprise. and I have my own set of curated graph of all the sources that I work with that I maintain on the side. So it's really helping me. I have my own personal agent running in the background, even now when I'm talking to you, that's going and scanning my emails and messages and alerts me on what I need to go look at. So I've done that in the last few weeks and it's becoming very addictive.

John Furrier

>> Well, we'll have to get that agent working for our next interview. We talked about that last time, so let's get some new topics. John, thank you for coming on our AI Luminary series. Really compelling. Again, we are at the dawn of what I would consider a graph wave, which is about time. And maybe a decade sooner could have been better. But everything's teed up. You've got the infrastructure. You've got the enablement. You have the software layer, engineering layer coming in. So it's a beautiful opportunity. Data lakes are now open source. You now have lots of pipeline and capabilities. So all really perfect. So we will have more conversations. Thank you for taking the time.

John Furrier

>> Likewise. Nice to meet you, John.

John Furrier

>> All right. The AI Luminaries series really is the cutting edge on the frontier of this new layer. general intelligence, specialty intelligence, all together, our brains and the human we'll look at what ontologies and graphs could do for an enterprise and a business in the AI future. I'm John Furrier, host of theCUBE. Thanks for watching.

Jeevan Pathuri

>> Thank you.

Jeevan Pathuri, Microsoft | AI Luminaries with Neo4j

(INTRO)

John Furrier

>> Likewise, John. Nice to talk to you.

John Furrier

Jeevan Pathuri

John Furrier

Jeevan Pathuri

John Furrier

Jeevan Pathuri

John Furrier

Jeevan Pathuri

John Furrier

Jeevan Pathuri

John Furrier

Jeevan Pathuri

John Furrier

Jeevan Pathuri

John Furrier

Jeevan Pathuri

>> So cut your workload in half.

Jeevan Pathuri

>> What part of it?

Jeevan Pathuri

>> Like the performance or is it the specific -

Jeevan Pathuri

John Furrier

>> Got it, okay, cool. Sorry to interrupt.

Jeevan Pathuri

>> No worries, yeah. So I think that's what we're seeing is spending time in building this layer actually pays off in getting the solutions out to users in a much quicker way.

John Furrier

>> So you got to do your homework.You got to lay down the right groundwork, plumbing, pipelines, Fabric. If you get that right, the spoils are for everyone to enjoy.

Jeevan Pathuri

John Furrier

Jeevan Pathuri

John Furrier

>> The bad data or the lack of data or the misconfiguration creates compounding effect.

John Furrier

>> Absolutely.

John Furrier

>> That blows up downstream.

John Furrier

>> Yes.

John Furrier

Jeevan Pathuri

John Furrier

>> Well, Jeevan, I love this conversation. What's your favorite thing that you've done with agents, graphs, that's been cool and relevant? It could be anything. What's the coolest thing you've done?

Jeevan Pathuri

John Furrier

>> Likewise. Nice to meet you, John.

John Furrier

Jeevan Pathuri

>> Thank you.