theCUBE + NYSE Wired: Data + AI: Turning Data Into Knowledge for Autonomous Systems | Jagjit Singh Flahy | AI Luminaries with Neo4j

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

Jagjit Singh

Founder & CEO

Flahy

play_circle_outline Clinical Knowledge Layer Unifying Genomic, Biological and Clinical Data for Model Integration

play_circle_outline Versioned Knowledge Graphs for Explainable Real-Time Clinician AI Decision Support

play_circle_outline Integrating time-series wearables into graph nodes

play_circle_outline Enterprise Integration of Data Lakes and Databases via Graph Embeddings for Non-Graph Sources

play_circle_outline Agentic AI traversing domain graphs for recommendations

Info

Jagjit Singh Flahy | AI Luminaries with Neo4j

Jagjit Singh

Founder & CEO Flahy

search

(INTRO)

John Furrier

>> I'm John Furrier, host of theCUBE here in our Palo Alto studios. Of course, we have our New York Stock Exchange CUBE studios connecting Silicon Valley to Wall Street. This is the Neo4j AI Luminary series. We talk to experts to unpack the latest in graph and database and the AI layer, the knowledge layer connecting all the data and putting AI to work. Jagjit Singh is here, founder and CEO of Flahy. Welcome to theCUBE, appreciate you coming on. It's a pleasure to be here.

Jagjit Singh

>> Pleasure, John. And thank you so much for having me.

John Furrier

>> You're coming in from India. I know you got great bandwidth there, so thanks for coming in. This series is really about unpacking the experts that are doing the work and getting AI in action. The focus of the knowledge layer has been a big thing, but the role of graph databases has been key. You're starting to see this ingredient be very critical in connecting data together, and now with the AI models, you're starting to see everything move into this area. These platforms are going to have a lot more intelligence in managing the data and the models are critical. Talk about what you guys do and then we'll get into it, because you have a great use case. You've got a lot of data, a lot of context, and rolling that up in real time and putting it to work is key. So set the table, give us some context. What are you guys doing and what's your approach?

Jagjit Singh

>> Sure, Flahy is the clinical intelligence layer that connects your DNA, biological, and clinical information for better decisions around prevention, early detection, and treatment selection. And there's so much knowledge that we have to process. And what do you think a knowledge layer is? that's basically for a company, the rules it has to play by, the relationships it has, or the constraints it has to operate around. And if you look at how people discuss these separate layers of the stack, so first is you have the data layer, which is what happened, right? And then you have the model layer, which is the general reasoning. Now, the knowledge layer is very critical because it connects your data layer to your model layer, and it tells the model exactly what facts matter and what a specific data point implies. And that's where we come in, wherein we process a big context of biological and clinical information that's spread across a big graph. And that decides across which for a given person what decision tree or what route to take from here in terms of better treatment decision -making. And that's decided by the model because the model has already analyzed the domain. And it's reasoning over that specific domain of connected clinical information and deciding, okay, where are you within that domain and what's the best path from here to take or what's the best tree to go down. And that's what we do best at, and we have done it beautifully across prevention and treatment selection and diagnosis.

John Furrier

>> I want to get into some of the specifics. But I had a conversation in San Francisco with the CTO of Neo4j, Philip, and we were talking about the evolution of software. You had one database to rule the world, then you had cloud, now you have a lot of databases, now you have the AI era where you have multiple data layers and software always has been the case, but he calls it the connective tissue. So explain the new stack implication because what's happening here, what you just laid out is essentially, normal IT architecture. Hey, I got a bunch of data and I write some software and I expose it to a GUI, but now you've got the models. Okay, you've got the data layers, but a lot of work's done in data platforms. And now you've got the model layer coming in. Talk about the specific connective tissue between the two and why it matters and what are some of the results and outcomes people see?

Jagjit Singh

>> Yes, exactly. Because if you look at the model layer, right? If you look at general reasoning models, they are being commoditized very fast, right? So where is your edge? Your edge is how you have structured your domain over which the model will run. How have you built those relationships? Because that is what's making your model kind of predictable and not black box like. If you look at Flahy, right? We're into clinical intelligence or clinical decision making. So we need to be very clear, okay, what is the domain my model is operating over? How is it structured? If my model is telling you, okay, you have this mutation that's linked to this disease, so give this specific drug, right? How did it get to this conclusion? So in our case, what we have is that we have that graph database of the terrain very well mapped in terms of nodes, and every edge in that graph is versioned and curated with top technical talent who have put years into building that big graphical information database over which our model has then learned the relationships across these specific data points. And so it's very important because that's what separates you, right? What's your underlying base on which your model thinks? And especially within healthcare, how are things connected? So I feel like that's very important as to how you structure the base information, right? And then how you train the model over it or what sort of agency or what sort of freedom do you give? Do you restrict it to your domain or do you make it slightly free, like give it some sort of freedom? But at the end of the day, you want the model to select the correct branch to learn or what exactly within your underlying domain does it have to go to? And that's very different from databases. For example, if you look at Databricks or Snowflake, right, they are aggregating across 50 million rows. We are in a very different space. You know, I'm looking for this one person, this guy had a genetic mutation and this specific high cholesterol marker given this information. Now, if I have a new clinical signal, would that change how this person should be treated? Now, that is a very different question. It's a graphical question by design. And so you have to, it's very critical that your underlying information is structured in a way so that a general reasoning model can operate according to that. And that's the connective tissue. How do you kind of make it or restrict it in a way to specifically just look at the target information and think in a very graphical manner and do that in real time while information is typed into the system, because this is what I have, now tell me what to do. Does this change treatment or what? So it's the speed, your underlying data architecture, it's an architectural problem, basically. It's not an alignment problem. And if you're able to solve that, I think you're winning the race. Yeah. that's what we're trying to do.

John Furrier

>> Yeah, and I know, first of all, we're both excited by this because in computer science, you got nodes and arcs, Graphs have been around for a long, long time. If you look at AI, the compatibility in the computer science side is actually beautiful because you have recursion, you have the ability to kind of go down these graphs and it really kind of sharpens the context conversation. Everyone's talking about context. When you have this specialty intelligence, because it's domain specific, you could have a huge graph, millions of nodes, but you have general data, but you can go and narrow in on that domain. And that's where a lot of the enterprise, that's where a lot of these use cases where you want to have context. You can eliminate a bunch of data because you're not even using other graphs nodes. So you can say, okay, I'm going to tap this. This fits into AI. Talk about that aspect of it because this is where AI shines because you can get the context super fast, highly relevant, reduce hallucinations. This is like a dream scenario for the AI era.

Jagjit Singh

>> Yeah, and this comes to this thing of function calling. Now, if you look at how a graph is, it's a domain or it's terrain. It's like a geography of information connected to one another, right? Now, at one point, if you look at when I'm looking at how do I leverage this graph? How does AI leverage this graph to make a better healthcare decision for this person? Now, it's going to go and look at this specific person, where in that domain is that person right now? What's the specific context and how that context is structured, which is very important. The context needs to be structured well. And then what is the best path to take right now? You know, how many people connected to this interconnected graph, under how many clinical scenarios did giving this person a clinical change or a treatment change actually give a benefit in terms of outcome? And that everything can be structured as a graph. So it's where to go on from here. And I think that's kind of, it's very. And how we are kind of doing it is that you can train the model or give it some specific schemas. It's like a mix in terms of how to kind of get it. And there is a lot of programming involved but the whole point is quickly finding, okay where in the domain is that person right now? And what's the branch to run on it? What path to walk down or what tree to traverse here? And so the model goes and looks at their specific branch a specific sub graph and then tells you, okay this is what you should go for.

John Furrier

>> You know, Jagjit, this is a really good conversation. Because one of the benefits of AI we've been seeing and certainly we've been doing a lot of videos and talking to everyone from the NVIDIA's of the world all the way up to innovators like yourself is that AI infrastructure is changing the user interface. It's changing how people interact with the data. And the best companies are the ones that are aligned with and are accelerating the way people think and work. And if you think about what you've been talking about your use case, hey, I want to solve a problem. This is a person. Here's the markers I have. Here's some new data. What's the reason? That's just common sense. That's how the brain works, right? So that's not how a database works. So the mechanism of databases is kind of old school, but now you have multiple databases. You have a time series database. You have unstructured data. You're going to have notes. Talk about the importance of the architecture and then the software side of it, why that works.

Jagjit Singh

>> Okay. So, just to get you right, so you want me to tell. How are we connecting multiple different databases and building a unified context and running over that specific search space?

John Furrier

>> Yeah, the knowledge layer. How do you get that knowledge layer?

Jagjit Singh

>> Yeah, so for us, what we have done is that we have created our own kind of proprietary engines, wherein we kind of use multiple specific tools like Neo4j as one. And then you could also structure time series as a node in the graph, if you know how to do it. At the end of the day, everything is like in healthcare, everything, if you look at clinical decision making or the problem that I'm trying to solve, right, it is always a traversal problem. So when it comes to longitudinal data, you have to find a way how to put it in a specific graph, right? So for me, that's been kind of a challenge which we are effectively trying to solve in terms of how to kind of connect the dots across separate modalities. So for example, if you look at wearables, now wearables is time series data. How do you put that in a graph? So we've been trying to find, okay, how to make that longitudinal data or to kind of split that signal in a way so that it can be structured in a clinical, like in a traversal problem, for example, finding specific averaging, you know, across that longitudinal database. And then, you know, for example, creating specific nodes, for example, after every six months or after every 12 months, this is where the person was, and then creating an edge relation to that specific context. So if the model knows that six months in, my wearable was giving me a high glucose rate, I have a gene mutation, a high cholesterol level, I have a diabetes onset. So trying to find ways to put everything in that specific direction like one thing implies the other, the other thing implies the other, the other thing implies the other.

John Furrier

>> You know, the wearables. Treating the web of correlation. The wearables is a great example because if you look at some of the use cases there, it does great data collection, but it's terrible at alerts. I was reading a tweet by Bill Ackman. I'll send you the link. His daughter had a brain hemorrhage, but she had nine hours of data sitting there and she had no signal and she's recovering. And she's on her way to recovery. but that was catastrophic. So the signaling, the fact that the data was there, okay. Yeah, exactly. And there's no alert. So this is your point about the reasoning piece. You have new data. You have existing data. Existing data. You can work on that. So, crack the code on that. That's a home run. But in general, the enterprise is still going, wait a minute, I've still got my Oracle database. I got some stuff in the cloud. I got Postgres over here. Maybe I just use Postgres. Why should I even? So there's a lot of confusion around how to lay this out. And once people get it, everyone that I interview in this series is like, it's like they're dying to share their knowledge because it's like a secret. It's not a secret secret, but it's a superpower because once you figure it out, the results are there. Take us through some experiences you've had around that discovery of, wow, this really is kicking ass.

Jagjit Singh

>> Yeah, yeah, because, even if, for example, the case that you just spoke of. If someone had already built a context engine where in that clinical signal was connected to her biology and to her disease, the flag would have come. It's just that the context space, or no company had designed that context space where they've effectively integrated that longitudinal or variable data into their context graph, and the model is looking at that data and looking at it in the graph as to does that matter or not? Does that change anything or not? So what we have done is that if you look at, we did have EHR databases, a lot of EHR information, and then we have done some sort of encoding and tried to sort of mathematically encode them into our graph. You can encode certain clinical information from other databases and map them and push them into nodes. So your nodes can be anything. Your node can be even a database if you're able to embed. So you have these things called graph embedding algorithms. And by usage of these graph embedding algorithms, you can embed these other external data sources, which are not so graph friendly into your graph. At least they can learn the latent space over these guys. And that is something which we have tried and we have seen some success over.

John Furrier

>> Yeah, get that metadata, and you can actually put that as a graph node and then go deeper if you need to. Talk about the challenge. And this is an objection I hear. Not an objection, but more of an IT religious argument around, we have this or that database. A lot of large enterprises have investments in data lakes. You mentioned Databricks and Snowflake. They have a variety of other things, maybe some hyperscale or cloud stuff. What's your advice when someone says, well, I already have an investment in all this. Why do I need this? Talk about how this works with existing stuff, whether it's brittle old legacy or modern techniques.

Jagjit Singh

>> Yeah, so if you look at the existing techniques, right? when you query graph data, when you query an SQL database or something, your query already has the reasoning. You've already done the reasoning and then put in a query, right? This is totally different. This is a model telling you this is what I have. Now think what to do, where to go from here, right? So that's a completely different ballgame. It's not, you're not thinking of the logic and then aggregating, give me a trend over these 50 million orders. You know, that's like a lake house query, right? You know what exactly you're looking for. Or I'm looking for, I want to see, okay, my sales went up or whatever went up. I have a trend in mind and I'm just, aggregating and proving that trend. This is something different. This is more like, this is what I have. I have this, I have this, I have this new. Now, given whatever I have done previously in the past, does that change my present state? Where do I go from here? So that is a very different question that you solve with this type of graphical underlying basis and a model trained on that context space. You're no longer thinking first, you're just giving it information and telling it, okay, guide me. Does this change what I'm doing? Or what do I do now, basically?

John Furrier

>> Yeah, it's awesome. It's completely different. I'm glad you highlighted that. Now that takes you to the next level, which is the agentic era, because now you're going to have agents. It's not a search query, it's not to get the fastest answer, it's to get the right answer fast. Take us through how this plays out for agentic.

Jagjit Singh

>> Agentic, yeah. If you look at Flahy, we also have a consumer offering wherein we have an agent which has been trained over the same graph for prevention so for example we have this offering where we tell people okay these are the health areas you need to focus on by considering their dna and their blood context okay now when they're chatting with their ai agent and i tell it that person they have a mutation which is flagged or they have a high cholesterol level which is flagged and for example when chatting with the agent our consumer is like i have a family history of heart disease right. Now that agent knows because it has the underlying decision tree of the domain so it knows for example if you have if a family history node is connected to a mutation or to a cholesterol node that increases your probability of having heart disease right? So that's how the agent is further reinforced by you know you. So you have your underlying decision tree, you have your underlying context and you tell the agent okay this is the new information that i have. And then that agent is using that tree or traversing a specific path in that domain to give you that answer. That, okay, we really think that you should go see a doctor right now, because this is not an early pattern. If you have family history, then you need to start treatment right now. And that really helps beautifully in terms of preventing disease and avoiding downstream burden. And it's a big care gap that we are solving in that sense.

John Furrier

>> I love that example. And again, you mentioned traversal multiple times. That's what graphs are. You traverse the node and the arc, then you move around. This is where the reasoning works, because agents love that format because they can then go find and do work to get data and make a decision.

Jagjit Singh

>> Yeah, and it's not a black box. In healthcare, uncertainty is understandable. You can have leading doctors, leading surgeons, we'll try our best, but we don't know. Okay, there are chances of success, but a black box is not acceptable, right? You can't be like, I don't know what I'm doing, right? So the traversal is very important in terms of what our domain is, what Flahy's domain is. So it really solves that the agent over a graph of information, a domain that's clearly mapped with everything connected to one another and a clear line of decision making. That's quite beautiful.

John Furrier

>> Yeah, and software, not easy to do. But coding is very compatible with traversing and recursing through. All right, now. Final point, come back to the knowledge layer because this is the tie in. You got the data layer. Years and decades of data work has been done. and there's tons of data everywhere. You live in a sea of data. I'm sure you had tons. The models come in, general models, frontier models with general intelligence. The role of specialty intelligence is emerging super fast. Companies like Fireworks and others are really taking advantage of these smaller domains that have really specific data in the enterprise. That's their proprietary data or their domain. That's their competitive advantage. Or in your use case, that's where the data is for the value. So take us through. How do you set up the knowledge layer? Take us through what people might have to expect to go through. Is it a lot of grinding? In your case, you guys have a lot of proprietary data, so I see that, but if I want to just get involved, how do I get going?

Jagjit Singh

>> Yeah, so, it's very important because models are commoditizing fast. So what is your edge? Your edge is underlying information in your specialty that it's operating on to give it a better answer than a general reasoning query to a ChatGPT, right? So, yeah, for us, so we have the proprietary, but we also have external databases. So in our case, we had a full team of engineers, biologists, manually, it started with manual sheets and reference sheets and building CSV three years ago. And since then, we've been compiling, global databases of genetic information linked to clinical information, linked to treatment outcome, linked to diagnosis, huge, huge CSVs that later were turned into graphs and where every node was versioned and curated. Okay. And because guidelines change all the time. New findings come all the time. So there's a lot of manual work and thinking that went into first designing how I want to collect, assimilate, and connect this information, and then deciding on how exactly to train the model over it. So there's a lot of human element involved to get your AI work perfect.

John Furrier

>> So basically what you're saying is if you're in spreadsheet hell, you're probably waiting for the knowledge layer to be implemented because.

Jagjit Singh

>> Slowly, slowly, slowly, slowly, yeah. You have your spreadsheet, now you've got to start connecting your spreadsheets, connect the rows, connect the columns across spreadsheets, build the implications of the correlations, and you have your knowledge layer running ready. And just make sure the core. Yeah, sorry. Go for it.

John Furrier

>> And we hear also some great examples from people saying, hey, I'm connecting my Slack channel with this chat or my emails. I can go in. And so there's always other processes that have emerged out of either institutional knowledge of the job or undocumented layers of data. In healthcare, it's probably tons of written notes or process that's not being ingested. This is where I think this also can be put into the graph.

Jagjit Singh

>> Exactly. Exactly. Yeah. Yeah. Everything can be put into a graph. That's the beautiful thing about graph. You just need to know, okay, this is connected to this. And in this specific way, like this implies this, like X leads to Y leads to Z. Once you have that relationship right, or that correlation right, you can get going and you're building a knowledge layer. Now every sea starts from a drop. So slowly, slowly, slowly you start connecting your dots. And the more dots you have, the more connections you have, the better you're going to be versus another player in the market within that same category you're trying to compete in.

John Furrier

>> Well, Jagjit, you're now alumni of the Neo4j AI Luminary Series. Really appreciate it. Give us a quick taste of what the momentum is for you guys in the company. What are some of the things you're optimizing for right now? What's your focus?

Jagjit Singh

>> Yeah, so we are right now working with leading clinical laboratories and health systems and trying to deploy our platform in terms of better clinical decision-making and closing care gaps. Now, if you look at familial hypercholesterolemia, that's basically genetically linked to cholesterolemia, and it has a 96 % misdiagnosis rate. and it, you know, costs a hundred billion dollars in annual spend because you don't know that because that changes treatment. So if you could simply connect the graph, you can reason in a correct way, like, the model can look at, okay, given the specific context, okay, this seems a genetic case. So let's just do a genetic assessment. You're flagging this, you're serving the right people for the right test at the right time. You're intervening at, you know, when along the care journey should you want to intervene with either insight or a test recommendation to effectively change how the care is served better for the patient. So that's what we kind of really going and super excited about and working with leading diagnostic laboratories and health systems within South Asia and then also America soon to come.

John Furrier

>> Well, I love what you guys are doing. Again, it's great work. This is a great example of some of the breakthroughs you're seeing with AI. And of course, we highlighted one easy win out of the gate, which is crack the code on the alerts and signaling off the data, getting from the edge, the wearable, a lot of new data coming in. Got to get these knowledge layers in the context set up.

Jagjit Singh

>> Yeah, yeah, yeah. Because it's not data, like 10 years ago, it would be a data generation problem. Less signals, DNA sequencing was difficult. Now it's a decision-making problem. The bottleneck is decision-making. Like better data is not yet producing better decisions. And how do you get to that? By connecting the data correctly itself. First, what in the data implies what? And under what specific context? Once you get that thing right, and you get your AI to learn over there properly, you can start making better decisions right away. And the data is about to explode. It's already exploding, but it's only going to grow exponentially. So you've got to find a way to connect the dots.

John Furrier

>> You've got to get on board or the boat will sink. you've got to get the bigger boat. That's the knowledge graph. Jagjit, thank you so much for coming on. Really appreciate your time. And have a great evening in India.

Jagjit Singh

>> Thank you. You have a great day too. It was great speaking.

John Furrier

>> I'm John Furrier, host of theCUBE here, the Neo4j AI Luminaries Series. And we're talking to the experts. This is a mixture of experts we're putting together to share the new techniques to make things go faster. And again, the innovations driving value, not just revenue for society and in the medical field. Obviously this is a big win. We're seeing more and more of these use cases in biology, in all sciences and of course, financial services and other great things too. Thanks for watching.

Jagjit Singh Flahy | AI Luminaries with Neo4j

search

(INTRO)

John Furrier

>> I'm John Furrier, host of theCUBE here in our Palo Alto studios. Of course, we have our New York Stock Exchange CUBE studios connecting Silicon Valley to Wall Street. This is the Neo4j AI Luminary series. We talk to experts to unpack the latest in graph and database and the AI layer, the knowledge layer connecting all the data and putting AI to work. Jagjit Singh is here, founder and CEO of Flahy. Welcome to theCUBE, appreciate you coming on. It's a pleasure to be here.

Jagjit Singh

>> Pleasure, John. And thank you so much for having me.

John Furrier

>> You're coming in from India. I know you got great bandwidth there, so thanks for coming in. This series is really about unpacking the experts that are doing the work and getting AI in action. The focus of the knowledge layer has been a big thing, but the role of graph databases has been key. You're starting to see this ingredient be very critical in connecting data together, and now with the AI models, you're starting to see everything move into this area. These platforms are going to have a lot more intelligence in managing the data and the models are critical. Talk about what you guys do and then we'll get into it, because you have a great use case. You've got a lot of data, a lot of context, and rolling that up in real time and putting it to work is key. So set the table, give us some context. What are you guys doing and what's your approach?

Jagjit Singh

>> Sure, Flahy is the clinical intelligence layer that connects your DNA, biological, and clinical information for better decisions around prevention, early detection, and treatment selection. And there's so much knowledge that we have to process. And what do you think a knowledge layer is? that's basically for a company, the rules it has to play by, the relationships it has, or the constraints it has to operate around. And if you look at how people discuss these separate layers of the stack, so first is you have the data layer, which is what happened, right? And then you have the model layer, which is the general reasoning. Now, the knowledge layer is very critical because it connects your data layer to your model layer, and it tells the model exactly what facts matter and what a specific data point implies. And that's where we come in, wherein we process a big context of biological and clinical information that's spread across a big graph. And that decides across which for a given person what decision tree or what route to take from here in terms of better treatment decision -making. And that's decided by the model because the model has already analyzed the domain. And it's reasoning over that specific domain of connected clinical information and deciding, okay, where are you within that domain and what's the best path from here to take or what's the best tree to go down. And that's what we do best at, and we have done it beautifully across prevention and treatment selection and diagnosis.

John Furrier

>> I want to get into some of the specifics. But I had a conversation in San Francisco with the CTO of Neo4j, Philip, and we were talking about the evolution of software. You had one database to rule the world, then you had cloud, now you have a lot of databases, now you have the AI era where you have multiple data layers and software always has been the case, but he calls it the connective tissue. So explain the new stack implication because what's happening here, what you just laid out is essentially, normal IT architecture. Hey, I got a bunch of data and I write some software and I expose it to a GUI, but now you've got the models. Okay, you've got the data layers, but a lot of work's done in data platforms. And now you've got the model layer coming in. Talk about the specific connective tissue between the two and why it matters and what are some of the results and outcomes people see?

Jagjit Singh

>> Yes, exactly. Because if you look at the model layer, right? If you look at general reasoning models, they are being commoditized very fast, right? So where is your edge? Your edge is how you have structured your domain over which the model will run. How have you built those relationships? Because that is what's making your model kind of predictable and not black box like. If you look at Flahy, right? We're into clinical intelligence or clinical decision making. So we need to be very clear, okay, what is the domain my model is operating over? How is it structured? If my model is telling you, okay, you have this mutation that's linked to this disease, so give this specific drug, right? How did it get to this conclusion? So in our case, what we have is that we have that graph database of the terrain very well mapped in terms of nodes, and every edge in that graph is versioned and curated with top technical talent who have put years into building that big graphical information database over which our model has then learned the relationships across these specific data points. And so it's very important because that's what separates you, right? What's your underlying base on which your model thinks? And especially within healthcare, how are things connected? So I feel like that's very important as to how you structure the base information, right? And then how you train the model over it or what sort of agency or what sort of freedom do you give? Do you restrict it to your domain or do you make it slightly free, like give it some sort of freedom? But at the end of the day, you want the model to select the correct branch to learn or what exactly within your underlying domain does it have to go to? And that's very different from databases. For example, if you look at Databricks or Snowflake, right, they are aggregating across 50 million rows. We are in a very different space. You know, I'm looking for this one person, this guy had a genetic mutation and this specific high cholesterol marker given this information. Now, if I have a new clinical signal, would that change how this person should be treated? Now, that is a very different question. It's a graphical question by design. And so you have to, it's very critical that your underlying information is structured in a way so that a general reasoning model can operate according to that. And that's the connective tissue. How do you kind of make it or restrict it in a way to specifically just look at the target information and think in a very graphical manner and do that in real time while information is typed into the system, because this is what I have, now tell me what to do. Does this change treatment or what? So it's the speed, your underlying data architecture, it's an architectural problem, basically. It's not an alignment problem. And if you're able to solve that, I think you're winning the race. Yeah. that's what we're trying to do.

John Furrier

>> Yeah, and I know, first of all, we're both excited by this because in computer science, you got nodes and arcs, Graphs have been around for a long, long time. If you look at AI, the compatibility in the computer science side is actually beautiful because you have recursion, you have the ability to kind of go down these graphs and it really kind of sharpens the context conversation. Everyone's talking about context. When you have this specialty intelligence, because it's domain specific, you could have a huge graph, millions of nodes, but you have general data, but you can go and narrow in on that domain. And that's where a lot of the enterprise, that's where a lot of these use cases where you want to have context. You can eliminate a bunch of data because you're not even using other graphs nodes. So you can say, okay, I'm going to tap this. This fits into AI. Talk about that aspect of it because this is where AI shines because you can get the context super fast, highly relevant, reduce hallucinations. This is like a dream scenario for the AI era.

Jagjit Singh

>> Yeah, and this comes to this thing of function calling. Now, if you look at how a graph is, it's a domain or it's terrain. It's like a geography of information connected to one another, right? Now, at one point, if you look at when I'm looking at how do I leverage this graph? How does AI leverage this graph to make a better healthcare decision for this person? Now, it's going to go and look at this specific person, where in that domain is that person right now? What's the specific context and how that context is structured, which is very important. The context needs to be structured well. And then what is the best path to take right now? You know, how many people connected to this interconnected graph, under how many clinical scenarios did giving this person a clinical change or a treatment change actually give a benefit in terms of outcome? And that everything can be structured as a graph. So it's where to go on from here. And I think that's kind of, it's very. And how we are kind of doing it is that you can train the model or give it some specific schemas. It's like a mix in terms of how to kind of get it. And there is a lot of programming involved but the whole point is quickly finding, okay where in the domain is that person right now? And what's the branch to run on it? What path to walk down or what tree to traverse here? And so the model goes and looks at their specific branch a specific sub graph and then tells you, okay this is what you should go for.

John Furrier

>> You know, Jagjit, this is a really good conversation. Because one of the benefits of AI we've been seeing and certainly we've been doing a lot of videos and talking to everyone from the NVIDIA's of the world all the way up to innovators like yourself is that AI infrastructure is changing the user interface. It's changing how people interact with the data. And the best companies are the ones that are aligned with and are accelerating the way people think and work. And if you think about what you've been talking about your use case, hey, I want to solve a problem. This is a person. Here's the markers I have. Here's some new data. What's the reason? That's just common sense. That's how the brain works, right? So that's not how a database works. So the mechanism of databases is kind of old school, but now you have multiple databases. You have a time series database. You have unstructured data. You're going to have notes. Talk about the importance of the architecture and then the software side of it, why that works.

Jagjit Singh

>> Okay. So, just to get you right, so you want me to tell. How are we connecting multiple different databases and building a unified context and running over that specific search space?

John Furrier

>> Yeah, the knowledge layer. How do you get that knowledge layer?

Jagjit Singh

>> Yeah, so for us, what we have done is that we have created our own kind of proprietary engines, wherein we kind of use multiple specific tools like Neo4j as one. And then you could also structure time series as a node in the graph, if you know how to do it. At the end of the day, everything is like in healthcare, everything, if you look at clinical decision making or the problem that I'm trying to solve, right, it is always a traversal problem. So when it comes to longitudinal data, you have to find a way how to put it in a specific graph, right? So for me, that's been kind of a challenge which we are effectively trying to solve in terms of how to kind of connect the dots across separate modalities. So for example, if you look at wearables, now wearables is time series data. How do you put that in a graph? So we've been trying to find, okay, how to make that longitudinal data or to kind of split that signal in a way so that it can be structured in a clinical, like in a traversal problem, for example, finding specific averaging, you know, across that longitudinal database. And then, you know, for example, creating specific nodes, for example, after every six months or after every 12 months, this is where the person was, and then creating an edge relation to that specific context. So if the model knows that six months in, my wearable was giving me a high glucose rate, I have a gene mutation, a high cholesterol level, I have a diabetes onset. So trying to find ways to put everything in that specific direction like one thing implies the other, the other thing implies the other, the other thing implies the other.

John Furrier

>> You know, the wearables. Treating the web of correlation. The wearables is a great example because if you look at some of the use cases there, it does great data collection, but it's terrible at alerts. I was reading a tweet by Bill Ackman. I'll send you the link. His daughter had a brain hemorrhage, but she had nine hours of data sitting there and she had no signal and she's recovering. And she's on her way to recovery. but that was catastrophic. So the signaling, the fact that the data was there, okay. Yeah, exactly. And there's no alert. So this is your point about the reasoning piece. You have new data. You have existing data. Existing data. You can work on that. So, crack the code on that. That's a home run. But in general, the enterprise is still going, wait a minute, I've still got my Oracle database. I got some stuff in the cloud. I got Postgres over here. Maybe I just use Postgres. Why should I even? So there's a lot of confusion around how to lay this out. And once people get it, everyone that I interview in this series is like, it's like they're dying to share their knowledge because it's like a secret. It's not a secret secret, but it's a superpower because once you figure it out, the results are there. Take us through some experiences you've had around that discovery of, wow, this really is kicking ass.

Jagjit Singh

>> Yeah, yeah, because, even if, for example, the case that you just spoke of. If someone had already built a context engine where in that clinical signal was connected to her biology and to her disease, the flag would have come. It's just that the context space, or no company had designed that context space where they've effectively integrated that longitudinal or variable data into their context graph, and the model is looking at that data and looking at it in the graph as to does that matter or not? Does that change anything or not? So what we have done is that if you look at, we did have EHR databases, a lot of EHR information, and then we have done some sort of encoding and tried to sort of mathematically encode them into our graph. You can encode certain clinical information from other databases and map them and push them into nodes. So your nodes can be anything. Your node can be even a database if you're able to embed. So you have these things called graph embedding algorithms. And by usage of these graph embedding algorithms, you can embed these other external data sources, which are not so graph friendly into your graph. At least they can learn the latent space over these guys. And that is something which we have tried and we have seen some success over.

John Furrier

>> Yeah, get that metadata, and you can actually put that as a graph node and then go deeper if you need to. Talk about the challenge. And this is an objection I hear. Not an objection, but more of an IT religious argument around, we have this or that database. A lot of large enterprises have investments in data lakes. You mentioned Databricks and Snowflake. They have a variety of other things, maybe some hyperscale or cloud stuff. What's your advice when someone says, well, I already have an investment in all this. Why do I need this? Talk about how this works with existing stuff, whether it's brittle old legacy or modern techniques.

Jagjit Singh

>> Yeah, so if you look at the existing techniques, right? when you query graph data, when you query an SQL database or something, your query already has the reasoning. You've already done the reasoning and then put in a query, right? This is totally different. This is a model telling you this is what I have. Now think what to do, where to go from here, right? So that's a completely different ballgame. It's not, you're not thinking of the logic and then aggregating, give me a trend over these 50 million orders. You know, that's like a lake house query, right? You know what exactly you're looking for. Or I'm looking for, I want to see, okay, my sales went up or whatever went up. I have a trend in mind and I'm just, aggregating and proving that trend. This is something different. This is more like, this is what I have. I have this, I have this, I have this new. Now, given whatever I have done previously in the past, does that change my present state? Where do I go from here? So that is a very different question that you solve with this type of graphical underlying basis and a model trained on that context space. You're no longer thinking first, you're just giving it information and telling it, okay, guide me. Does this change what I'm doing? Or what do I do now, basically?

John Furrier

>> Yeah, it's awesome. It's completely different. I'm glad you highlighted that. Now that takes you to the next level, which is the agentic era, because now you're going to have agents. It's not a search query, it's not to get the fastest answer, it's to get the right answer fast. Take us through how this plays out for agentic.

Jagjit Singh

>> Agentic, yeah. If you look at Flahy, we also have a consumer offering wherein we have an agent which has been trained over the same graph for prevention so for example we have this offering where we tell people okay these are the health areas you need to focus on by considering their dna and their blood context okay now when they're chatting with their ai agent and i tell it that person they have a mutation which is flagged or they have a high cholesterol level which is flagged and for example when chatting with the agent our consumer is like i have a family history of heart disease right. Now that agent knows because it has the underlying decision tree of the domain so it knows for example if you have if a family history node is connected to a mutation or to a cholesterol node that increases your probability of having heart disease right? So that's how the agent is further reinforced by you know you. So you have your underlying decision tree, you have your underlying context and you tell the agent okay this is the new information that i have. And then that agent is using that tree or traversing a specific path in that domain to give you that answer. That, okay, we really think that you should go see a doctor right now, because this is not an early pattern. If you have family history, then you need to start treatment right now. And that really helps beautifully in terms of preventing disease and avoiding downstream burden. And it's a big care gap that we are solving in that sense.

John Furrier

>> I love that example. And again, you mentioned traversal multiple times. That's what graphs are. You traverse the node and the arc, then you move around. This is where the reasoning works, because agents love that format because they can then go find and do work to get data and make a decision.

Jagjit Singh

>> Yeah, and it's not a black box. In healthcare, uncertainty is understandable. You can have leading doctors, leading surgeons, we'll try our best, but we don't know. Okay, there are chances of success, but a black box is not acceptable, right? You can't be like, I don't know what I'm doing, right? So the traversal is very important in terms of what our domain is, what Flahy's domain is. So it really solves that the agent over a graph of information, a domain that's clearly mapped with everything connected to one another and a clear line of decision making. That's quite beautiful.

John Furrier

>> Yeah, and software, not easy to do. But coding is very compatible with traversing and recursing through. All right, now. Final point, come back to the knowledge layer because this is the tie in. You got the data layer. Years and decades of data work has been done. and there's tons of data everywhere. You live in a sea of data. I'm sure you had tons. The models come in, general models, frontier models with general intelligence. The role of specialty intelligence is emerging super fast. Companies like Fireworks and others are really taking advantage of these smaller domains that have really specific data in the enterprise. That's their proprietary data or their domain. That's their competitive advantage. Or in your use case, that's where the data is for the value. So take us through. How do you set up the knowledge layer? Take us through what people might have to expect to go through. Is it a lot of grinding? In your case, you guys have a lot of proprietary data, so I see that, but if I want to just get involved, how do I get going?

Jagjit Singh

>> Yeah, so, it's very important because models are commoditizing fast. So what is your edge? Your edge is underlying information in your specialty that it's operating on to give it a better answer than a general reasoning query to a ChatGPT, right? So, yeah, for us, so we have the proprietary, but we also have external databases. So in our case, we had a full team of engineers, biologists, manually, it started with manual sheets and reference sheets and building CSV three years ago. And since then, we've been compiling, global databases of genetic information linked to clinical information, linked to treatment outcome, linked to diagnosis, huge, huge CSVs that later were turned into graphs and where every node was versioned and curated. Okay. And because guidelines change all the time. New findings come all the time. So there's a lot of manual work and thinking that went into first designing how I want to collect, assimilate, and connect this information, and then deciding on how exactly to train the model over it. So there's a lot of human element involved to get your AI work perfect.

John Furrier

>> So basically what you're saying is if you're in spreadsheet hell, you're probably waiting for the knowledge layer to be implemented because.

Jagjit Singh

>> Slowly, slowly, slowly, slowly, yeah. You have your spreadsheet, now you've got to start connecting your spreadsheets, connect the rows, connect the columns across spreadsheets, build the implications of the correlations, and you have your knowledge layer running ready. And just make sure the core. Yeah, sorry. Go for it.

John Furrier

>> And we hear also some great examples from people saying, hey, I'm connecting my Slack channel with this chat or my emails. I can go in. And so there's always other processes that have emerged out of either institutional knowledge of the job or undocumented layers of data. In healthcare, it's probably tons of written notes or process that's not being ingested. This is where I think this also can be put into the graph.

Jagjit Singh

>> Exactly. Exactly. Yeah. Yeah. Everything can be put into a graph. That's the beautiful thing about graph. You just need to know, okay, this is connected to this. And in this specific way, like this implies this, like X leads to Y leads to Z. Once you have that relationship right, or that correlation right, you can get going and you're building a knowledge layer. Now every sea starts from a drop. So slowly, slowly, slowly you start connecting your dots. And the more dots you have, the more connections you have, the better you're going to be versus another player in the market within that same category you're trying to compete in.

John Furrier

>> Well, Jagjit, you're now alumni of the Neo4j AI Luminary Series. Really appreciate it. Give us a quick taste of what the momentum is for you guys in the company. What are some of the things you're optimizing for right now? What's your focus?

Jagjit Singh

>> Yeah, so we are right now working with leading clinical laboratories and health systems and trying to deploy our platform in terms of better clinical decision-making and closing care gaps. Now, if you look at familial hypercholesterolemia, that's basically genetically linked to cholesterolemia, and it has a 96 % misdiagnosis rate. and it, you know, costs a hundred billion dollars in annual spend because you don't know that because that changes treatment. So if you could simply connect the graph, you can reason in a correct way, like, the model can look at, okay, given the specific context, okay, this seems a genetic case. So let's just do a genetic assessment. You're flagging this, you're serving the right people for the right test at the right time. You're intervening at, you know, when along the care journey should you want to intervene with either insight or a test recommendation to effectively change how the care is served better for the patient. So that's what we kind of really going and super excited about and working with leading diagnostic laboratories and health systems within South Asia and then also America soon to come.

John Furrier

>> Well, I love what you guys are doing. Again, it's great work. This is a great example of some of the breakthroughs you're seeing with AI. And of course, we highlighted one easy win out of the gate, which is crack the code on the alerts and signaling off the data, getting from the edge, the wearable, a lot of new data coming in. Got to get these knowledge layers in the context set up.

Jagjit Singh

>> Yeah, yeah, yeah. Because it's not data, like 10 years ago, it would be a data generation problem. Less signals, DNA sequencing was difficult. Now it's a decision-making problem. The bottleneck is decision-making. Like better data is not yet producing better decisions. And how do you get to that? By connecting the data correctly itself. First, what in the data implies what? And under what specific context? Once you get that thing right, and you get your AI to learn over there properly, you can start making better decisions right away. And the data is about to explode. It's already exploding, but it's only going to grow exponentially. So you've got to find a way to connect the dots.

John Furrier

>> You've got to get on board or the boat will sink. you've got to get the bigger boat. That's the knowledge graph. Jagjit, thank you so much for coming on. Really appreciate your time. And have a great evening in India.

Jagjit Singh

>> Thank you. You have a great day too. It was great speaking.

John Furrier

>> I'm John Furrier, host of theCUBE here, the Neo4j AI Luminaries Series. And we're talking to the experts. This is a mixture of experts we're putting together to share the new techniques to make things go faster. And again, the innovations driving value, not just revenue for society and in the medical field. Obviously this is a big win. We're seeing more and more of these use cases in biology, in all sciences and of course, financial services and other great things too. Thanks for watching.