Joseph Geraci, CSO/CTO & Co-Founder at NetraMark

In this episode, we sit down with Dr. Joseph Geraci — mathematical physicist, medical scientist, and AI innovator — to explore how artificial intelligence is transforming the future of medicine and clinical research.

Dr. Geraci began his career in quantum computing and mathematical physics, earning his PhD under renowned scientist Dr. Daniel Lidar with research supported by DARPA and the U.S. Army’s Disruptive Technology Office. He later expanded his work into artificial intelligence for medicine, oncology, and neuroscience, where he uncovered a critical problem in modern healthcare: traditional disease labels often fail to capture the true complexity and diversity of patients.

This realization led him to found NetraMark and develop NetraAI, a groundbreaking AI platform designed to learn from small, noisy clinical datasets. Using a novel mathematically-augmented framework and long-range memory mechanisms, NetraAI can identify hidden and explainable patient subpopulations with extraordinary precision — helping pharmaceutical companies improve clinical trial outcomes and better match treatments to the patients most likely to benefit.

In this conversation, Dr. Geraci shares insights on quantum algorithms, machine learning, the limitations of modern clinical trials, and how explainable AI could redefine our understanding of disease, health, and intelligence itself.

Transcript

Jesus Moreno (00:01.944)
Welcome back to Global Trial Accelerators, the podcast where we dismantle the barriers to clinical innovation. I am your host, Jesus Moreno, and today we're diving into the hidden geometry of clinical trials. The industry is currently obsessed with large language models and deep learning. But what if the very mathematical foundation of those models are fundamentally misaligned with clinical reality? Joining us today is Dr. Joseph

Jirachi, a mathematical physicist, medical scientist, and AI innovator. Dr. Jirachi earned his PhD in Mathematics specializing in quantum algorithms, worked support by DARPA and the US military before transitioning into oncology and biological psychiatry. He is the founder and CTO of Netramark, where he developed and

Neutra AI, a mathematically augmented platform designed to rescue drugs by uncovering hidden patient subpopulations in small noisy datasets. Joseph, it is an absolute pleasure to have you here to challenge how we think about health, illness, and machine intelligence. Welcome to the show.

Dr. Joseph Geraci (01:26.478)
Great to be here. Thank you.

Jesus Moreno (01:30.23)
I would like to start by exploring your career trajectory. It's fascinating because you have

You have circled the problem of human biology from some of the most rigorous abstract disciplines. You started studying neuroscience, moved into pure math and quantum computing, and then brought the mathematical firepower to cancer research and neuroscience. Going from providing that... Excuse me.

Going from proving the mathematical theorems of quantum computing to mapping the heterogen... that word, sorry... heterogeneity... going from proving mathematical theorems in quantum computing to mapping the heterogeneity of cancer is a massive leap.

Dr. Joseph Geraci (02:20.685)
heterogeneity.

Jesus Moreno (02:35.746)
Was there a specific moment where you realized that the complex mathematics you were developing for DARPA could solve the biological puzzle you first encountered in neuroscience?

Dr. Joseph Geraci (02:49.357)
That's an interesting question. So when I was at the University of Southern California, and thanks to DARPA's support, I was working on a problem where we were trying to figure out like, what are the outer limits of the power of quantum computation? It's always a question, especially back then that we were curious about. Is this...

assuming we're going to be able to build these systems, is there really an advantage for certain problems? So I used as a model this problem in statistical physics. And this problem is basically a toy model that we call the Ising model or the Potts model. It's a toy model for complex systems. And when I was

playing with this and starting to understand it better and better from the classical side, nevermind quantum with the classical side. I had the revelation that I was like, I have this great idea. I think that we can model cancer with this, right? Or, you know, the way viruses spread, whatever, like, you know, and then I realized that other people already had that idea. So I was working on a problem that because

when you generalize it, can, you know, it models complex systems quite well. it, I started having this feeling that, Hey, this can actually lead me back to where I started, which were in the biological sciences, biochemistry, and all that. So yeah, I had this, I had this inkling and then I had the hope that quantum computing one day might be able to start touching these domains. Maybe.

you

Jesus Moreno (04:44.496)
Hopefully it will when it becomes more mainstream and available to us. But that points us in another direction and something that you have emphasized or pointed out previously which is how fundamentally flawed our approach to medicine is when we're talking about labels. The labels we use like small, non-small cell cancer.

Excuse me, non small cell lung cancer or major depression are actually obscuring the true heterogeneity of patients. Can you explain how our reliance on this broad labels is setting clinical trials up for failure before they even begin?

Dr. Joseph Geraci (05:34.882)
Yeah, so let's talk about that a bit. and you know, humans are not to blame for this. It's really tough to make observations, but to me, patient population with a disease label, like you said, non-smell cell lung cancer or major depression or Alzheimer's, there's a lot of different types of patients in these groups.

And so what happens is, you know, historically what we've done is we've made observations about these diseases. and, know, we've made detailed reports about, you know, how these diseases start and how they progress. But what's happened now through the lens of machine intelligence and artificial intelligence, we're starting to realize, and we've realized this for a while, right? But we're starting to get more clarity that.

there's a lot more complexity and heterogeneity, in these diseases, just to say that there's, there's, know, two patients you can find in this disease that you give them the same label, but in fact, they're quite different biologically, right? And they respond differently to medication and treatments. And that gets to the point that you brought up. So it's kind of like trying to hit five target boards with one dart.

Right, so a pharma company comes and says, oh, I got this great drug for Parkinson's. And what ends up happening is when you do a clinical trial, you try to apply this to that label, and then you find out Parkinson's is actually a bunch of different diseases in some sense, or that the mechanism that the drug targets won't work for everybody, and you end up failing your trial.

And so then obviously, if you had a better definition about who your drug actually works for, you get a pass and a win.

Jesus Moreno (07:39.169)
And I think we saw this during COVID-19, where certain predisposed causes would drastically impact the outcomes for patients. As we move forward, how must the industry shift its mathematical paradigm from a broad, let's call it, network of

biological markers to a more dynamic system to actually capture the patient groups or not capture but segregate, separate or

specify the cohorts of patients.

Dr. Joseph Geraci (08:25.889)
Yeah, that's great. So look, to my mind, there are different layers, right? So at one layer, you can go work at the layer that physicians work at now, which is with these labels, right? Like this person has heart disease, this person has schizophrenia. But then what happens is if you lower that down, if you go down a level and just look at, you know,

the efficacy of drugs, you already start seeing some separation. But in another way, there's some unification that happens. And the reason is, is because if you go to an even lower level, which is what's going on at the biochemical machinery, you end up kind of blurring the divisions between these diseases. You actually have some subtypes of a disease you would never imagine is related.

next to something else because there's similar genetic or epigenetic or microbiotic. I mean, there's so many layers to it. You understand that. It's important for people to understand that as a machine, we have genetic machinery, proteomic machinery. There's micro RNA that does a bunch of other stuff.

which is very complex. And then there's epigenetic changes, which you do to yourself as you live. If you're someone who smokes, you're gonna cause negative epigenetic changes to your DNA. If you exercise, you give yourself positive epigenetic changes. And then of course there's the microbiome, right? The poop, right? And how this is actually a real nexus between the metabolism, the brain, and the immune system.

So there's all these layers. And what happens when you start looking at it like this, you start realizing, wait a minute, there's a whole other fine tuned resolution with which you can see these patient populations. And so I think now the problem with all that is it's a scientist's dream, right? And luckily for me, because I got to play in academia and industry,

Dr. Joseph Geraci (10:37.289)
I had the opportunity to model with all these different types of data. But from a business perspective, like you brought up clinical and pharmaceutical, this is very tough. This is a dream I have to wait for because it's very expensive to get all this data about someone.

So instead, what we're doing is we're using as much data as possible to create a new taxonomy of disease. And that's exciting, but in a practical sense, what it means is that pharmaceutical companies, I think, are going to have to harness artificial intelligence and different modalities of artificial intelligence to understand that their drugs have specific

they have specific advantages depending on different people that can be characterized quite accurately if you look at this the right way.

Jesus Moreno (11:38.423)
I see. But that raises another obstacle. I don't want to say a problem, but I think it's an opportunity. the obstacle is nowadays we want to use deep neural networks to study this cohorts of patients that if we were to classify or label them by those very specific features.

of their genetic composition would make the cohorts minimal. We could be talking about two, three, 30 patients. And now this neural network strategy that relies on considerable amounts of data falls apart. Why is throwing traditional deep learning at small patient populations not just

Ineffective, but mathematically impossible. Can you help us walk through that?

Dr. Joseph Geraci (12:41.963)
Yeah, so deep neural networks are one of my favorite things in the world. I love these creatures because we can do a lot with them, as everyone has played with chat GPT or Gemini, know, Anthropic, whatever, whatever you like. I'm even building my own versions of these things.

But the reason they behave so well with language is because they're trained on massive amounts of data. mathematically what happens is that their ability to predict the next word or to understand that this is a picture of a cat and not a school bus, is that that power comes from its ability to generalize towards the average.

And so these things require huge amounts of data. Now, everyone should be aware that cutting edge convolutional neural networks, you can literally change a few pixels on a page and you would still know it's a picture of a cat. But the deep neural network will now decide that it's a school bus. And it's, this points to a problem.

It's a foundational problem with the technology. going to your question, right? So the hope now, what a lot of people are doing is they're training these large, they're training large models to think about biology, which I'm a believer in because I think that you can encode a lot of biology in this way. But we're not talking about that in this specific question that you asked. We're talking about a system.

Okay, think about this, okay? You have a group of people you're bringing together in between 30 to a thousand people for a clinical trial, right? And I think those numbers are pretty good. representative of the trials. I've looked at trials for almost, you know, two decades now, right? So you have all of these people that you're bringing together. It's not a big number for a neural network to learn from. So,

Dr. Joseph Geraci (15:02.795)
But let's talk about a clinical trial, okay? Each person is complex. They all have moods, they all have a disorder they're dealing with, and they're all on past medication, and they all have comorbidities and you know, whatever, right? And now you have a group of doctors that come together and give them medication or placebo or control, and they interact with the doctors.

and they interact with the ambient information, which means like, what's going on in the world right now? Is there a war in Iran? How is that affecting you, right? So that system is a very contemporaneously locked thing that is being affected by the people that are interacting with it, interacting with the drugs that they're being given, and with the overall

happenings of the world. You get what I'm saying. It's an extremely complex system. now think about this. Now what you want to do is you want to take a deep neural network that's looked at clinical trials over the last 50 years. You train it on this. It starts understanding what a patient is and you want this thing to tell you about how this patient is going to do in your trial. Despite the fact that there are factors affecting that patient,

that the model's never seen before. OK, so then the question is, OK, well, we just trained on the small data. Well, that's a problem too, because training on that small data requires understanding something very special about patient populations, which is, goes back to the first thing you brought up, which is heterogeneity. The heterogeneity means that the labels that you give the patients for the machine to learn from are already faulty. They're under, they're like,

They're too simple. So what I've been proposing then is a system that can handle this. And mathematically, we'll go on to the next question and I can get into this. But mathematically, what happens is that when you have a small data set, like less than 5,000 samples, but a huge number of variables,

Dr. Joseph Geraci (17:25.581)
Mathematically, what happens is, and you can prove this, I'm actually working on a paper right now, I don't want to give too much away, but you can demonstrate that the data set will segregate into different subpopulations. Whereas if you have a huge data set, let's say with a million patients and 500 variables,

you end up getting something that will generalize quite nicely around the average effect. So instead of having one average effect with a big data set, a real clinical trial has the opposite. It has clumps of different patients. And how do you understand this intuitively? It's easy. Each, there's, you scoop a data set, like a clinical trial,

Some people are gonna have other people that look like that person according to the variables that you have. And other people will not look like everybody else. Some people call those outliers, they're not outliers. They're just different types of patients that don't have representation. And you have to treat them very specially or else you'll just overfit, specifically for these reasons I just gave you. Does that kind of clarify what I'm getting at here?

Jesus Moreno (18:41.78)
Yes, yes sir. And I think you've further captured these concepts in a beautiful analogy I heard you use previously, was, and feel free to correct me here. You said that training LLMs or image recognition models is like looking at water, where it's smooth and predictable, but

Dr. Joseph Geraci (18:42.167)
Thank you.

Jesus Moreno (19:09.214)
clinical trial data is like stone under the water, clumpy, chunky and concentrated. Maybe can you help us unpack how the concept of geometry and how does that concept apply to the different causes of... Excuse me. How does that...

The concept of geometry and the difference between what large datasets look like in relationship to the clumpiness, the chunky concentrated data of a clinical trial dataset causes this average effect to fall out of the intended use.

Dr. Joseph Geraci (20:06.859)
Yeah, yeah, and thanks for reminding me of that metaphor. What happens is when you have a lot of samples, a deep neural network is very good at simplifying the data itself. It simplifies it. And it simplifies it correctly. And mathematically, there's a reason why it does this. And it's because, for the mathematicians listening, it's because neural networks have a tendency towards

the reason it doesn't overfit when you have a large amount of data is because the functions that it's approximating are simple. For non-mathematicians what this means is that it really starts to get the essence of the phenomenon it's looking at and these are very close to the average effects become very strong when you have a lot of data a lot of samples.

It's the law of large numbers. You can flip a coin a few times and get a bunch of heads. You can get way more heads than tails. But I promise you, if that's a fair coin, if you keep flipping it, you'll get very close to 49-50 % of one and 49-50 % of the other, or 51-50, right?

So that's that you I think that's very easy to understand that as a phenomenon happens more and more the average becomes the dominant right and so you start really understanding like a cat really looks like that and we're not just talking about a simple average like heads and tails we're talking about an average across multiple dimensions whiskers and the fur and the eyes and the tail whatever features right

In a small data set, that's not what's happening. There's not enough room for that to happen. What you need is a system that's intelligent enough to say, I don't understand this person, but I understand these people. And by using that as a leverage to say, okay, in this clinical trial, the dominant, or maybe not even dominant, but the three dominant patient populations,

Dr. Joseph Geraci (22:09.805)
can be characterized like this and their statistical significance. So the difference is, is because you don't have enough degrees of freedom for that overall average effect to occur, you end up having these clumps. Now what's going to happen is this. Eventually, as we start collecting more and more data, we're going to start learning actually that there is this clumpiness in the large data sets. The problem is,

Right now we have a predisposition to oversimplifying those labels. So what I'm proposing is that mathematically we need to do a better job at implementing what I call the modularity hypothesis. Pardon me. The modularity hypothesis is something that nature's taught us. And it's this. You don't just have a large language model in your head.

you have multiple computational substrates. You have a frontal cortex, which is for reasoning, have parietal lobes for situational, you have occipital for vision, and you have this integration, have temporal, have all of these different lobes that our neuroscientists have been teaching us, right? And I think AI needs to evolve in the same way. I think that you can't expect one type of learning

to just do everything. And so what I've been focused on is constructing mathematically alternative learning substrates that can work with the deep known networks. See where I'm going with this?

Right? Anyway, I think I addressed your question. Tell me if it's not clear, but that when you have huge amounts of data, averages start to dominate. When you have small data, you can't take advantage of a system that's very good at discovering those averages and encoding it in a generalizable way. You need a system that's able to say, I don't know about these people, but these people have a strong signal because there's people that look like them. And because clinical trials are so well designed,

Dr. Joseph Geraci (24:19.777)
These things can appear in the smaller data sets. They're there, right? You just need the supervision, the way the machine learns, to not be so obstinate, right?

Jesus Moreno (24:33.284)
Not have that averaging effect just obfuscate or overpower the smaller cohorts of or sub population of patients within that big data set and the real life effects of ignoring that unintended effect could be that we scrape a multi-million

dollar therapy that just wasn't adding up in the average but that could be very clinically effective for a subpopulation.

Dr. Joseph Geraci (25:17.709)
100 % and you got it exactly right. Let me just add, the math can get quite simple here and it's just this, right? What I'm talking about is in the complex scenario, which is the reality for patient populations, but I've seen in some cancer data sets, some.

where there's a dominant genetic marker that explains everything. And that's the way to think of it. There's the average effect. People that have this polymorphism or over-expression of a gene, they have this particular effect or response to a drug. It's very rare to happen, but that can happen. But that's a good way for you to make sense of it. A great example is this.

If I have five feathers and I have five...

let's say bullets, okay? And I drop the feather in a normal environment and I drop a bullet. The bullet will fall a lot faster than the feather. And because the effect size is so massive, one looks like this, the bullet would just drop, right?

You don't need a lot of samples. And any machine learning algorithm will learn from the five and five, very small samples. And it will get it right, because the effect size is massive. Now, imagine this, that hiding inside a small data set, there's this feather and bullet scenario.

Dr. Joseph Geraci (27:04.191)
A deep neural network will not find it because it's trying to find something that works for everybody and all that. if the effect size is not overwhelming, finding that subpopulation with that effect or the multiple with a combination of variables, that's the other problem here, is the combinatorics of the problem as well when you need explainability. Anyway, I'm leading you astray, but we can get there.

Jesus Moreno (27:30.564)
You can tell you're very passionate about this subject. I would like to, okay, now that we have a better understanding of what the problem is, which is that averaging effect for a highly complex, highly variable data sets, where the sheer amount of information obfuscates or hides potentially interesting subpopulations. Let's transition from that to

what the solution to the problem is. Instead of relying on that data hungry deep neural network, I understand that neutral AI utilizes a completely different branch of mathematics. And please lead me through this terms because they're quite abstract. I've read through them and tried my best to understand them, but

We need your guidance to get through this technical segment of our conversation. I remember two terms specifically, dynamic systems and novel large range memory mechanisms. I understand it as the magic in this systems being the ability to partition

even in even small amounts, the data in a way that represents the data very honestly. So is that accurate? Is that what it's doing? And how does that effect help us overcome the obstacle we have discussed?

Dr. Joseph Geraci (29:16.065)
That's great. Yeah, you do have it. So just for clarity, Netra AI comes from the Sanskrit word Netra, which is to see our beautiful eye. Because the first time I built a system based on this principle, I was actually able to see the way the patients related to each other and why.

Right. And it was beyond principal components or these other algorithms that are out there. And what I was noticing is that this thing was able to reorganize the data and use the labels you gave it to guide it, but not to be restrictive and to relabel the data sets in its own way saying, yeah, you have this label here, like, aggressive non-small cell lung cancer, but really, that breaks up into three different structures.

And I was able to even to like put in stuff where I knew it was how like leukemia data, we know some of these diseases come in multiple forms and just blinding the system and putting it in and it was able to see those substructures without me telling it. And then I was like, this is what we need. And so the mathematics behind this thing comes from a branch of math that we call dynamical systems. And what these things,

The simplest dynamical system or the most fun dynamical system to think of is of the solar system, right? We're just, you know, through time, planets are revolving around the sun because of some dynamical principle, which is the gravitational force and the fact that we're in motion, right? And so we have these elliptical orbits.

And so if you can back, you know, fast forward time or backup time, you can see that the planet will traverse around this arc back and forth. And then there's stability, right? There's stability. The stability is the fact that, you know, Mercury, Venus, Earth, so forth, have stable orbits. Okay. Now imagine an abstract version of this.

Dr. Joseph Geraci (31:33.676)
that can take a patient and find its orbit in a very weird geometry. It doesn't look like this, right? But these each you throw patient in and they orbit, they go through an orbit of this thing where because of the way I designed the algorithm, they're guaranteed to converge to some path, right? In other words, computer scientists, we need to have some sort of promise that our algorithm is gonna finish.

Right? And so we have a promise that this thing's going to organize the system, the set of patients, in a way. That's number one. Number two, it's kind of like being on a golf course and it's windy, right? So the trajectory of the ball will move according to how hard you swing, the pitch of your, of the, you know, the head of the, of the,

you know, of the golf club, the wind, and all of these factors, they accumulate to end up projecting where the ball is going to end up. Very simple, right? It's the force, the pitch, the wind, and a bunch of other factors, right, that come together that end up mapping your ball from the tee to where it lands in the grass or in the sand trap.

Now, I want you to think about the sand traps. It turns out that there's a way to design a system like this golf course thing where the sand traps are your friend, not your foe like in golf. So all these factors come together to throw the patients out into space and if they get near one of these sand traps, these sand traps actually emerge from the data. It's very hard to imagine.

But these things emerge and what they do is they suck the patients in if they're similar according to the combination of those variables. So here's the real magic, okay?

Dr. Joseph Geraci (33:46.625)
If I give you a deck of cards.

Jesus Moreno (33:47.095)
It's like a selective black hole. To build on your analogy, what I'm imagining is when I throw a quote unquote patient into space, there's a bunch of little black holes all over that are selective to a particular feature on that patient. And if it's a match, they pull them towards that center of gravity.

Dr. Joseph Geraci (34:09.847)
Right, it's an emergent. Those black holes? Right, the black holes are an emergent property of the data and the patients interacting. Okay, now here's the killer, okay? If I give you a deck of cards, 52 cards, and I ask you to shuffle it for about a minute, you'll create a signature of your existence in this universe that will probably persist.

for the whole time the universe exists. Are there aliens playing cards? Whatever, because there's so many combinations that I actually encourage people to look this up. It's actually, somebody does a great job at explaining how many combinations there are. Like you picking up a bucket of water every million years. Anyway, right, so.

So there's so many combinations. Now, what chance, if that's true, what chance is there for you to find the right combination of variables when you want to characterize your patient population? This is the problem with explainability. Right? This is something that you didn't really touch upon, but is critical here. And it's related to what you're asking. So these, this system, what it does is it sets up a space, a collision space for the variables.

so that these black holes, as you call it, these sand traps, as I call it, open up as it pertains to the combinations of these variables because of this interaction space. In other words, you cut through this combinatorial craziness. Not completely, it's still hard. Still a hard computation, right? But it does something differently than the deep known networks. So to go back to your question, this system creates a playground for the variables to interact with each other, for these things to emerge.

for the patients to get grabbed, for those machine to say, I don't understand these people, they end up over here, but I understand these people, they end up here. And what it does is it gets rewarded for finding pure subpopulation, meaning all of these people are 90 % or 75 % of them were responders to your drug and they have this one, two, three, four characteristics in common.

Dr. Joseph Geraci (36:23.925)
And then what you do is you set up a substrate where this becomes statistically tested and the best models are not eliminated and you have something robust at the end. And then what you do is you fuse this with these deep neural networks, these large language models, and you have this beautiful synergy where this system, the Netray AI, extracts through this black hole kind of metaphor. It finds the driving variables and then the large language model.

What it does is it tries to do the same thing by the way it fails because it's it's too hard for them But then it understands the context You give it the answer the Netri I gives it's a beautiful pairing Right and so now what we've done and Netri mark is we've built an agentic System for this where this Netri I that just described to you your black holes are just a tool

that LLMs don't have, that now we can give them and we have an agentic environment where they can interact. And we have a paper coming out soon that describes this interaction, actually. Anyway, sorry.

Jesus Moreno (37:33.251)
No, please don't be sorry. This is very exciting and beautifully illustrated. But I want to bring it back to, let's say, more pragmatic side of the conversation. In practical terms, what I understand is that this technology cuts through the noise to hand a sponsor, a company that's running a clinical trial, a clear, explainable patient subpopulation.

Dr. Joseph Geraci (37:45.463)
Sure.

Jesus Moreno (38:03.456)
essentially allowing them to pre-specify a winning cohort to a regulatory body like the FDA and potentially even salvage what might otherwise be a failed trial just by applying this technique to an already captured dataset.

Dr. Joseph Geraci (38:28.951)
Yeah, so the FDA was very kind and they invited, they agreed to meet with us for CEPA meeting, this critical path innovation meeting. And I was very lucky to have an ex-regulator.

train me on how to think these things. He's our regulatory innovation officer, Dr. Luca Pani. And Dr. Luca and I have been working very closely over the last couple of years, along with Dr. Larry Alps, who's a big name in clinical trials. And so I've benefited from learning about taking all this math and thanks for bringing me back home. Sometimes you have to do that with me, right? So, and so they...

So together they've given me this this vantage. then Dr. Luca Pani, who's himself a psychiatrist as well, but regulatory expert, was able to teach me what was really like missing from our system and how to reason this way.

We got this great meeting with the FDA. The FDA was really gracious, heard me out. They invited us because our technology is not Bayesian, it's not statistical, it's not standard neural networks. It's novel, they recognized that. And so they heard me out and they gave us this great feedback. And the feedback was this. They're like, you guys are actually novel, but what you should be doing is, and they said this because they developed a path.

for pharma companies to do what you're saying. I'm just gonna clarify the pragmatism here a bit, okay? So they came back with a real practical thing. We can do a lot with this technology. What they said is this, listen, when a pharma company does their phase two or phase three and they need to do another trial, you're right.

Dr. Joseph Geraci (40:27.115)
You study their data because you have that capacity to reveal that black hole, that sand trap of the advantage of the drug. These patients, these types of patients characterize in these very simple ways, like literally blood pressure in between this and this with this genetic marker or with this item on the ham D like scale, whatever, all kinds of data, urinary markers, anything you can imagine, imaging. And we come back and say,

You want this brain region to be slightly bigger in these patients and for them to be more depressed, specifically. What they do is you don't tell the pharma company to get rid of everybody and just take those people. You tell them, take your big shot on goal. Try to make your drug work for as many people as possible. But use Netra AI's insight, this model derived subgroup that that's what we deliver to our customers.

use that to pre-specify another cohort, put that in your statistical analysis plan. And so if your trial doesn't pan out, you have another shot on goal. And then the FDA will work with the company to provide a label and so on and so forth. And so it's a no-brainer for sponsors, for pharmaceutical companies, to say, OK, we're going to take our shot on goal.

We'll pay Netramarkt to get this insight and then we'll include this just in case. It's an insurance policy. On top of that, they get the added benefit because of these systems practically to understand their patient populations better and to interact with the sites where they collect the data in a much more effective way. there's companies now using what we've delivered to do exactly that, to interact with their trials because there's a bit of control there without cheating.

Jesus Moreno (42:27.982)
So instead of manipulating the results with statistics is hedge your bet and make sure that you're including the right patient population that is going to respond positively to the treatment.

Dr. Joseph Geraci (42:48.555)
Right, so that's exactly right. It's hedging a bet. And then if the signal, so for example, there's a company we worked with where we found a strong cognitive factor, we handed it back to them, their statisticians went away and came back and verified, yeah, we actually agree this happened. The pharma company can do what they want with what we deliver, right? I just gave you the most practical.

solid like FDA understands, right? But they can do other things with it. Like one group literally was just like, you know what, we're going to change our inclusion exclusion criteria. But really.

Our purpose is not that. Our purpose is to give you a stratification that will inform you so that you can design your next trial and produce statistical analysis plans that are aligned with the FDA's path that they already produced specifically to help pharma companies through this personalized medicine journey. This is where it starts. Real precision medicine.

Jesus Moreno (43:48.913)
So the real breakthrough here lies in fusing the LLMs with this specialized geometric technology you've built. Now that the regulatory door is open and there's that seal of approval, say, and the market is finally losing up to this concept, what is the end game for NitroMark and NitroAI?

Dr. Joseph Geraci (43:58.766)
That's right.

Dr. Joseph Geraci (44:19.691)
Yeah, that's a great question. mean, right now, even though I've been doing this for eight years and evolving the product, I'll tell you the way it started because it kind of answers where I think it's going to end. When I first raised money from Netramark, I drew this really crappy picture compared to what our people can do at Netramark.

I drew this crappy picture of natural AI in the middle of a wheel with spokes coming out of it. And I had different machine learning algorithms on the other side. And on one of the boxes, so I had like gradient boosting, deep neural networks, right, support vector machine, whatever, right? But I had one that had a question mark in it because I knew something was coming. You see like,

Me and my colleagues in this industry, we all realized that machine learning had a lot of potential, but we needed a way to have long range information traverse the learning so that, and the most obvious place where this showed up is in the efficacy of the large language models. ChatGPT is so good because there's this long range information that tells you the context of the word.

So it can predict the next word. That's how it works, right? It predicts the next word, next word, okay. And so that comes from something called an attention mechanism and from a famous paper called Attention is All You Need. During that time when those scientists were developing attention, I was developing a long range information connecting piece as well. But for patient populations, I didn't care as much about language. I cared about

How do I get these variables to come together? So when I successfully saw that I did that, it took us all this time to evolve the tech so that it speaks the right language for the regulators and for the pharma companies and all that. But you see in this picture that I saw that it was going to be fused with other methods. And so I think where we're going now is

Dr. Joseph Geraci (46:44.267)
Because of the agentic revolution, what we're seeing is that AI is now taking a form where you have a group of AI agents working together towards a goal. And this is what we now have at Netramark, where we have a regulatory agent. We have a scientist agent.

We have the Neture AI agent, which does this miracle of preventing, know, dealing with trillions of combinations through these sand traps and whatever long range memory mechanism. Right. So we, that's, that's what we're doing. We're building this agentic space. now that companies actually have a lot more freedom to use our technology, either within their own doors or they deliver the tech to a safe space and

I think the other place, think that this is where we're going now and we're still in the breaking through the customer stage right now. And so we, I think that it's a very exciting time for us, but I'll give you something more. I think that because of the partnerships we're forming now in oncology and that we already have, for example, with the Mayo Clinic and other work we're doing.

with sponsors I can't tell you about, that what's evolving is a much better understanding of the heterogeneity of disease, a new taxonomy is emerging, and this is gonna have downstream effects for drug discovery and also for gene editing.

and other things like that. This is gonna reveal, even though it's not our primary mandate is, you have a clinical trial, we're gonna help you succeed, right? But there's these downstream effects that has always been my dream that's starting to emerge now. That's where I see us penetrating into the market, becoming a real partner to these pharmaceutical companies, biotech companies.

Dr. Joseph Geraci (48:54.987)
But there's gonna be other consequences down the road.

Jesus Moreno (49:01.316)
Joseph, thank you so much for sharing that with us today. You've shown us that to truly accelerate clinical trials and save lives, we can't just throw big computers at the problem. We have to change the underlying mathematics to adjust, to mold, to adapt to the data sets and the real world we're facing. For the innovators and sponsors listening who want to explore

this clump he concentrated data and understand better how the tools you're building can help them. Where can they learn more about your neutral mark and neutral AI?

Dr. Joseph Geraci (49:50.177)
So the Netra AI tech and our materials are available at Netromark.com. Just go to the posters paper section and you can look there. We have a nature paper that was just accepted a while ago, just a few months ago, that actually gets into this a bit.

We have a new paper coming out, hopefully soon, on the modularity hypothesis. I'm writing a paper right now on the deeper, this is very technical, of what you and I just talked about, the geometry. And for fun, we're putting out a paper very soon about how quantum mechanics might play a role in the psychedelic experience.

Dr. Luca Pani had an interesting idea and I kind of figured out how it might be possible in a testable framework. I think that's gonna be a lot of fun for people to look at, especially now that psychedelics are an emerging therapeutic and we work with some of these companies as well. So that's a more exciting thing, right? So yeah, I'm not just all math.

Jesus Moreno (51:05.796)
Yeah, work hard, party harder.

Dr. Joseph Geraci (51:11.249)
Hahaha

Jesus Moreno (51:13.7)
To our listeners, thank you for joining us on this deep dive into AI and new technologies. Until next time, keep accelerating.