Jeff Gombala from Helios and Evidexa on AI behavioral models, digital twins, and shaping patient experience | The Dev is in the Details #19 S02

Innovation runs in weeks, evidence cycles run in quarters or years. We are introducing digital into a very highly regulated environment. AI is a truly, if it's used correctly, is a really great augmentation tool, especially in healthcare. So it's not really a job replacement in that sense, but it unlocks opportunities for us to better support people that are suffering from disease.

It's really empowering the innovator, the builder. It's helping us make better decisions and make more informed plans and products. Understanding customers and patients has always been a challenge. Companies spent billions in surveys, interviews, and research programs, yet the insights they get often fail to capture how people behave in real life.

This is especially true in healthcare, where patient journeys are complex, emotional, and deeply personal. But what if we go beyond surveys? What if we could model the behavior itself, simulate decisions, and use AI to forecast how real people will act long before we launch a product or service? This is the promise of AI-powered behavioral modeling, digital twins and simulation, great evidence.

I'm Lucas, and this is the Dev in the Details, the podcast where we look at technology and business through a practical honest lens. My guest today is Jeff Gombala, co-founder and CEO of Helios Innovation Lab and co-founder at Evidexa. Jeff has spent more than 10 years working in the intersection of digital health, behavioral science, and patient experience. He helps organizations turn early concepts into scalable digital products, build new service lines, and improve outcomes of AI-powered insights.

Jeff, great to have you here. Let's talk about how AI is reshaping customer and patient understanding and how behavioral models are becoming next generation of research. I appreciate it, Lucas. I look forward to the discussion and thank you for having me on.

Fantastic. Welcome. So maybe for starters, if you could explain how you come to the place you are, how did your journey span over the years and the interest in digital health and all of the aspects of that that led to your co-founding of Helios and Evidexa? Yeah, no, it's quite the journey, and I'll spare the details because it could go quite deep quickly.

But you know, I've been, as you introduced, I've been involved in digital health now at the intersection of both the innovators, the people building the solutions, as well as the adopters. So we say innovators are those digital health sponsors, be it a startup, be it you know, through VC funding or be it pharma-sponsored. And then you have adopters, which themselves could be, you know, large health systems, be it private or public in nature, as well as pharmaceutical companies as well.

But I've worked in an intersection for 10 years with the main mission to support people suffering from disease. So very mission-driven person. And what I've seen time and time again is the same repeating failure pattern, right? We are introducing digital into a very highly regulated environment where it's very difficult to understand the disease, let alone the regulatory complexity that we operate in.

So you have digital on one hand that evolved quickly, as we know for the value what digital is. And on the other hand, you have this evidence body that needs to exist, which evolves very slowly. Innovation runs in weeks, evidence cycles run in quarters or years. And that's what kind of brought me to this intersection of founding Evidexa really to help close that gap in how we can launch health tech solutions and make sure that they can reach the skill and the promise that they have.

Helios Innovation Lab is a means to an end in many ways, where it's our bootstrap consultancy firm. So we believe in the power of bootstrapping to build Evidexa and you know, so we use Helios to support our clients with consulting services at the same time. Thank you. That's fascinating.

Do you feel like maybe as I was listening to you, that question across my mind. You know, as in software we have all of these different environments, including staging and testing and whatnot. Would it make sense to do something similar in a legal aspect for healthcare? So we could actually have a group of people, an audience, right, for a clinical trial where people agree to be part of something, even if it's uncharted territory for sake of speeding up, you know, innovation?

Yeah, there's definitely so there's tools out there that we use today, right? The same tools that we'd use in non-healthcare settings, right? We use things like user testing.com, we use personas in many ways to help us navigate, and we use small patient panels.

The risk with all of that is they're all state-based, meaning if I'm a part of a panel as a patient and I raise my hand, I'm part of this small group of people, I as Jeff represent, you know, one perspective, and I'm influenced by the state I'm in, not necessarily the disposition of my behavior. And what we know from behavioral science is that our state changes based on environment. I might have woken up in a bad mood, I might not have had my coffee or something. But what we're trying to do with Evidexa is model the disposition and the core traits of behavior over time, which does a couple things.

One is it allows us to simulate in a repeatable evidence-based structure versus where you use personas or user testing. I don't know the state that people are in. And then I have to do a lot of power assumptions around how to apply that to my digital product and assume it's going to work for everybody. The other big risk is diversity, right?

Typically, those hand raisers are raising their hand because, you know, for the social good, they're interested in research. Or, you know, some raise their hands because they're looking for the incentive. And that bias is the population and it necessarily doesn't find the people when you think about massive scale of populations that suffer from disease, it doesn't really represent the whole. And with the power of digital twins, what we're trying to do is democratize that, create an opportunity, not replacement, right?

Because you'll still have to go through the regulatory great evidence, but at least make sure you're spending your capital in a wise way. And in your early product life cycle and your commercial decisions, you're able to have access to those insights, the foresight, in order to tune your product, change your messaging, support a commercial decision, those sorts of things. Okay, great, great explanation. Thank you.

Digital twins, I would like to understand a little bit more and for our audience a little bit more about them, because in production, in the car industry, and various different aspects of production, different industries, we have this concept for testing and simulating behaviors, but it's slightly different for humans, right? We by definition are slightly not predictable, like you said. So how do you design system to actually be a digital twin of a specific human being with a specific condition?

We could probably spend a whole episode, Lucas, on defining digital twin as a concept. You know, we've received some early feedback from you know our customers and in the market that to be careful with the word digital twin, rightly so, because I think it is used in many different contexts, as you said, right? From you know, simulation of industrial production lines and creating digital twins of how things flow through that to in the healthcare space, the sort of the ultimate digital twin is a whole representation of the self.

So the the digital specimen, if you will, which mimics how proteins fold and fantastical stuff that we don't do. Okay. What we create essentially is focused solely on behavioral aspects of the human. And we like to say it's you know creating a simulation environment much like a wind tunnel where you can test and repeat and whatnot.

So our digital twins are based solely on the fundamentals of behavioral science. So we're able to extract from the population through some of the tooling that we have those behavioral dispositions and traits. And then the one thing we don't do as well is we don't create a true twin. So Lucas and Jeff are not represented as twins in the environment, but what's represented is the attributes that imbue us as a human with our twin or our behaviors, and those are what we then resaturate into the simulation environment that represents the population, if you will.

So in practice, right? So you come to me and say, hey Jeff, I have a population of migraine patients that I want to simulate and test my new symptom management app, right? So I've developed an app that's gonna help you know people suffering from migraine manage and manage their condition. They would come to me and say, you know, in the population, I'm making some numbers up, but there's a million migraine patients suffering, or sorry, there's a million people suffering from migraines.

Can you simulate those? We would say, okay, a million is a bit extreme. We're not gonna do that, but we're gonna look at the population curve and we're gonna say, okay, for that, there's you know, this diversity of patients, which is context and state, and then we're gonna say there's these dispositions of behaviors, which is that core thing that we're trying to simulate, and then we're gonna launch that into our simulation environment for you to test. Fantastic.

That that sounds really great. And how does the just out of curiosity, well, maybe it's too deep, then tell me definitely too deep. How do you build the attributes on your population? Is this done by AI or is this done by an actual still human being, that data engineer who also understands healthcare specific problem domain?

We have clinicians on as part of the founding team, right? So we have practicing psychologists that have joined the team and they're part of the founding founding team of what we are. Their responsibility is solely for understanding the behavioral dynamics of the population and how to translate that into machine learning algorithms that then populate our meta model, if you will. So what we've done, we've looked at a lot of data, because you know, in healthcare, there's a ton of data, but unfortunately, a lot of that data is not at the specificity or the level that we need that captures human behavior.

So we have to go generate our data. So we actually launched a whole other product, which we call Evidexa Nexus, which is an AI-to-human speech-to-speech psychologist in many ways. I'll be very careful with that term because my sure my co-founder doesn't like it, but I call it that because it's a simple way to grasp the concept. But the AI actually will call you.

It will walk you through what we call a behavioral insight interview. And from that interview, we then have a series of agents that are trained in their domain, you know, for a specific behavioral disposition that look at the raw data, and their sole job is just to look at data, extract attributes, and populate our meta model. So we have one agent, which we call DEXI, and Dexi is designed by psychiatrists and based on behavioral science, and it conducts the interview. And then we have a series of agents that look at that data and then extract those behavioral dispositions and populate our meta model.

And that's the front end of the process. And then we have a whole other set of agents and AI that take those attributes and create the simulation environments for which we run, which we can dive into as well, on how that works as well. No, no, that's brilliant. That's yeah, that's a fantastic explanation.

Cool. Just one last question on this topic, I promise. Do you have a model that discovers emotion of the of the individual that is being transcribed to from voice to text? Yeah, like can we detect a mode of like you know certain dispositions of the voice?

We haven't explored that necessarily because it's not sort of core to creating the you know the attributes that we need to power our model, but in theory, yes, right? I mean, Evidexa and Nexus is built in such a way where it's in many ways it's a campaign engine, and you can deploy new tools against the data that we've collected. Tool being in you know a trained machine learning regression learning you know algorithm, or it could be you know some other tooling that we've designed, or you know, we're happy to explore with partners, right?

There's a lot of partners out there that have focused on using voice to do diagnostics for example for major just depression disorder and things like that. And so there's a rich opportunity in that environment for us to partner to help us expand our data set. You know, our data set two, I think just you can touch on this quickly, is fantastic you know, we're generating a sampling of 2,000 people, general population people as our initial data set. So that's actively running.

But our goal over time is to represent as many people as we can. So we go through the Evidexa and Nexus product, we want to collect millions of people's data and support them in this journey. And , I think that creates a rich opportunity for partnership and discovery and all sorts of great things. Yeah, totally.

I see that. And with your current and past experiences, right, in commercializing these kinds of technologies, what's your take on the strongest demand currently in AI, enhanced behavior on insights in healthcare? Yeah, so demand right now is you know, it shows up where you know the evidence is required early, but those live studies are too slow or too expensive. So where we're seeing our current demand is with those companies that are building rare chronic disease programs.

You know, so those are the focused in specialty pharma, those sorts of areas where there's an opportunity to help a person suffering from a disease either get an earlier diagnosis or there's an opportunity to support a person in their treatment journey once they've been diagnosed. And the reason being is developing that early evidence both for building the solution as well as making sure that it'll be adopted is really cost prohibitive and capital expensive and resource intensive.

So we're that's where we're seeing a lot of demand for this type of simulation evidence. And it's resonating quite well on both sides. So one side is the hospital systems, the health systems that want to adopt. They're looking at Evidexa as what they call a you know, lack of better term, a bench test.

So you can imagine on their side, they're getting pitched hundreds of solutions from startups that say we have the next best thing that'll help your patient population. But they don't have enough time to run all every pilot. Pilots are expensive. And so they have to, you know, create a signal from the noise.

Of course. And then on the other side, you have the innovators that we want to empower them because we want to make sure their solution scales and that they're getting adopted. So we're trying to really close the gap between those two. But the demand is strongest right now for us in the digital health solution space, mainly.

And then interestingly, we're starting to see as we expose the solution to more and more stakeholders at these customers, they're bringing some really unique use cases to us, which we're extremely excited about, but being a startup, we have to pace ourselves. But I'll give you a flavor of one. One being is workforce burnout. Sure.

So in the health systems, there's a big issue with clinician burnout, you know, and I think we all know the problem. There's an aging population, there's more disease, there's more demand on clinicians, and unfortunately, there's not enough nurse practitioners and clinicians to go around. And that creates a high burnout pretensity for these clinicians, which causes all sorts of bad effects in the system downstream. And there are health systems that are interested in using Evidexa because we're able to capture those underlying dispositions of behavior to model their clinician population to understand the burnout potential or how to and to use that data to optimize their workforce or where clinicians are placed.

So there's where I'm heading is we've stumbled onto something really important and critical for the digital health space, but there's a lot of applicability in other domains as well. That initial data intake about those potential patient profiles, can this happen independently from any sort of research or request from your innovator side? So for example, at the hospital intake. Yeah, it could.

So for the and you're saying specifically around how we generate the data. So what because I see a great potential in trying to understand at which point your data is you have a general population data with any without any specifics, but potentially someone comes in and asks you about burnout or a specific condition that they want to well simulate in those profiles that you have collected so far. Because it's either this or you're producing the data as I understood it, right?

Correct me if I'm wrong, or you're producing the data case by case basis for each innovator, which then I guess I'm asking about the reusability of each sample, yeah, how anonymous it is, and how can you do the intake without risking that it's just for one specific case. Yeah, no, great, great question too, because what we're doing is so the Evidexa data is essentially what's we're licensing. There's obviously an application around it that allows people to run the experiments and studies.

But the data itself is sort of what's core and unique. So the data is applicable to all, is really our goal. And we're investing in building the initial data set. But as we onboard a customer, they might say to your point, well, I have a unique population, or I have some other data I want to collect, or I have my own data that I want to add to the mix.

And we can do that and carve off their data for them and for what they need. So that's the power of Nexus, right? Nexus in many ways is a campaign engine where you can launch these assessments and you can launch these AI human agents to do all sorts of different types of data capture, data collection. So at the core of it, the data is the economy, right?

We're using that to make sure it's applicable to all of our customers. How you use it is the power of Evidexa as a core platform, if that makes sense. Okay. Brilliant.

Thank you. Yeah, that makes a lot of sense. That would be the core business then of the data licensing. Yeah.

And I understand that after the intake, because you keep the recording or transcript, you can still rerun other models or plug in different learning functions or data enhancement functions afterwards, right? Retrospectively. Yeah, well, this is where we have to be careful, right? Because when we conduct these interviews of humans, right?

If I was to call you Lucas or your loved ones or family, the nature of the interview is quite personalized and has a lot of health information and mental health information. So we necessarily don't retain that data. What we do is we got it, we extract what we need, de-identify it, and we aggregate it with the population double data, which then becomes the power of Evidexa. So there is the opportunity based off of the attributes that we extract to rerun different models and whatnot, but we were very careful about the data processor, data controller relationship of how Evidexa is a data processor and what data it manages, you know, to adhere to data privacy laws like GDPR and data privacy laws, you know, at the state level in the United States, so that we don't trip that up, right?

So most of our data collection, at least the initial covert that we're generating, we actually there's a separate legal entity that exists that is the data controller. And their whole responsibility is you know, explicit consent, how we use the data, how it's licensed to Evidexa downstream, so we can build those models responsibly. Okay. Yeah, that's one of the major complexities, right?

Hearing you will see it as well. Yeah. The regulatory aspects. That's what I meant when initially I asked if we could have like some sort of you know test environment where people would consent to say, look, as long as my name is not mentioned anywhere, you can reuse this data for improving your models for improving your algorithms in an anonymized way, right?

But I guess that's we're still not there, right? That would require regulatory change for mindshift, right? Yeah, I would think so. It's a complicated area, you know, to your point.

You were walking the some novel lines of innovation, you know, in terms of how we use you know frontier models like you know, Cloude and OpenAI and Gemini, etc. Walking a lot of novel lines around regulatory and compliance, how this data is used to create the twins. It's a complicated space, and I think there's the journey of Evidexa is starting, right? We've launched this year and to great success, both techno technical success and customer success.

But there's a journey to go, right? I think when we think about regulate regulatory landscape being a big one. So we'll it'll be interesting to see how Evidexa can influence if we have that power, you know, to make those changes. But yeah, be part of the ongoing discussion, right?

That's always interesting. Yeah. Well, I mean, to that point, we have a dream where you know Evidexa are currently to the where we're seeing demand. It fits in a particular space where you need that early evidence, and that helps you optimize you know downstream investments you'll make in other research, which is will always be required, right?

There's a concept in regulated software where you have notified bodies or the agency being the FDA. Sure. Where you have to signal safety and efficacy of these software as a medical device. And you still need to run that evidence.

But the goal for Evidexa is we help you optimize that run ultimately, as one of our value propositions. But I do see a world where these notified bodies and agencies potentially will say, you know, simulation evidence is equal or better than what they can run in the real world. So it becomes in many ways real world evidence based off simulated humans, in that we can trust that data and make regulatory oversight decisions. That's sort of the dream, you know, not tomorrow, but you know, in the near future, that we want to run towards.

And that way we create this whole new novel way of generating evidence and insights to unlock the power of what digital health can be at scale. Understood. Yeah, pretty cool. And on a patient level, you don't have any, so to say, B2C experience, right?

Because you always do it on the behalf of an innovator, right? Or regulatory aspects. So there's no need to have like a I don't know, I'm gonna make this up, gamified experience for your patients to participate in even simulated data intake for sake of you know building those profiles, finding out about the attributes in different cases. So for example, I have ADHD, right?

So knowing that I could participate in some sort of AI conversation. I do that with surveys and of some sort, right? And psychologists or psychiatrists are asking if I want to do it, I say, well, yeah, why not? Right?

They're anonymized, and I fill in the surveys. I suppose that could be the conversation with your voice-to-voice model instead, and that could be much more interesting for even individuals like myself. Like, I would be keen on doing this just to progress on the general understanding of the condition, right? Yeah, our co-founder Maitri Patel, she's our clinician, and she likes to say, you know, humans are very happy to talk about themselves, you know.

And I think it speaks Lucas to your point, right? Like you know, we want to collect that data because that's sort of the role of psychology and psychiatrists that practice is that they engage in these open-ended conversations, and you know, through those conversations, they're able to profile us as individuals and learn all about our behavioral dimensions, traits, traits and dispositions, etc. etc. Much less our behavioral health as well, be it ADHD or be it depression or other conditions.

But to your point, the B2C play really comes in that angle. So Evidexa as a core product is really a B2B play where we're you know selling to founders, innovators, and adopters. The other side of the house, back to my point, like ability to collect the data, we created a separate legal entity. In this case, we created a concept called Pathways to Health Index.

And that's what we're using as the B2C experience. So Pathways is really a website, a landing page where you can go and register and you can sign up for your profile, basically, to talk with the AI interviewer. And then what you get in return is what we call a Pathways Index. It's essentially a report about your underlying behavioral dispositions that affect how you make health decisions.

So it's an insights report that we're generating, giving back value for people participating in the Pathways. Let's call it a study. But over time, that Pathways becomes the B2C, at least what we're offering to the market. Now, the beauty of it is the way we built Evidexa Nexus is again, we could partner with all sorts of, you know, be it nonprofits or patient advocacy or even, you know, people interested in studying human behavior, and they could use Evidexa Nexus to launch their own studies.

The value is they can they control the data, they can do whatever they want to do for value exchange with the participant, but what Evidexa always gets is the anonymized aggregated data to power our digital twin models. So we're kind of building a data data economy, data flywheel with with that aspect of the solution. Right. Now, our B2C play, again, like I said, it's Pathways, but we can imagine you know, there could be particular you know, consumer plays that are specifically targeted to ADHD as an example.

And can anyone use this Pathways landing page, or is it just per invitation? It right now it's per invitation. So we're controlling the experience because again, that it's part of the the startup life cycle, right? You know, we're testing and learning as we're launching.

And right now it's US focused, mainly because we're in the US and we understand those privacy and... Of course. transparency laws better than we would in EU countries. And it's 22 and plus right now.

So we're being very specific around making sure people are adults and that we're not getting a bunch of students just participating for the fun of it. So we're learning about how to tune Pathways and create the value exchange and control privacy and the fairness of it. Cool. And beyond regulatory challenges, in your view, what are the most common barriers that organizations on either public side, healthcare side, or innovator side face when they try to adopt AI in their research or in their products?

Yeah, it's you know, AI is it's a fascinating space right now, I would say, right? Because there's so much innovation happening quickly on these frontier models. Yeah. So there's two sides of the barrier.

One is like how do you adopt Evidexa? And you know, what's the barrier to that entry point? And the other is if you're an organization considering AI, you know, whether it's to augment your current product or envision a new product, how does that responsibly fit in and what's that barrier? So from my perspective, you know, the barrier is really rarely the technology.

It's always comes back to AI as an augmentation tool and making sure that there's real issues you're solving, and there's you know an intended use and whatnot. Specific for Evidexa, it's all about the misaligned outputs tied to the innovation cycle as a barrier. So that extends, right? You know, as we think about those.

Oh, go ahead, Lucas. Sorry. Sorry, I'm not sure what these are. So if you could just clarify for me, what do you mean like those outputs, misaligned outputs, like what would be the example?

Oh, yeah, so misaligned outputs. There are when you run evidence cycles, you're early on a digital, you have those signals that help you make decisions of do we add this feature or how do we tune this message, those sorts of things. So those outputs tend to be early, or either bench research. So we're you know, we're using either prior knowledge that we have or we're using prior case from published materials and those sorts of things.

And those outputs are misaligned because it doesn't really keep pace with the decisions you need to make, if that makes sense. Got it. So the intended use is there, you know, we want to use evidence and research, but the outputs are being misaligned to the decisions you're making. And so that creates a complication.

But that's more, I think where I'm going with that. It's more of an opportunity for Evidexa and not necessarily a barrier. The barrier within that is the change management, right? So Evidexa is entering the scene.

We say, hey, we have this new way to generate foresight and insights and new evidence modalities, and we can help you optimize spend. That's a risk, right? Because people say, well, how do I adopt that when I know I have these small end panels of patients and I have this desk research that I could use to make decisions, and so disrupting that is a barrier for entry for us overall. Conversation kind of drifted there, Lucas.

But I think where I see is there's two barriers, right? Yeah. One is like there's people that want to adopt AI, and that's a barrier itself because AI is fantastical and it has a lot of promise. And then there's a we've leveraged AI to build up Evidexa, but at the same time, we face barriers for people to adopt us because it's new, it's novel.

So it there's this there's this paradoxal loop, if you will, that circles around these two things. Yeah, absolutely. And funnily enough, I think healthcare environment is particularly slow on that, right? Because there's a lot of fear, maybe, or doubts about you know what happens to this data, the regulatory aspect of some of the conversations about the bigger LLM providers and how they have trained their models, right?

So that's all open questions. I think it's more to the latter point, though, Lucas. Like the I would say people in healthcare in general, be it life sciences or healthcare in general, we're a suspicious bunch, right? So we always question if there's something new and in how it works and making sure there's evidence that it works before we adopt, which is why we have the whole situation with digital health adoption and scale to begin with, right?

And why we asked for evidence to prove this thing works. And bringing a novel solution like Evidexa is the same. So we have to prove to them that it works, that we represent, you know, the population of behaviors, and that those simulations can support them in the decisions that they need to make. So we have our own validation and evidence paradox to use that word as well, to drive that adoption.

It's rarely about the , because they do have they do evidence generation today, so it's rarely about the you know, the the ability to go talk to people. It's just that you know, doing traditional research methodologies today is just time expensive, resource intensive. You know, it just takes too long in order to iterate and build a digital product. Absolutely.

Yeah, I mean years and five, two digits in the millions, usually, right? So absolutely. I wonder, I guess one of the most common aspects that dive here and see as a pushback at the I don't know whether it's a hospital or clinic or patient side, is how do we avoid bias for you know, I guess that's one of the main reasons why any sort of automatic or AI-driven diagnosis tools are currently not so widely used, right? Because you can't rely solely on them.

And then in my experience, doctors would say, listen, this is an extra step and effort for me to get the data, even if it's just a recording, put that in the system, and I already know what my diagnosis is. So why would they listen to something which either exaggerates or downplays the symptoms into something which is not the correct conclusion? That being my experience, I wonder what is your protection against this kind of bias in your view in the Evidexa or any other AI system that is applied to healthcare.

Yeah, it's an interesting debate and conversation to have of where bias shows up and why, right? So if we built Evidexa in a way which solely relied on frontier models, meaning we ride the backs of Open AI and the training they've done, there's inherently bias in the data they use to build their model. In many ways, that's a good thing and bad thing for us, and we can talk about that. So that's one area where bias would show up, but we don't do that, right?

We just don't go out, you know, and say, hey, here's this LLM, we're gonna use them and prompt engineer and create a digital twin based of the training set that lives in that model. That's high risk, right? And there's actually competitors out there doing that, where they will they'll go basically scrape data from the web and you know, build quote unquote an authentic persona and then claim that as a twin and people use that to generate insights. But there's a lot of danger in that because you they don't really control it, it's not repeatable.

Repeatable meaning if I put my simulated thing into it on Tuesday and then I go back Friday, it's gonna give me very different responses. And how do I make decisions off of that? Let alone the bias that it introduces. So that's one area that bias shows up is at these frontier models.

The other is in the data we collect as well. And so as we build our data, there's we have to be very careful that we don't collect data that's overly biased towards a particular population, you know, or identity of people, right? And that we have a representation and we truly understand the disposition and the dispersion of behaviors across that population and make sure it's diverse based on what we know about populations, you know, through epidemiology data and other data sets.

So we have to control bias on the front end and then we have to control bias in the middle. What I mean by the middle is we take their data and then we put that into our meta model and then reconstitute populations. So being very careful not to do a one-to-one mapping back to the point of we're not creating a twin of Lucas. We're taking your attributes and the dispersion of the population behaviors and repopulating that into a simulated environment to control that bias.

So we have more control over how the population looks. And then we do have to ride the frontier models as well, because that's where agents and the novel technology exist. But we do something different in where we do LLM switching. So we don't rely on one model.

We'll switch between models because we want sort of the chaotic environment of different LLMs as we run our simulated populations, if you will. And so we hope that helps control bias from our perspective. Now, priest in the pudding, we have to validate that and demonstrate that. And and that creates a journey for us as well, you know, how we do that.

All right. And your intent is to create your own models and train your own models on that data as well, on that filtered out intake data that you collect? Yeah, we won't so we won't ever build our own LLM. I t's for where we are as a startup in this journey, it that's out of reach for us in many ways.

What we are building, what we own is the translation of raw data to the attribute data. So the machine learning process there, and there we use you know, open-weighted models, and we tune those models and we make those models ours, and we do our own reinforcement learning on those with our staff. And so that middle layer we control. We don't want to really control the LLMs because it's to our benefit that you know, Gemini and you know, Cortana and whomever's, right?

Go whatever next wave is coming, you know, they built their models trained on their data sets, and we can in many ways leverage that to our benefit because it creates different behaviors when you run the same the same agent on different frontier models, if you will. Do you get always deterministic results with this flow for any intake or the same intake multiple times going through your entire pipeline of models and processes that machine learning processes that you have designed?

Yes. Yeah, and that's the goal. So we're reaching a point where we're seeing the repeatability through the ML ops process. The output, you know, so when we put when we imbue the agents, the frontier model agents with the parameters to control them, we're seeing better control there as well.

So that becomes more deterministic. And now we don't own and we don't want to control it, right? So the whole point is when we run a simulation, we want to know we imbue those agents with the behaviors that we want to give them, and then we observe them. And so we're learning about things like comprehension, intent, knowledge, action.

And so those are the patterns we're looking to get back. And we don't really want that to be deterministic, we want it to be more organic in a sense that we want to see how the agents behave in the real world, real world being simulated world. But you see what I'm saying? But the front end of the process is deterministic, like we know the pipeline of getting raw data to the attributes we want, and that's repeatable.

Now, how those when you ask for a population back to the example of a thousand migraine patients, we'll say, okay, we want to represent this dispersion of behaviors, populate the twins, and then go run that in a simulation environment. We're observing that as much as we would observe you and I in the real world in what we do. You can't really make determined deterministic logic on that because I don't know what Lucas is going to do next. True, but I mean, like for attribution and for identification of certain symptoms to results or response to a medication, is that for a specific input will always give the same output, right?

That's how I meant it by the literal definition. But we don't so that's where why we stop short at behavior, right? Because if we start trying to model therapy response to you know a migraine medication as a drug product, that's a very different, a very different solution than what we are. So we're we live at the experience layer, not necessarily the biologic layer, if that makes sense.

Oh okay. If you could elaborate on this a little bit more. So your how I understood it so far. You have patient population details.

You will find a company who is interested in this either for research or testing my solution. It doesn't have to be a pill, it could be in some sort of you know meditation, maybe an app, right? How would you then run that pipeline for them to produce hints that it can go left or right, whichever way at the end, with that publication? Yeah, so maybe just to walk back to one point, so the experience layer, right?

So when you think about digital health, digital health is a simple equation in my mind, right? It's data science plus behavioral science wrapped around an experience layer. Experience drives engagement, it makes sure people you know on board and whatnot. Drug products, you like an actual pill or a subcutaneous injection, you know, for your GLP1 or whatnot, that experience is pretty limited, right?

A pill, I take it out of a bottle, I swallow it, is probably the most basic experience in the world. Yeah. Now injecting is a little different because you bring in the fear of needles and sort of other aspects of the experience. We don't really need to model that because it's well known, it's been proven.

Now, digital health, we need to model, and the solutions that we engage with every day do this for a living, right? Be it Netflix, be it Instagram. They use behavioral science to their advantage to create engagement and you know reoccurring, you know, use of the solutions. Digital health hasn't had that because of the again, back to the point of the regulated environment that we work in and our inability to generate real human interactions at scale, right?

So we can't go to a 10,000-person population pool and do A-B experimentations with digital health. It just doesn't work. That scale breaks. So Evidexa is filling that need by giving empowering digital health innovators and adopters the experience layer to say, okay, well, you can take your solution, run it through our environment, and we can help you understand is the population comprehending what you're doing?

Do they understand the feature that you're just wanting to launch? Will they take, do they have intention or are they taking action? And then more importantly, we can simulate over time, you know, how an agent will go through your app, if you will, and we know where the breakpoints are, where they're hanging up in terms of reading, you know, because of comprehension level issues, those sorts of things. So we're helping tune the digital experience layer, not the mechanism of action.

So the mechanism of action could be like if you went through this experience, I could reduce the number of migraine days you have. And that's a clinical claim. But to do that, I need you to engage with the solution. That's a behavioral experience claim, and that's where we help.

My next question is about one of the biggest fears that is going through people's minds these days when they hear AI, especially in general populace, which is the loss of jobs, right? But I really kind of feel that what in what you do, there is no equivalent. It's not like you're really substituting anything or taking away from an existing setup, right? Because it can be done at the scale.

It's just not that you're gonna replace behavioral scientists. I think you're just gonna empower them if I understood the entire conversation. Yeah. So but if you could comment on that for me.

Yeah, I know it's exactly how you framed it, right? It you know, AI to me is always the great augmentation capability, right? I don't necessarily see it replacing, at least in today's AI world, it's not gonna replace large populations of jobs at scale. You know, for us in particular, you know, the human has to define the question, the research question.

They have to interpret the trade-offs and make the decision and decide what matters from the results that we're giving them. To your point, we're an augmentation tool because the way they would do that today is they're doing it in very low-powered studies, low-powered meaning populations of 10 patients that represent a million patients. And that's a huge trade-off for these innovators, right? To make pretty critical feature go-to-market decisions and commercial decisions based off really low-powered studies.

And so we give them the power and help augment what they do, not as replacement, but truly as a multiplier, to make better informed decisions, you know, of what should the solution do? How should we tune our messaging? How should we tune our onboarding experience? How do we bring evidence to a potential adopter that says, yes, this will work for your population?

And the goal ultimately is to scale digital health at the end of the day. You know, I'll say one thing, like related to us, you know, our mission statement and views that, right? Our mission statement is to shape the future of health by empowering the innovators building it. And I think that speaks to like how we see AI playing a role as well.

It's really empowering the innovator, the builder, not necessarily replacing jobs. It's helping us make better decisions and you know make more informed plans and products, you know, that that will scale. I feel the same. My observation also is that we are in pretty much any specialisation help or significantly understaffed these days, at least in Europe.

You know, from simple technicians who run X-ray machines to to specialists. So anything that can paralyse or speed up or simplify in that area is more than welcome for public. Yeah. But it doesn't change the fact that the general populace is still afraid of it.

So yeah, I think it's we don't know, right? I mean, there's a lot of media out there, and there's a lot of promise of what these solutions are, even from the frontier models building it, right? You know, I don't know if you have these these advertisements in the EU, but in the US, we have Open AI ads on TV, which is fascinating to me. Like, Open AI is like you know, doing consumer marketing and saying, you know, hey, use our solutions and we can help you know in many ways.

What they're subtly saying is we can replace your travel advisor. I'm like, can you? Like it's a good claim, but I think backing it up is key. But Lucas, you brought up a really good point, though, is I think AI is a truly, if it's used correctly, is a really great augmentation tool, especially in healthcare for all the reasons you said.

There's the population isn't getting necessarily healthier, and there's a higher demand on health services, but there's less people to support them. So AI can show up in many ways across that journey and help remove things that we shouldn't just be doing, right? Administrative burden, you know, for prior approvals of a procedure. Oh, yeah.

Like spend the time with the patient and you know, helping them navigate. Don't spend time with the you know, the administrative burden and let AI handle that, right? So it's not really a job replacement in that sense, but it's, it unlocks opportunities for us to better support people that are suffering from disease. And Evidexa plays a role in that.

I mean, obviously, we're not dealing with administrative burden and those sorts of things with what we're doing, but we're making sure that technology is responsibly built and responsibly adopted and giving people access to you know evidence that they have or giving people the opportunity to generate evidence they prior priorly didn't have. Brilliant. Thanks for sharing. And , you know, we don't have ads in U E on the streets for AI, but I know what you mean.

I've been to New York and seen some of these. They're pretty cool. And sometimes really I think they , it's a bit of a marketing trick over actual reality of what they can do. I'm not gonna name the company, but they were literally saying like our AI never sleeps, would you still employ people knowing this?

You know, it's really direct messaging. Yeah. I mean, there's a beauty in a and a curse to the US advertisement system, but yeah. On the other hand, you know, it doesn't happen in here because we are overly regulated to not use AI, which is significant downside.

But to your other point, I wonder if you see yourself or market someone filling in the void for data duplication. One of the biggest nightmares that I see in EU in terms of patient journey is that you answer the same questions in multiple places on the intake, on the form, paper form still, that someone then moves manually to computer or two different systems because even the new chars are not properly talking to one another. And then doctor still asks you the same question afterwards, and then maybe nurse when she's you know injecting you with something, or she still asks you that.

So it's really interesting to see if your yeah, first of all, your perspective on this, because there that's it seems from someone like myself who comes from a tech industry, obviously I'm tremendously in the bubble, right, and biased about these things. How many things can be simplified immediately, not even before we even start the conversation about the AI. Yeah. So I wonder what do you feel Evidexa or your other company could, you know, what's your perspective personally or your business's perspectives on that?

Yeah, maybe from the Evidexa standpoint. We I don't know if we'll ever move into that space because it is a complicated space. You know, sort of who owns the data has always been a question. And when I say the data, in this case, I mean patient data, my data about my disease, my journey.

Yeah. Ultimately it's mine to own, but I'm never given the power to do it. So in the US, we have some interesting things, right? There's a thing they call the blue button, which is an initiative where I can access my patient records and I can own those.

Apple, you know, through their through their innovation, has given us some control over that, but it's still not complete. And then there's some other innovators around the space that have thought through it around how do I bring together Jeff's patient records into one single view for the benefit of him, but also for the benefit of his treating clinicians. There's really no holy grail there yet, I would say that that's scalable. Evidexa's position in that is if I am building one of those innovations, we want to make sure it's going to work for the masses by again giving them the support for the behavioral experience evidence they need to build it.

My perspective is there's data repeated everywhere within healthcare. So there's patient data repeated, there's market research data that's repeated, and there are opportunities for Evidexa to support in market research data, you know, and consolidating that and preventing that from being repeated. Patient data, I don't necessarily have a good answer for. I think it's an interesting intersection of patient engagement.

Like how engaged is Jeff to own and collect his data, and how aware is he of his disease? And you know, I'm talking about myself in the third person here, but just as an example. Because if I'm not engaged and if I don't really care that I'm about managing my disease, I'll never own that data. It's just never an incentive for me.

Sure. And then as a clinician, I only see that clinician, you know, once every year, maybe they're not really incentivized to own it, they're incentivized to process my billing for the claims, and my health insurance is the same, right? They're just trying to manage risk for me as a a person in their population and not really own my data. So that nobody's really in the ecosystem incentivized to own patient data.

Nor the patient isn't, the treating clinician isn't, the health system isn't. And I think until we've solved that incentive problem of like who truly owns it, it'll always be broken. At least the patient data aspect of it. I always thought that it's that electronic health record companies have the incentive, right?

Because that's their core business to deploy those huge EHR systems to hospitals and everywhere. Yeah, but think about it, like EHR data, very powerful data, but most of it is billing code data. Oh, okay. I saw Lucas.

Lucas did this procedure, this procedure, and we can make a lot of health understanding of those procedural codes, but at the end of the day, it's just code management. You know, now we have new technologies like ambient listening in the clinician office, which are picking up the nuanced discussions that are happening between doctor and clinician, but we don't really know what to do with that data, right? So back to the physician burnout conversation, right? We're now we have more data.

What are we doing to the clinician? Like, how do they make decisions on that? How do they not rely on their intuition and their education to make decisions? And now you have these other signals coming at them.

So it creates a really interesting conversation to be had around again, who's really incentivized to own it and manage it? Some sort of innovation or innovators should proceed with solving this, basically, right? And that's business case that seems to be built. Well, you would say that, but that becomes dangerous, especially in the US market, because why would an innovator do it?

They would only do it to monetize it. Yeah, certainly. Right. And I don't know if the general population wants their health data monetized for adverts or whatnot, however they use it.

It needs to be solved. I 100% agree with you, Lucas. I just don't know who's incentivized to solve it in a responsible way. Absolutely.

Well, let's park this one for the next episode. That's yeah, that is a deep dive and see a lot of intellectual conversation there. Super. Jeff, thank you so much.

Thank you for joining today and sharing your thoughts and were a very fruitful discussion. Really appreciate that. No, I appreciate it, Lucas, and I really enjoyed the conversation. We we explored a lot, but I was very happy to share the Evidexa story and what we're trying to solve.

Sounds amazing. Super fingers crossed, and we'll put all the details about that as well to the footnotes so people can find both companies and information about them. And the other one, what was it, the Pathways. Perhaps one day it'll be open and anyone can volunteer to participate.

Yeah. I'll be very keen on trying this out. Yeah, look for it. If you have a wait list, I'm happy to jump on it.

We are looking for early customers, and , yeah, I'll send you the survey link, Lucas, and maybe if anything, we could publish that to the population. Super. Your audience, I mean, and we can go from there. Thank you for sharing your experience and perspective.

It is clear that AI-powered behavior models are opening a completely new chapter in how we understand customers and patients. This approach has the potential to accelerate research, improve predictive power, and better align with real human behavior. To everyone listening, thank you for joining us on the Dev is in the Details. If you enjoyed this conversation, remember to subscribe and leave a review.

If your organisation is exploring AI or considering how behavioral modeling could support your strategy, feel free to reach out. We're here to help you navigate the details and build what truly works. Thank you for listening and see you in the next episode. Until next time.