Let's chat!
Click HERE to book a call
Episode 109 (Season 5)
November 10, 2026

Your AI Strategy Is Erasing Your Market Edge

with Alexandra Pasi, Lucidity Sciences

Generic AI adoption is erasing competitive differentiation. Lucidity Sciences founder Lexi Pasi explains why proprietary data is the real edge. Learn more.

Most leadership teams now track how quickly the company is adopting AI. The number feels like progress. Lexi, founder of Lucidity Sciences, points to a harder question underneath it: where does the company's competitive edge actually live, and does the way it adopts AI protect that edge or dilute it?

As foundation models grow more capable, many CEOs are asking whether their offering will eventually be absorbed by general AI. Lexi's answer is that the edge sits in what a general model cannot see: the company's own data and its earned understanding of its customers. The risk is that companies adopting AI the same way as everyone else in their market erase that edge on their own. As she puts it, "I think a lot of people just sort of washed out their differentiation."

For a CEO, this reframes the AI decision. The question is less about speed of adoption and more about what the adoption is built on.

Where Competitive Edge Lives in the AI Era

Lexi's view starts with how foundation models are built. They are trained on nearly all the data their builders can obtain, which makes them broad and capable. It also means they have not trained on the data that makes any one company distinct. Her example is a manufacturer with a specific configuration of fabricators, each with its own quirks that affect production. A general model has no view into that, and if any of it has leaked into a training set, it has been diluted among everything else.

That is where she locates the edge. In specialized domains, especially over large bodies of structured data with complex interactions, models trained on a company's own data still outperform the larger general models. Beyond the data itself, she points to the logic of a company's workflows and the language of its niche. A general model can layer in some of that. In narrow, cloistered markets, it falls short.

The strategic consequence is direct. When a company says its insight into its customers, its workflows, and its data is its differentiator, the AI architecture has to carry that insight forward rather than flatten it. Lucidity's approach is one expression of that logic: small, specialized models trained on the company's structured data, which AI agents can call for answers that are deterministic, traceable, and auditable, while the agent keeps its flexibility on top. The technology serves the strategic point. The edge has to be built into the system, not assumed to survive it.

Why Fast AI Adoption Did Not Create an Advantage

When Lexi first asked executives about their edge, many answered with their AI posture: they were adopting aggressively. She does not dismiss that, but she is clear about what it produced. Companies that adopted early models built their adoption strategies around those models' limited capabilities. When newer models arrived, they had to rebuild those strategies from the start. In her assessment, they were not that far ahead of everybody else.

The deeper cost was convergence. Without specific insight or specific data guiding its use, AI applied across a sector tends to homogenize the companies using it. Lexi describes companies diving in to use it for everything and watching their brand tone and product differentiation disappear. The pattern reaches strategy as well. A team that asks a general model what its business strategy should be will get an answer shaped by the general case. A more targeted conversation is possible, she notes, but only if leadership already knows its differentiation well enough to direct it. Otherwise, "you end up looking like everybody else."

She is equally direct about thin product adoption: "if you're just gonna go to a foundation model and reskin it and put it out there, like so is everybody else in your vertical." Companies that applied AI as a superficial layer flavored for their market had a hard time keeping their edge. Her conclusion is that preserving differentiation requires a deeper integration.

A related belief compounded the problem. For a period, a view circulated in some investor circles that data assets might no longer be a moat, because large models would be able to do anything. Lexi is careful not to single anyone out, and she notes the view was not held everywhere. Her observation is that the view hurt the companies that accepted it, because it led them to discount the asset she sees as their clearest edge.

Being Early to the Market

Lexi saw this dynamic from the vendor side before many buyers saw it in their own businesses. Lucidity Sciences has existed for about two years, though the research behind it ran far longer and required developing new types of math and computing paradigms to make compact models work. The founding logic was a deep-tech bet. In her words, "The bet was sort of that if you make better, faster, smaller models, the market's gonna need them somewhere."

The timing tested that bet. "I think we were probably just a touch bit early to the market," she says. When Lucidity first approached potential commercial beta partners, most were still in what she calls the honeymoon phase of large language models. Those who had hit the early limits were only starting to experiment with retrieval-augmented generation. They had not yet built enough with AI to see where the technology fell short, or where they needed to carve out their own differentiation.

That left a constraint no pitch could remove. Lucidity had to wait for buyers to "burn their hand on the hot stove" before they could understand why their approach was not working and what they needed instead. For a founder, the decision logic here is worth noting. A buyer who has not yet encountered the problem cannot be argued into it. The market's own experience has to do part of the work.

That experience has since arrived. The issues Lucidity raised early, including data sovereignty, access to GPUs, and the cost of running large models, are now pressing concerns for the companies it talks to. The shift also shows up in how executives describe their advantage. "People go to their data a lot these days. They didn't at first."

What Customers Valued That the Team Did Not

Lucidity's early positioning was a better, faster, cheaper single model. It found natural buyers in settings where a fine distinction at high precision carries large value: trading, insurance, and large international scientific organizations. One early clinical collaboration improved the prediction of a single adverse event in hospitals by what Lexi describes as pretty significant margins, and the work was published with those collaborators. That demand still exists. But the team had a much looser idea of what most customers would do with the technology. "We weren't sure what people were gonna do with that yet."

The answer came from customers rather than from the roadmap. Lexi says the insight about where the product delivers the most value was not something the team dreamt up and then tested. Conversations with early commercial users kept surfacing the same request. Buyers had data, dashboards, retrieval, and agents, but they wanted their systems to predict what would happen and to recommend what to do. After the single adverse-event work, others in the hospital space asked to predict several events at once and use those predictions to optimize their operations. Lucidity could build a model for each need, which raised a new question for customers: what to do with all of those models.

The demos made the gap visible. Lexi describes the team building toward what it considered the big reveal, only to have customers acknowledge it and then lean in on something else. They wanted to know whether training was really that easy, and whether the models were really that small. Those features had been built as table stakes. One-click training and lightweight models let customers see that they could store hundreds or thousands of specialized models and manage them like a database, with an agent pulling the right one when needed.

Lexi gives the credit plainly: "the people that we were talking to saw this as something more significant than we realized." She calls it one of the company's big breakthrough moments, because those table-stakes features opened a way to bridge analytics and agents, a gap most people had not yet named. The decision lesson for a CEO is specific. The features a team rehearses as the headline are not always the ones buyers act on. Listening closely to what customers get excited about can reveal where the real value sits.

The Catch: Clarity Around Your Differentiation

A technology that is better, faster, and cheaper invites a fair question about the cost. Lexi raises it herself. "There is a catch. Like there's definitely a catch." Her answer: "And the catch is clarity around your differentiation."

She extends that clarity to the customer, the product, and how to bridge them. None of it can be delegated to a model. The engine Lucidity built synthesizes the patterns in the data it receives. It says nothing about which inputs matter or what to do with the outputs. Her clinical example makes the division of labor concrete. A clinician identifies the inputs believed to raise a patient's risk of delirium without knowing exactly how they interact. The model finds the statistical patterns and produces a prediction. What happens next is a clinical judgment. "Those decisions remain with the humans as I think they should."

This is where the earlier failure pattern and the solution meet. A company without clarity about its edge will feed a general system general questions and receive general answers. A company with that clarity can decide which data goes into which model, keep the results auditable and in-house, and retain control over its own decision processes. Lexi frames that control as a form of sovereignty over the business, and she contrasts it with outsourcing decision-making to systems that are not trained on the company's data, which she expects to bring weaker performance and higher costs.

What This Means for CEOs Making AI Investments

Lexi's argument does not oppose AI. She supports agents and large language models, provided they run on top of specialized models trained on the company's own data. Her point concerns sequence and foundation. The decision that determines whether an AI investment builds advantage is made before the technology is chosen.

The risks of skipping that decision accumulate quietly. Differentiation erodes as products, brand voice, and strategy converge on what everyone else in the market produces. Dependence grows on token-based costs and on GPUs that are hard to secure. And more decisions shift to systems that have never seen the company's data.

"What's your edge" is a question Lexi says she asks executives often. It is a useful test for any leadership team weighing its next AI commitment. A strong answer names a specific body of data or customer insight the company owns and a general model cannot know. An answer that could describe a competitor signals the clarity gap she calls the catch, and it is the place to start before the next AI investment, not after it.

Soundbites

“I think we were probably just a touch bit early to the market.”
“We weren’t sure what people were gonna do with that yet."
“And so you you end up looking like everybody else.”
“I think a lot of people just sort of washed out their differentiation.”
“People go to their data a lot these days. They didn’t at first.”

About the Guest

Lexi Pasi is the founder of Lucidity Sciences, maker of the LumaWarp ML engine, which lets companies build libraries of small, specialized models from their own data. She holds a PhD in mathematics, brings 10 to 15 years of machine learning research, and has worked in enterprise corporate finance and startup data teams.

Connect with her on Linkedin: https://www.linkedin.com/in/alexandrapasi

‍

__________

About Jeff Holman and Intellectual Strategies

Jeff Holman is a CEO advisor, legal strategist, and founder of Intellectual Strategies. With years of experience guiding leaders through complex business and legal challenges, Jeff equips CEOs to scale with confidence by blending legal expertise with strategic foresight. Connect with him on LinkedIn.

Intellectual Strategies provides innovative legal solutions for CEOs and founders through its fractional legal team model. By offering proactive, integrated legal support at predictable costs, the firm helps leaders protect their businesses, manage risk, and focus on growth with confidence.

__________

About The Breakout CEO Podcast

The Breakout CEO podcast brings you inside the pivotal moments of scaling leaders. Each week, host Jeff Holman spotlights breakout stories of scaling CEOs—showing how resilience, insight, and strategy create pivotal inflection points and lasting growth.

Listen and subscribe on your favorite podcast platform:

Apple

Spotify

YouTube

__________

Be a Guest on the Show

‍Want to be a guest—or know a scaling CEO with a breakout story to share? Apply directly at go.intellectualstrategies.com.

‍

Book a free strategy call with Jeff's team - https://www.intellectualstrategies.com/strategy-call/start-here

Get the weekly newsletter - https://podcast.intellectualstrategies.com

‍

‍

TRANSCRIPT

TRANSCRIPT SUMMARY:

00:00 Differentiation in the AI Era
01:14 Mathematics as a Cheat Code
10:03 From Research to Entrepreneurship
16:50 Leveraging Proprietary Business Data
21:47 Building Small Specialized Models
29:00 Creating a Super Executive Team
33:50 Arriving Early to Market
39:05 Preserving Your Competitive AI Edge
42:20 Customers Shaped the Product
45:02 Predicting What Happens Next
52:15 The Catch Is Clarity
58:28 Solving Compute Through Mathematics

‍

‍

FULL TRANSCRIPT:
Alexandra Pasi (00:00)
And the catch is clarity around your differentiation. And so you you end up looking like everybody else. People go to their data a lot these days. They didn't at first. And there is a catch. Like there's definitely a catch. And the catch is clarity around your differentiation.

Jeff Holman (00:14)
Welcome back everybody to the Breakout CEO podcast. I'm your host, Jeff Holman, and I am here today with Lexi Posse from Lucidity Sciences. Lexi, great to have you on the show. Yeah.

Alexandra Pasi (00:24)
It's great to be here. Thanks, Jeff.

Jeff Holman (00:25)
Uh yeah, glad to have you. And I'm gonna give everyone a heads up that you're way more advanced technologically than most of us, probably. So I'm gonna ask some questions and we'll we'll get into some of the I'd love to have you explain your your your business and your your technology that you're building. Um

If if I have to I hope I don't slow the conversation too much from the momentum that you typically have when you're talking about about this with your customers and such.

Alexandra Pasi (00:51)
Happy to go wherever the conversation takes us.

Jeff Holman (00:54)
Yeah, it'll be good. Um so before we get into your into your business though, uh you've done some you've done a PhD in mathematics, which I think has probably set you apart from a lot of the CEOs we've talked with. Tell me tell me kind of what what took you down that path? What what made you want to go get a PhD in mathematics?

Alexandra Pasi (01:14)
Ooh, that's an interesting question to start with. I love it. you know, I I think I've always been a a pretty curious person, but I've always had a lot of different interests, you know. And uh I think math was kind of my shortcut way of doing it all, right? So mathematics is sort of the uh the study of structures and patterns in generality. And so if you study math, you kind of get this little cheat code to, you know, I'm gonna study this particular type of, you know, game theoretic construct, and now I get to

actually secretly be studying the market and crowd dynamics and, you know, all sorts of, you know, games in terms of social media engagement. I mean, there's there's these uh very high level paradigms that let you study patterns and then apply them in all sorts of different contexts. And I think that was part of the draw uh for me towards math. Yep.

Jeff Holman (02:03)
what point do you would you say that that because I think a lot of us who there's a couple of camps, right? Like I did I did fine in math. My daughter uh actually graduated with a degree in math and wa is doing her PhD in finance. So she's a she's exceeded, excelled past me in in my math. But I have an engineering background, so I did a fair amount of math. But I think there's I I feel like there's kind of three camps, right? You have the people who say, Yeah, math's not really for me. I did what I needed to and I got out.

Then you have people, maybe like me, who were decent at it and got good at it and we we see the usefulness of it. But then we have you and you're like, when people start talking about math helps you see patterns, is that a different level of thinking about uh about it just a different perception of what's around you than the rest of us?

Alexandra Pasi (02:49)
Yeah, I mean, I think that's a good question. And and not to make this a referendum on education, because I think it's a really uh nuanced conversation, but I do think that a lot of the way most people experience math through their educational arc, I would compare it to as if you studied art without ever being allowed to pick up a paintbrush. And so I do think that there's a lot of like rote memorization kind of machine type manipulation.

Jeff Holman (03:15)
And cheats. Yeah.

Alexandra Pasi (03:17)
Exactly, exactly. That is focused on um without ever really getting to ent the phase of okay, now you're gonna go create something that has this sort of shape, and then you describe it mathematically and then you figure out the machinery. Uh there's a lot more kind of license and autonomy at that level, but unfortunately we don't invite people into that side of math until typically around grad school.

Jeff Holman (03:41)
Okay. Okay. So I might have I might have uh finished my math studies a little too soon to get to that point. But but I love it because because when you start saying, Okay, we're gonna talk about patterns, now you're really math feels uh early on so specialized. I'm gonna be a math. And when my daughter said, I'm gonna do math, I'm like okay, w you know, you were in a you're gonna do accounting, you you consider engineering. What are you gonna do with math? You know? Um i was my question. But but

Once you understand it the level you do, I I believe the answer is you can do whatever you want with math shows up everywhere from economics to to I I mean, I don't know, probably food science, probably totally. Definitely the AI stuff that we're gonna get into, but but everywhere, right? Math is math is just the language that underlies the physical world. Is that a good way to s is that okay to say that? Yeah.

Alexandra Pasi (04:34)
Absolutely. And I I think, you know, that that's kind of a a philosophical point that some people might have, you know, uh some nuanced opinions on. But I think math underlies our descriptions of the physical world, but also our descriptions of kind of everything, right? Math underlies descriptions of language. Math underlies, I mean, we'll get to this when we get to LLMs, but you know, far beyond the physical world, I think we're even seeing the patterns of math applied to the space of concepts and concepts.

creativity and all sorts of symbolic representations. Um, and I I think that it's that's actually the secret behind why AI is able, you know, AI as we understand it today, is able to do what it's able to do.

Jeff Holman (05:19)
I I think we get we're gonna have to dig into that secret a little bit more here in in this episode. that that to me is uh like that's where math um in a way kind of crosses over into entrepreneurship. If if I can take some liberty here. Like I think what you're describing is that that people get into math and all of a sudden you can say, well, I like I want to describe what's happening in the world better, so I'm gonna come up with a construct.

And I'm gonna say this is how this is one way to describe that thing that we see or hear or you know that happens around us. We're gonna just describe it with math. And and so you're you're kind of creating something out of nothing, or you're creating something t that didn't exist to describe things that do exist. It business to me seems seems kind of the same way. You're saying, listen, there's I I see things, I see pattern, maybe patterns or problems or

Whatever you want to call it, I see these things happening around me. Why don't I just create this new layer, this new construct, this new algorithm and and and you know, offer it to people to address the problems they're having? Is is there a I I feel like I'm I feel like I'm stretching here a little bit to to reach to your level that you're at, but but to me that it seems like you're creating something out of nothing in both i in a way in both of those environments. Is there some parallel there? Yeah.

Alexandra Pasi (06:44)
I I completely agree. And it's very much the way I've approached entrepreneurship. Um, and you know, it's interesting. My first sort of corporate job at the, you know, um, where I really dug my hands into the what was then called the big data, you know, boom. And I I mean a lot's happened in the ecosystem, but was embedded within the corporate finance team at an enterprise. Okay. And um that was, I think, you know.

A financial model, financial projections, kind of what-if scenarios that you use to guide a lot of executive-level strategic decision making, um is a mathematical model. I think oftentimes in in finance or in business, we have certain templates that we're used to using from, and then we just kind of fill them in as we're thinking about projections. We don't think about the fact that

Somebody at some point sat down and was like, okay, this is the kind of shape of these sorts of business decisions. This is what CAC looks like. This is what COGS looks like. This is, you know, how this sort of like scaling law is going to happen. And then you put it in a template, eventually it just becomes numbers that you can fill in in a spreadsheet. But you know, if you're trying to do something really creative or you're trying to capture a nuance about some interdependencies within your business, that modeling exercise has to step out of the template, you know, because you're kind of

breaking the mold, you're going outside the box, and you have to then figure out what the way of describing all of that extra stuff is that fell outside of your original Excel template. That's what happens when we do like financial modeling and financial projections. And then I think so that's kind of maybe one that probably everyone can relate to that's worked in a in a sort of executive setting. But then to your other point, kind of which is more architectural, you know, we sit within this, I mean the

The depth of the technology ecosystem that people are building their products off of grows every year. And so you can certainly think of that market and that ecosystem as, you know, this large system describable by math, understanding where the gaps are, understanding where the connecting pieces are, understanding kind of like the volume of demand that needs to go from point A to point B and like what track is it going through. That's all.

a fundamentally mathematical question, or at least it can be modeled mathematically in a way that lets you capture more complexity than you would able be able to encapsulate within your mind as you're just kind of thinking through these problems.

Jeff Holman (09:10)
Yeah. Man, fat uh so it makes me it begs the question as you're in corporate finance, you're dealing with templates. At what point did you say to yourself, I think I'm gonna go and create something that doesn't exist? I'm gonna I'm gonna move from templates to the next, you know, the the next evolution of of my career. How what was that process for you? Just a quick note about our guests. I host the breakout CEO podcast to share behind the scenes insights from scaling businesses.

As an attorney, I see the real challenges leaders face long before success becomes public. But clients' stories have to stay confidential. So we invite guest CEOs to share their own moments of struggle and success. I'm so grateful to our guests and my team at Intellectual Strategies for making this show possible. Now, let's get back to the show.

Alexandra Pasi (10:03)
Yeah, it's so I I think it it was kind of the the merging of two worlds. You know, I'd been working on machine learning research uh for a long time in some form, um, for probably like a decade at fifteen years at this point. Okay. But I was also, you know, I had started to do consulting for startups, um, got brought in with one that was growing very quickly, sort of built up their data science and analytics team. So I got to see a little bit more of that entrepreneurial side, you know, moving out of a more

kind of backroom corporate type role. But then on the side had this kind of continuous arc of machine learning research. And one of the interesting insights that it gave me, I think, was these three layers of need within the marketplace. Right. So understanding kind of strategically the evolving role that data assets were having for businesses. This was, you know, starting back in 20 uh 2020.

So I think that was when it became a little bit more mainstream to be doing these large scale um analytics activities as like just table stakes within the enterprise. And then at the kind of architectural level, you know, because I was building data teams, I was watching people's workflows, I was working really closely with those tech stacks, um, saw a lot of the gaps in the tools that people had available to build the products that they were taking to market.

And we we'll I'm sure we'll talk about that a little bit, but you know, as a little bit of a spoiler, a lot of it had to do with how pattern finding scales and kind of the nature of that uh both relative to data and compute. But that that last level and that kind of fundamental level was tying this back to the mathematical bottlenecks, right? So I knew a fair amount about how machine learning and you know, AI, and that those terms have evolved quite a bit, but um knew how that worked theoretically.

And knew that there were some pretty big theoretical gaps. And I think the observation that, you know, those theoretical gaps actually had a continuous thread all the way through, you know, what's the tech stack? What are the tools up to the strategy? What can people do? Um, gave me that insight to try to create a better path from A to B that, you know, companies were not equipped to undertake because the level of mathematics.

that had to be engaged with to close those gaps was far beyond anybody they would have staffed.

Jeff Holman (12:32)
So so just so we have context here, people listening, who who are the types of companies that I mean when we're talking about big data, I think a lot of us think, okay, huge enterprises, very large uh data analytics teams. I who who are the pe the companies that you're talking about? And what what prob what business problems were they solving with with their data? Yeah.

Alexandra Pasi (12:52)
So I I think, you know, I I've worked everywhere from kind of startup to enterprise. And one of the big shifts that I think we saw around 2020, especially with like larger third party data markets, um, was that big data was available to more people, right? And I I think a lot of companies, um, both trying to use their internal data and also purchasing data solutions from the outside were were um trying to use this new you know, kind of

gold like asset to create products um from you know everything to like specialized you know hyper specialized marketing agencies to um you know solutions within life sciences or maybe even for their own internal use so I might be trying to do something like revenue forecasting or I might be trying to

optimize my spend across different channels or I might be trying to create a system by which to gate, you know, what customer is a good customer, what customer is a bad customer, how much should I invest into each of them in all of these different touch points. And so, you know, some I I I would say there's kind of a distinct the biggest distinct difference I would say is not size of company, but what they're doing with the data. Um so some were repackaging that data as a product to sell.

Right. So I'm going to take all this different data, some of my own, some from third party sources. I'm going to combine it. I'm going to put it in a nice little UI so that I'm synthesizing really custom high-level insights for a specific type of end user. And I'm going to do all of that work so that they don't have to, right? Because it is hard. Like if if that company wanted to build that, they would have had to have one of those big custom analytics teams. But you saw a lot of vertical SaaS companies finding these really specialized niches and kind of tackling that.

problem within that particular vertical or that particular niche um for their customer for their market. But the other thing that you saw a lot of was these companies start to get into a higher level of data informed decision making uh using just their own internal data, right? So the second piece is internal analytics. And I would say that companies that use their data for products um and companies that use their data for internal analytics and decision making.

Are usually fairly different in terms of the way that they approach the question of their data assets. Okay. The the one exception that kind of lives between those two is if you're playing what I would call kind of like a casino game that's like super highly dependent on making really fine decisions at high volume. So, you know, you see this a lot in uh financial institutions or trading um insurance companies would often fall into that camp beyond a certain

us, but it there's like a spectrum.

Jeff Holman (15:47)
Okay. Yeah, th that's interesting. So as you're as you're describing that, I'm thinking of again back trying to relate this back to what I'm familiar with. My daughter as she's doing PhD research, she's talked to me many times about how um how valuable these data sources are. She's like, I've got the data for this, so I can actually do this. Or or I've got a question, but I don't have the data, so I can't really address that question until I find a data source that that works. And so and when we heard what was it, I don't know, three or four years ago, people were talking about

Um or maybe maybe it was maybe it was two or three years ago. People were as AI's, you know, coming out, you're people are talking more about I have, you know, data, how am I gonna leverage my data? Um, we're gonna package it, we're gonna make it available, we're like data's gonna become one of the primary IP sources for for companies type of thing. So are uh so so Lucidity Sciences and and what you're doing there, are are you looking to make

Are you looking at to address one of these factors or is it something else that you're doing?

Alexandra Pasi (16:49)
Yeah, that's a great question. So, you know, the the gap that we fill in the market right now, and I'll talk first about what the product is, what we're doing for our customers. If we have a little bit of time, we can get more into the technology and kind of where we can go longer term. Cause that's also very exciting. But um, I think the the obvious place to start is what are we doing for our customers and users right now? Yeah. Um and it's exactly like you said, this big gap is

How do I best leverage my unique proprietary data assets, right? Um, in an agentic world, in a world driven by AI. I think there was, there were some answers to that, um, you know, say five, 10 years ago, that don't really apply uh in the same way in a world of agents.

Jeff Holman (17:41)
And the whole tool set's changed. Like, we're

not we're not using hammers and screwdrivers anymore. We're using something totally different.

Alexandra Pasi (17:48)
Like

we're we're we're well out of just the dashboard era. Right. I mean, not the dashboards don't have a place, but like for a while that was that was the big thing. That the big thing was the semantic layer and your dashboards and your executive reports and so many things are agentic now. and and there's there's a an another another side of this too, in terms of kind of competitive pressures, which is that, you know, as these foundation models get better, a lot of people are wondering, where's my competitive edge? you know, is the offering that I'm providing to the market going to get swallowed up?

By the expanding capabilities of these foundation models. Yeah. And in that light, the question of well, I have my proprietary data assets. What can I do with that? How can I leverage it in a way that brings value that these large models, you know, being very large general solutions, being expensive solutions, and not having insight into that data cannot provide to my customer. And so that's where this question of how how do I use my data?

To greatest effect in the agentic era really becomes kind of an existential one. And so what Lucidity Sciences does is provides a sort of layer between your data and your agents that lets you create really specialized models, especially on structured data, that the agent can call and provide like a deterministic, traceable, you know, auditable answer from your data.

but you still have all of the flexibility of the agent on top. And we have very lightweight specialized models that you can train, um, you know, because of our engine, our Luma Warp engine, that lets you build out a library of hundreds or thousands of those models. Um, so basically you're powering your agent on models trained from your data for specific functions, but you still maintain all of the flexibility of the agentic interface.

Jeff Holman (19:42)
How does that differ from what other people are doing right now? Or the uh uh you know, I've I've read your website, I've got some hunches, but but it seems to me that the the trend has been everybody out there is saying, I've got the most data. I've got the biggest data sets, I can therefore probably do more than you can do, or you know, we're better than you 'cause we got a bigger, bigger amount of data. Is that is that the trend that's been happening?

Alexandra Pasi (20:08)
well, I think that's certainly some of the argument that is made about the foundation models, right? Is that and the way the foundation model works is you just get pretty much all of the data you can get your hands on, you train a very, very, very large neural net to the entirety of that data. Yeah. Um and some of the places where it doesn't perform as well is like if you want to find the patterns that are in a very large corpus of data that the LLM was not trained.

Right. So maybe I have like I'll use a manufacturing use case because it it it makes sense. Maybe I have a really specific configuration of fabricators within my plant or something like that. And they all have their own little quirks and they affect my production somehow. Well, the LLM is not going to have trained on that. I hope not. Maybe somebody slipped up at some point and ended up in the training and said, That is a thing. But like I hope if I've been good about my security, like this is my data. Um

Jeff Holman (21:05)
even if it had besc had escaped, it was mixed in with everything else out there and it's it's been diluted, right? So

Alexandra Pasi (21:11)
That's exactly right. And so in in a lot of these specialized cases, especially over larger bodies of data, structured data, complex, nuanced interactions, the specialized models do still outperform in those domains the larger foundation models. and especially, you know, you hope if you have that edge. So, like that's your edge, right? This is mine. I do it better than anyone else. I see it better than anyone else. How do I make sure that level of fidelity and insight

Is actually carried over into the agents that I'm producing to run my business, to offer insights externally as products, that sort of thing. And, um you know, what we're able to do is create much more compact models with top level performance around these specialized domains. So the ability to train specific models in a specific domain from your data has been around for a long time. That's kind of what you hear called now machine learning.

Okay. Kind of a weird evolution of those those words. But usually when you hear machine learning, it means taking my data and training a model to it specifically around a really particular use case. The the challenge with that is that in the old way of doing things, it was either very labor intensive or the models were very large and cost intensive to use. And so people would usually have one or two flagship models, you know. So I'm like,

Okay, my revenue projection model is my flagship model, or my um customer selection model or you know, customer product pairing model is my big flagship model, 'cause they were just very expensive to maintain and produce.

Jeff Holman (22:52)
Meaning if I was a business and I said, Okay, we put our our team of data scientists on the data we have and they've they've come up with buckets of data. Again, I hope I'm not dumbing this down too much. Come up with buckets of data and they've come up with some applications for that data. And we're like we're gonna just because of resources, we're gonna put our resources on one bucket for one application. We'll build that out as it'll be our premium, you know, go to model because we're gonna use it a lot. It helps drive our business.

But we've got these other things, th th this other data that that's not really we're not really optimizing for. And so everything kinda gets skewed towards one type of data, one application. Is that what you're saying?

Alexandra Pasi (23:33)
Yeah, absolutely. Um, and it leaves a lot of it leaves a lot on the table. But the other thing that's interesting about that too is that when you talk about agents, um, an agent is often being asked to do all sorts of dynamic functions. That's kind of what the chat based UI gave us, right? Is that you know you can ask anything and it's gonna do a thing with it. And so you can imagine that the agent would really like to have more models under its hood that it can call in the appropriate context.

So, you know, if you're asking a question that involves understanding revenue and cost and, you know, the shape of my machines on the ground, like it needs to be able to synthesize all of those pieces. And um, especially if you're talking about a setting where not only am I using my agent internally, I'm letting my customers use it. My customers are putting data into my system. You know, it this can quickly blow up from I need one or two models and I'll get away with that.

To, well, really, I need hundreds or thousands of models. And maybe they even need to be able to be created dynamically. And so there's this question of from a labor perspective and from a cost perspective on compute. So actually using models, how do I scale and support that? Um, and that's been kind of part of the big open question that's meant that you couldn't bridge between analytics and agents in a way that incorporated these specialized models really in the way that they

Would need to be used.

Jeff Holman (25:02)
And so you're sitting there saying, Okay, I've got a math background. I see the patterns happening here. I see you know, you've been in machine learning for ten or fifteen years, and you're and you're saying, Well, w the the models are kind of a limiting factor that if we could if we could create more models, maybe on demand or faster, or so they they run cheaper, so they don't cost as much to create or use or you know, uh replicate whatever the word is. If we could if we could create

if we could address the model creation point and the model usage point in the system, now we can unlock different levels of data usage.

Alexandra Pasi (25:39)
That's exactly right. And so the constraint then becomes from this kind of tool level, right? Product level. Yeah. Um, the gap is I need models that are like one click to train. So I don't need a data scientist that's very specialized being like, maybe I'll use a neural net, maybe I'll use you know, trees, maybe I'll use like because that's I mean, not how people work anymore. That's very expensive. That doesn't scale if I'm talking about 10, you know, hundreds or thousands. So I need all of that kind of one clickedness.

Um, but I also need the models to be really small and fast and lightweight. Yeah. So I need an agent to be able to pull from a library inference very quickly, and I need that not to be able to blow out my budget. Right. Um and so the question becomes, well, what kind of thing can do that? And the answer that the machine learning world had provided up until now was basically like there's not a thing that can hit both of those marks of.

Jeff Holman (26:33)
What does it unlock by by by being more nimble with the m the model layer?

Alexandra Pasi (26:38)
Yeah, it basically means that you can have agents doing what they need to be able to do, which is uh, you know, operate from a large number of specialized models. Um, the the other thing that this allows you to do too, is that if you are letting those specialized models do most of the heavy statistical thinking, you know, projection simulation, you don't try to force the larger model to do that.

One, the larger model is probably not going to be as good because it's not trained on your specific data. It's architecturally limited in terms of how it can synthesize those patterns. But two, if you really do want to try to make an LLM reason over that, you're going to be stuck with the larger models, right? And so that's a lot of token expenditure. That's trying to figure out how to allocate GPUs for those larger models. that can result in a lot of latency. So one of the things that this lets you do, beyond just, you know, create

better models, faster, cheaper, is actually lower your token costs for the LLM that sits on the top layer because you can use a lighter weight model because you're not asking it to do as sophisticated of reasoning.

Jeff Holman (27:48)
Which which is the biggest thing that people are tackling right now, right? Is is oh yeah is the data usage, the cost of that usage, the the data centers that need to be built and put underwater or put into space or whatever wherever we're contemplating right now.

Alexandra Pasi (28:02)
Well, and and the other thing too, I mean, you know, yeah, I mean, there's there's this huge world level impact in terms of the necessity of compute, you know, and how do we kind of how do we get the same result but keep that down from cost perspective, from an ecological perspective? but I mean, from an operations perspective, a lot of people are reporting difficulty finding GPUs, you know, for for their um operations. And so just because 'cause of the shortage.

And so one of the things that's really nice about our models is that once they're trained, they inference on CPU, um, which is not supply constrained. So you have, yeah, you have an additional amount of compute flexibility.

Jeff Holman (28:45)
You're actually able to move where the compute is happening uh if in in the physical hardware so that it's you've got access to more computing power.

Alexandra Pasi (28:55)
That's exact there's always a CPU available. There's often not a GPU available immediately.

Jeff Holman (29:00)
Interesting. How interesting. So I mean this is this is complex stuff. Um the uh like the the analogy that's popping into my mind is uh I'm thinking of an executive team in a business and and this is kinda like saying, Hey, we've got we've got one, you know, mastermind CEO and and that's the go to model. The the CEO has to make every decision and every decision takes a while 'cause there's so many things they have to do versus what what you're describing is a bunch of, you know, you're able to provision

you know, a thousand different executive leaders all on the team and you get the right one for the right task at the right time and uh and they can, you know, I'm sure they can talk back and forth. So all of a sudden you have like a a super executive team uh that's that's different from uh just a normal executive team where you have kind of limited limited capacity. Or worse, uh you know uh not not worse, that's the wrong word, but more constrained is a smaller team where you have

you know, one person or two people or three people instead of seven or eight or nine, you know, people all with the experience and bringing bringing the context in. So I that that's probably a bad analogy, but

Alexandra Pasi (30:08)
No, no, it's actually a really good one. I mean, it that's exactly the right way to think about it. And and like, you know, if if I'm like, imagine I'm running a company and I've got that CEO. Do I think it would be better for the CEO to receive from their executives, like, here's my revenue projections, here's my cost analysis, here's, you know, all of these different pieces that are auditable, that are clear, that deterministic, that are kind of specialized, you know, trained by a specialist in that area.

And then the CEO takes that and it makes decisions from there. Or do I think it's better for the CEO just to be given access to terabytes of data?

Jeff Holman (30:47)
Well

good luck. You'll figure it out. Yeah. Yeah. Yeah. I'm sure some CEOs are like, I've been there. I've had Yeah, exactly. It seems with the complexity here, is is this an easy business to build with what you're doing? I mean, what's what's worked so far? I I think I read that you've been doing this for the last maybe four or five years. Is that right?

Alexandra Pasi (31:10)
So it's actually only been it's actually been about two years um that the company has existed. Now the the research technology has been going for um quite some time before that in various forms. You know, we had to and this is where the math comes in, but we had to develop like whole new types of math and computing paradigms to make that model compactification work.

Jeff Holman (31:32)
Okay. So you're you're you're inventing new new algorithms out there. You're you're inventing new models. Um it so in the last couple of years, that's a short time, wha what's been working well? What w where are you finding your wins?

Alexandra Pasi (31:45)
Yeah, you know, I think that what's really interesting to us is we're we're vertical agnostic. Um, you know, basically if you have a bunch of the structured data and this is the kind of thing you're trying to do where you you need to do it faster, cheaper, better, um the product can work in any of those settings. Yeah. But what's been interesting to us is that organically there's been some kind of concentration of vertical interest rate and engagement. So a lot of our customers come from

either finance settings or sort of like life sciences, healthcare, bio settings. Yeah. And I think that some of the reasons for that are that those people are a lot more concerned about uh sort of data sovereignty and privacy. Yeah. Um, and they're a lot more concerned about understanding why models do what they do and making sure they do the same thing every time. You know, so uh it's it's no good to get two different projections um, you know, if I'm if I'm a day trader.

Like yes, buy the stock or no, buy the stock off the same prompt and the same data with no change besides that I just ran, you know, ran it twice. And that's that's what you do get with LLMs. And that can be kind of um, that can be a whole new layer of risk that is difficult to juggle depending on the deployment setting. So so that's that's been an interesting thing that we've seen in terms of who's interested in understanding um data privacy, you know, having a model that has sort of

transparent outputs, that sort of thing.

Jeff Holman (33:13)
Maybe accuracy, but at least consistency.

Alexandra Pasi (33:15)
Definitely accuracy. Well, and and that's another good point too is that a lot of people are in settings where um every marginal bit of accuracy improvement uh translates into, you know, millions or sometimes even billions of dollars and increased returns. So so, you know, you're not gonna just settle for the quick and dirty solution necessarily if you're trying to, you know, if you've got, you know, many millions of dollars on the line.

Jeff Holman (33:41)
Yeah, of course. But as you as you've built, um where have you where have you run into the r roadblocks along the way?

Alexandra Pasi (33:49)
Well,

I think starting out was probably the most interesting one. Um, so I think we were probably just a touch bit early to the market. Um, because when we first started talking to potential commercial beta partners, um, most people it depends on the geography, but most people were still very much in this era of um kind of what I would call the honeymoon phase of LLMs. Yeah.

Um, my gosh, this is amazing. They can do anything. And then people who were kind of running into the limits of that were only just starting to do what they call rag, right? Which is basically just, well, give the LLM access to your data in kind of chunks and then pull it in whenever it needs. retrieval augmented generation. Yeah. So it's it's like generate a response based on data that you retrieve based on the user query. And that was like

the big pattern at the time. And people think very poorly on RAG these days. I think there's several reasons for that. Some that I agree with, some that I don't. But, you know, point being, um people hadn't done enough, built enough with AI yet to understand where, you know, where both the limits of the technology were, but where they were supposed to carve out their competitive differentiation.

Right. So one one of the things I think we've kind of alluded to is that, you know, AI as it's used across whole sectors has a kind of homogenizing effect in the absence of specific insights, specific data. And so a lot of people were just diving headfirst into like, well, we'll use it for everything. And their their brand tone disappeared.

their product differentiation disappeared because they just started to look like everybody else. And so we did have to wait um for people to kind of burn their hand on the hot stove, so to speak, before they could even understand, like, oh, this is why this wasn't working. This is what I need to do.

Jeff Holman (36:01)
Because everybody was just so enamored and they're like, I love it, I love it and then they're like, Wait, so so you're saying the AI slop happens in data in data analytics, n uh like on the data team side, not just showing up in the the day to day use of, you know, w drafting whatever, you know, blog posts, LinkedIn posts, whatever people are

Alexandra Pasi (36:21)
Yeah,

definitely. I I mean, I think like um and I would say it shows up even at the strategic side, right? So when you think about tech strategy or when you even think about your business strategy, I I see a lot of people going to it to brainstorm. Tell me what my business strategy should be.

Jeff Holman (36:35)
And then it's a and it's coming up with the the av the average strategy, if you will.

Alexandra Pasi (36:39)
It's exactly it. Yeah, that's exactly it. And so you you end up looking like everybody else. Um, it's not to say you can't guide a conversation that's more targeted around your specific differentiation opportunities, but you kind of have to know those already, right? You have to prime it in that way. Um, and I think a lot of people just sort of washed out their differentiation. I even heard so for a while, people were

And I don't think this is true in every every geography. And I'm not trying to call anyone out in specific, but you know, in there is this kind of um aphorism in investor circles for a while that like, well, maybe data assets are not gonna be a moat um going forward. Cause like LLMs will just be able to do anything. Yeah. And some people bought into that and that really hurt them because now you go talk to them and they're Yeah, like data assets are, you know, or some sort of, you know, but there's not a lot of places moats can be anymore. It's not gonna be like

software development in the same way that it was. Right. Um, so I I think there was just a lot of confusion and noise because uh the technology, you know, AI is a technology. To really understand it, and I know a lot of people don't like to hear this, and I understand why, but to really understand it, you have to understand technology to some extent. Or you're just going to be listening to kind of the the chatter, you know, um, the chatter on the airwaves without any way of discerning

What's real, what's grounded, what's not. It's not to say that you're not still guessing at the end of the day, but it's har it's really hard to do that without some technical understanding.

Jeff Holman (38:14)
In a way it sounds like you might be you might be saying that that uh there's uh not just AI slop outputs, but maybe there's AI slop uh usage. So the the the average user without understanding the technology kinda also ends up being the average generator of average uh outputs from the system itself. So so you can't you can't be on the leading edge if you're if you're not uh technologically savvy. That that makes that makes perfect sense.

Alexandra Pasi (38:43)
That's exactly right. And and so it's like, you know, what's your what's your edge is a question I I ask executives a lot.

Jeff Holman (38:51)
What do they like to what do they tell you when when like today, now that things are maybe settling, you were early to the market, the people are catching up to where you were, you know, and your technology, they're probably seeing the usage of it. That w what do they tell you uh when you ask that?

Alexandra Pasi (39:05)
People go to their data a lot these days. They didn't at first. Um, at first it was, well, we're very aggressive with AI adoption, which is fine. But honestly, it didn't give people a lot of an edge, right? Because there were people that adopted the early models that had more limited capabilities, and they built their adoption strategies around those early models. And then when the new models came out, they had to build an entirely new adoption strategy. We weren't that far ahead of everybody else. Um, but I I hear a lot of people

talk about their data. I hear a lot of people talk about special insights, um, which I I think has always been true, right? Like, like we know better than anybody else what our very nichey segment needs. Like we know our customer better than anybody else. Yeah, we know what they're trying to do.

Jeff Holman (39:51)
Is that what special insights are? I haven't I haven't heard that phrase before. Is that a is that a phrase that's being adopted or

Alexandra Pasi (39:56)
I hear it more in kind of like like there's like this earned insights concept.

Jeff Holman (40:00)
Okay. Yeah. So i in other words, y you've finally gotten into your data to the depth and degree that you can pull things out that you didn't have before and now you know things you didn't know and

Alexandra Pasi (40:13)
Or or even just um I mean data is a huge part of it, definitely. And and then I think there's this other part too where it it's more of kind of the logic, the logic of the workflows, the logic of the language. Um, and a lot of that LLMs can layer in, but some, especially in like more niche cloistered areas, it's not gonna get as well. So and that I I I would say that between those two things, that's where this paradigm that I've been describing really shines. Because

If you're managing your own agent, you get to micromanage to you know your own degree. So however much you think is appropriate, the agentic logic. So what kind of thing does this provide in what settings? What kind of function does it run? What skill does it go to? What kind of supplementary abilities does it have? but then you're also pulling in your data and synthesizing that in a way that can interact with the agent appropriately. And so if you're saying, my unique insights into the

way this language and logic and all of this comes together and my data are my differentiated edge. Yeah. Um, then that's the only pattern that makes sense to use. Because if you're just going to go to a foundation model and reskin it and put it out there, like so is everybody else in your vertical. And and there was the the death of sort of the the what did they call it? Like the SAS racker up rapper apocalypse or something like that. Very shortly after

And and that was like the burning your hand on the stove, right? People are like, ooh, there's AI, gonna adopt it, gonna reskin it very kind of superficially or or top level with sort of a the flavor of my vertical or my market. And uh I think those companies had a really hard time. It needs to be a a deeper integration to preserve the edge. And that is that's a challenge.

Jeff Holman (42:01)
Th those those companies sparked and then flamed out a pretty quickly. I think they were the the early ones got a little bit of funding, but the they couldn't keep their edge. Well so uh a as you've built this as you've built this business in addition to being early to the market, w where else have you found that you've had to put extra attention and effort to to get to where you're at?

Alexandra Pasi (42:20)
Yeah, I mean I I think that so this whole insight about where architecturally the solution lives and is, you know, provides the most value wasn't something we dreamt up and then tested in the market. Um

Jeff Holman (42:34)
What do you mean by that?

Alexandra Pasi (42:35)
well, I meant it it came from our early our early uh commercial users, right? So it was it was conversations with them, engaging with the market, talking to people, and we'd be like, hey, here's because we had a much looser idea of like,

Hey, these are better models, they're faster models, they're cheaper models, they're trained specifically on your data. That's great. We weren't sure what people were going to do with that yet.

Jeff Holman (42:58)
Were you kind of a tech first company? I I I know people talk about, Hey, we built the tech, now we gotta find a f we gotta find customers to buy it. Is that was that kind of what you were doing?

Alexandra Pasi (43:07)
The the technology ex I mean, 'cause we're a deep tech company, right? The bet was sort of that if you make better, faster, smaller models, the market's gonna need them somewhere.

Jeff Holman (43:19)
You'll fi yeah, and and the customers will come or you'll or you'll find that

Alexandra Pasi (43:23)
Yeah. But where we had kind of started, and and I think there's still there's still use and demand for this. So I'm not saying that this this is not, you know, a place that people use it because they do, but was exactly what you had pointed out, which is if you're working in settings where every little bit of accuracy or every little bit of speed becomes your edge, like trading applications or like a canonical example, then you just need the one model. You need it very fast, you need a very good. And so that's a pattern that was easy to see.

Right. Who out there needs that kind of razor's edge differentiation? Um, and and or they need fast answers. Um, and is great or is fine, but it's you know, maybe settings like trading, maybe settings like insurance, but maybe these really high powered settings like, you know, scientific analysis or manufacturing, where like you're really trying to make a fine distinction at high precision.

And so we, you know, we did a lot of those collaborations with like very large international scientific organizations that that got a lot of use out of it because they were trying to make those razors edge distinct edge distinctions. And they were trying to um get the accuracy as high as they could. And sometimes they had enough throughput that they needed to be able to do it fast. So like there was a lot of value there. But

Jeff Holman (44:39)
Which comes from better math, right? Like I mean better math, better solutions, even if it's kind of a very targeted ex systematic approach or w whatever you want to call it. A very singular focus. Uh but the math math makes that better. I think you're gonna I think you're gonna tell us though the the more complex solution now where math also makes it easier to do more complex solutions than a singular solution.

Alexandra Pasi (45:02)
Yeah, well, and it's kind of funny that I started with the complex solution, but I I think it's because that's what resonates more with a broader group. Because definitely it's out there that, like, okay, if I can improve this particular metric by five percent, that means, you know, three hundred million dollar return this quarter for me. Like awesome. Do it. And and a lot of businesses do have something like that. And that's really easy to see. It's like, if I can do this better, faster, here's my returns. Okay.

Just the one model has all the returns I could want. But as we started talking to more of the market, we saw this need kind of emerge, right? Where people were, I mean, this was this was kind of an organic reveal for us. People were really struggling with, I have all this data, and we do these aggregations and we have RAG and we do traditional analytics, and we've got these dashboards, and this agent can use all of this. But what I'd really like to do.

Is instead of it telling me what happened, I'd like to tell me what's gonna happen. Right. Or instead of it telling me, you know, what happened, I want it to tell me what I should do. Like what should I do to make this future outcome better? Or like, what would happen if I made this policy change? So there were all sorts of gaps that um people were like, can you help us fill that? And we're like, Yeah, we could do this. You can make a model here, you can make a model here, you can make a model here. And then the question started to arise like,

What do I do with all these models? Like this feels like a lot. And they're like, wait, these are all really small. I can just manage them like a database. Um, I can just store all these models and treat these models like their data and have my agent pull them in when it needs that piece of data, that piece of information, which is really that model. And and so, you know, I think it was kind of, you know, us having this product that did this thing that we had just put up there for convenience. Um,

But the people that we were talking to saw this as something more significant than we realized. You know, we'd go and we'd do the demo and we'd think like, here's the big, here's the big reveal, like the ta-da. And you're like, yeah, that's cool. But are you telling me that this, you know, this or this in terms of, you know, it really is that easy to train? Or it really is, you know, the models really are that small and easy to pull, and like, I could do this here. And so I would say that kind of, you know, one of our big breakthrough moments was just realizing that some of the things.

that we had just put into the product as kind of like, you know, table stakes. We're like, well, if it's going to be fast, it needs to be like one click to train. And if it's going to be, you know, fast, it needs to be like small models manageable here. Suddenly opened up the ability to bridge a gap that most people hadn't even put a name to yet.

Jeff Holman (47:42)
Yeah.

I I love that you're describing this. I hope I don't forget forget some of my thoughts though, but but you're like I could it seems like there's a parallel here here. You were in finance early on, but you know, you you talk about being on the back end and and taking now taking your solutions out, say, hey, we can give you more accurate data. Sounds like an accounting function. And then you got some customers are like, Hey, wait, what about forecasting? Like I like accounting, but I like forecast forecasting. Can you do some predictive things? And so it seems like maybe there's a natural um

Like that was that was meant to happen with some of your background and and talking with customers that you'd that you'd reach the point where they're they're saying, Well, let's talk about predictive behaviors in addition to to just accuracy, uh accurate data and and stuff like that.

Alexandra Pasi (48:27)
Well, and another example kind of along those lines that was one of our first, you know, we'd worked on with some clinical collaborators was predicting adverse events in hospitals. And we had just started out trying to predict, you know, a single adverse event better, which we did by by, you know, pretty significant margins. Yeah. Um, and end up publishing some of that work with them. But then, you know, some other people in the hospital space came to us and they said, Well, this is great, but but I don't just want to predict this adverse event. I want to predict this and this and this, and then I want to take these inputs.

To help optimize my operations. And again, we just kept seeing this pattern. Um, it can be in finance, it can be in bio, but everybody wants to do the same thing. And it really just comes down to the fact that it's not just about this being better and faster, it's about the lightweightness while maintaining the perform, you know, the more accurate performance is what lets these this whole model suite scale out to do what people are trying to do with agents that they can't do right now.

Jeff Holman (49:25)
Yeah, I I I can't articulate it well, but it what you're describing sounds like all of a sudden we're we're we're starting to see the first glimmer of capabilities to have an agentic uh company that that where where agents really are uh not just reacting but are making proactive decisions in the business and helping to run the business uh more like a more like a human would and and taking

more context into account. Like that it sounds like that's where that's happening or or what you're describing. The the other the other thing that that from your prior comment that really struck me is, you know, having worked with a lot of innovators, that you really see, and it's kind of magical, you see the innovation happen when somebody really gets into the problem and they really get into the customer experience and and they start saying, well, we had this we had this general solution with well or or technology that we assumed would be a solution to people.

But once the once the applications came to light, once the customer said, Well, I wanna do this with my data or I wanna apply it to this, you start to say, Well, wait a s okay. Yeah, that's where that's where a lot of the real innovation happens. And you mentioned something um almost like now there may be this opportunity for a uh an inventive layer of not just creating, you know, uh large scale, very efficient models, but maybe this

you know, models as a database itself, model model management layer within the system that didn't exist before 'cause nobody was nobody's creating models the way that you're creating them. I mean, that to me, it all of a sudden you're saying, okay, wait, that that could shift the entire paradigm.

Alexandra Pasi (51:07)
Yeah, that's exactly right. I mean, I think that, you know, a lot of these larger companies do internal direction to different, you know, large language models. But like, what if that wasn't just something they were doing as a black box and you have no control or insight over token cost and it's also not using your data? What if you had perfect control over this data goes into this model, this data goes into this model, this and it's all mine and it's all auditable and it's all, you know, repeatable and replicable?

And I store it in a database, and then that's what the agent is pulling from. And I know exactly what it pulled when, and I know what data that came from. There's just like a higher degree of control, but also of sovereignty over your own business processes and decision making that is not accomplishable if you're just, you know, outsourcing all of that decision making, you'll also get poorer performance.

Jeff Holman (51:57)
Yeah, by the I'm gonna get I'm gonna it's gonna cost me twenty percent of what it used to cost me. Not not only do I have all the exactly all all the control, but I've uh my costs are better.

Alexandra Pasi (52:05)
That that's exactly right. And and so you might ask like, Well, what is the cost, right? 'Cause it's like, Okay, it's better, it's cheaper, it's faster, like great, great.

Catch and and there is a catch. Like there's definitely a catch. And the catch is clarity around your differentiation. Clarity around your differentiation, clarity around your customer, clarity around your product, clarity around how to bridge that. But I I think what's what's nice and what we kind of love is that I think that's where humans are at their best, right? Understanding that kind of context. Yeah. And so from my perspective, we're we're putting the people where their time is the most valuable.

and I know a lot of people say that, right? But I I really think that this is this is true. Um, the machine that we've created is just the engine that synthesizes the patterns that come into it as best as possible, both from an accuracy and from a lightweightness perspective. But it says nothing about what you do with those outputs and predictions. And it says nothing about like the world of inputs that you need to.

you know, kind of gather to put in. Like you can use the engine as part of that discovery process. But you need, for instance, a clinician to say, um, you know what I think increases a patient's risk of delirium is like this, uh, these these different inputs, right? I don't know exactly how those inputs interact together to produce this outcome. But I know that these are relevant. So I'm going to put that in and I'm going to let the machine find these statistical patterns. And then, you know, if I want more transparency,

I have some tools to do that with. But and then and then, you know, so it predicts the patient's gonna be delirious and it's a good model. Like what do I do with that as a clinician? Those decisions remain with the humans as I think they should.

Jeff Holman (53:52)
Yeah. Sorry, I may be a bit distracted here because I'm thinking I I have an extended family member who went through a an episode of delirium earlier this year and it was pretty traumatic and that would have been highly beneficial to uh to everyone to uh have a better prediction for that. anyway, total total total tangent there.

Alexandra Pasi (54:12)
Huge

impact though. And it it's it, you know, it's often these little things that um do have actually solutions at our fingertips. There's a couple different things that keep us from tackling them. You know, many are related to kind of the financial incentive structure. Luckily in cases like pressure injury injury and delirium and stuff, like there's there's, you know, insurance alignment and stuff like that with with the incentives. But I I think the other piece is just we don't quite have the right

architectural patterns in the zeitgeist yet to know how, you know, we spent decades collecting all of the data. Like everybody had their own little perspective on this is the data I need to do this. Yeah. And everybody went out and they got it and they created their little kingdoms. And then what happened? Well, all of that data has been concentrated into the large models. And now the solution that companies are being pitched is, you know, you spent, you spent 10, 20, however many years collecting

this data and this perspective. And then it all, you know, through one means or another, kind of got souped up into the big, you know, these, these big large neural networks and these foundation models. Yeah. And now your decision is is just how many tokens of that to buy to run your business. and by the way, it's been mixed in with everything else. Yeah. Um, I think a lot of companies don't like that answer. I think a lot of companies are are really trying to preserve

their differentiation, their edge, their sovereignty, keep a lid on their costs. I mean, everyone's worried about what the end if token subsidization looks like. People are worried about how they get access to GPUs. And so I think that that's one of the places where, you know, we were a little early to market, but now all of the things that we were talking about early on are like the hot button issues. Um and I think it's it's time and it's it's right for people to start developing architectural patterns that put that

back in their hands.

Jeff Holman (56:10)
Wh how d how does that play in to to lucidity sciences? Uh like are you guys are sounds like you're positioned well to to continue doing what you're doing, but maybe layer in these other these other column services or or or where does this take you?

Alexandra Pasi (56:28)
Yeah. So I I think, you know, for for better or for worse, we're very interested in allowing our customers to take their data and serve their customer base. Sure. Right. So I I think largely the people that we like to work with are the companies that are trying to preserve their particular edge in a vertical. Now we'll we'll often work with platform partners that they're already on, right? In that kind of, you know, broader infrastructural horizontal slice. But

Um, a lot of the people that we work with that we talk to are trying to solve really difficult problems really well. Like they have a high bar of excellence in their particular markets because they see that bar of excellence in security or in performance or in cost as being the differentiator that they'll sink or swim on. Um, and so you know, we're in the business of enabling that because we think that's that's the only way you can carry forward.

As I mean, certainly as a vertical software company, but certainly for for companies in many other kind of tech enabled settings. And, you know, our our mission with LumaWarp ML, which is this ML engine that sits at this layer infrastructurally, is to let companies do that in a way where their decisions were either stay in the past and do analytics or just outsource everything to these foundation models and and like, you know, kiss your cost structure goodbye.

Jeff Holman (57:53)
But now Luma Luma Warp is the solution to build your own your own model and model management system.

Alexandra Pasi (58:00)
Yeah, model layer, exactly from from your just on top of your data. And you can still keep all of the agents, but guess what? You can lower your token costs on those since you're not asking them to like statistically reason through terabytes and terabytes of your data.

Jeff Holman (58:15)
Seems like a fantastic solution and and maybe next stage it'll be exciting to see where that goes. Um if you had a if you had a you know, a magic button you could push and be like, Hey, the the one thing I would like to do with my business now what would that be?

Alexandra Pasi (58:28)
Yeah, so now we're talking a little bit more about future, right? Which it really has to do with, okay, what is this technology? This technology is a way to find patterns as well as they ever have been found, but at orders of magnitude smaller size and therefore orders of magnitude smaller cost. And so, you know, we'll be expanding in areas that include

vision edge robotics where these, you know, that lightweightness and that that inference is really important. But I think that one of the things I'm I'm very eager to see is, you know, how does this help address some of our very real constraints around compute? do we really need to build out all of this additional infrastructure at this level? Or do we just need to find better mathematical ways?

of expressing and leveraging the patterns in the data that we're studying. Because neural nets very you know, these these neural nets that are running these foundation models, they're very, very, very large because neural nets don't compress super well natively. You know, if you're trying to capture a high degree of nonlinearity, you need a lot of nodes. And and so there's basically this this mathematical limit to how high fidelity um these models can maintain at smaller size.

And if you come up with a new mathematical way of representing it, you break that bottleneck, which means the solution to the compute bottleneck could actually be a mathematical one. Yeah. Implemented on hardware, but that would solve a lot of problems for the ecosystem on a whole different level.

Jeff Holman (1:00:07)
that's there's a there's an analogous path here in in alternative energy, right? W where people have been trying forever to to create m micro scaled uh alternative energy solutions because we you know, you have the solar farms and the wind farms and it it's been really hard to crack that nut for for people to bring that into the, you know, on site, home generation type of model, but, you know

With data, math seems like that's if if there's an option, math should provide that solution.

Alexandra Pasi (1:00:41)
Exactly. I mean maybe the answer to the hard tech problem is actually just hard math.

Jeff Holman (1:00:47)
man, we're all in trouble. No, we're not, because we've got people like you and and and lucidity sciences to help us out. Lexi, it's been fantastic uh having you on here on the on the show. Um and you've you've spared me from from having to show all my mathematical ignorance here, but so I appreciate you for that.

Alexandra Pasi (1:01:08)
I really appreciate where you took that discussion. I I think it was it was really fun, went a lot of really interesting places.

Jeff Holman (1:01:13)
Well, it's a f fascinating things that you're doing. Um, if somebody wanted to to get in touch with you um and hear more about what you're doing, or maybe, you know, maybe chat with you about uh hiring your team to to help create a solution for them, how would they do that?

Alexandra Pasi (1:01:27)
Yeah, so you can go to lucidity sciences.com. Um if you're ready to try out LumaWarp, let us know. We we do have a credits program for qualifying companies, so we can we can get you in there and hands-on. Um and then uh I'm also on LinkedIn, so feel free to follow me, Alexandra Posse. Um I'd love to hear from anyone uh that listens to this podcast.

Jeff Holman (1:01:48)
Awesome. Well, thank you so much for joining and sharing your insights today.

Alexandra Pasi (1:01:52)
Yeah, thank you so much again for having me. It's been a great time.

Jeff Holman (1:01:55)
Great. And thanks again to those of you in our audience who joined us on another episode of the Breakout CEO. Be sure to follow or subscribe on your favorite podcast platform. And if you enjoy the show, a rating or a review goes a long way. Our mission is to promote the stories of breakout CEOs in scaling SaaS, e-commerce, and tech companies to equip peer CEOs with valuable perspectives and confidence.

Thanks again for joining us on this episode of the Breakout CEO. I'm Jeff Holman, and I'll see you next time.

‍

STRATEGIC MOVES OF SCALING CEOs, IN YOUR INBOX

Subscribe to The Breakout CEO podcast newsletter.
Intellectual Strategies © 2025