Inna Tokarev Sela on Approaching Data Challenges with Generative AI

Welcome back to Data Driven, the podcast that peeks into the

rapidly evolving worlds of data science, artificial intelligence,

and the underlying magic of data engineering. Today's guest

is someone who's redefining the rules of the game in AI and data,

Ina Tokarev Saale. She's the CEO and founder of

Illumix, a company pioneering the use of generative semantic

fabric to make organizations AI ready. We'll dig into how

Ina's background as a frustrated data user sparked her innovative

journey, why 80% of enterprise decisions still aren't data

driven, and her bold vision for a future with app free workspaces

where AI copilots handle the heavy lifting. Oh, and we're

tackling the ultimate question. If the future is already here,

why does it still feel so delightfully chaotic? Sit

back, grab your favorite coffee mug, or a Maryland state flag

one if you're feeling fancy, and let's dive in.

Alright. Hello, and welcome back to Data Driven, the podcast where we explore the emergent

fields of data science, artificial intelligence, and, of course, it's all made

possible by data engineering. And with me today is my most favoritest

data engineer in the world, Andy Leonard. How's it going, Andy? It's going well,

Frank. It always warms my heart when you introduce me like that. Well, you are

my most favorite data engineer. Well, that's cool. You're well, you're my

most favoritest. I like, there's so many things. Right? Data

scientist, developer, evangelist.

I mean, there's all sorts of cool things that you do. Super,

certified person. What are you up to in certifies in certification?

12. Wow. Yeah. I'm in I'm in the

New York City area code now. So that's good. Next

up, the Bronx area code 718. So Wow. That's a

big jump. Yeah. Yeah. We're we're working on we're working on it, and I'm at

760 some odd consecutive days. I'm at the point now

where when I post anything on about Pluralsight

or, my number, the search or the number of

days, Pluralsight always sends me a congratulations, Frank. Keep

going. So, like, I'm on their radar now. So which is really

nice. I don't know. It's super cool. Yeah. It is super cool, which reminds me

I still have to do 2 days. But in the

virtual green room, we were talking about coffee mugs. We

were. And, we're we're I don't have a coffee mug with

me today, but, there's an

interesting anecdote from a previous show, which I think the show is live now, about

the Maryland state flag coffee mug, which is, pretty funny.

So today we have with us a very special guest,

Ina Tokarav Sala. She's the CEO and founder

of Illumix, and a pioneer

of generative semantic fabric, which I wanna know more about that, but it

empowers organizations with AI readiness throughout her career

leading data products, monetization, and as a data

stakeholder. Ina recognized the oxymoron of our

domain. Despite huge investments in data and analytics,

most business decisions are still not based on these data or

insights. And when I read that, I felt that one.

So she, she works she founded this company,

Lumix, which is, the the byline says, get your organization

data generative AI ready. So So welcome to the show, Ina.

And, tell us about this. Like, because I think this is a big problem

with generative AI. Well, first off, let's tackle the big

one, which is the idea that despite all this money that's been

thrown at data and analytics for at least 2 decades, probably

longer, a lot of decisions are not data driven.

Yeah. Fine. Can you hear me? Because

I see a little bit Yeah. We can hear you. Okay.

So yeah. Thank you. You're totally right. The the benchmark says

only 20% of decision making in enterprise is based on data.

And to me, I I have been around for a

while. So 25 years in data analytics, and it was

always about cloud, big data. But

what it actually boils down to? Are you able to

pull out whatever analysis of data you need when you have, like, question on

hand? Not really. And this is a situation in majority

of enterprises, right? Even if those huge data

teams and huge investments in infrastructure and all of that.

And to me, the biggest promise

of of LLMs in enterprise setting is to

to bring the contextual and relevant data

to the stakeholders in need.

Right? In this experience which is impromptu which

means it's improvised, it's governed and hallucination free, it's

transparent. So I I would totally love have to

have this experience where I'm in my Slack or Teams, right, and

I've been able to to chat with my data copilot

and ask a question and get the answer I can base decision happen.

Right? Not just an answer. I should be reverse engineering

with, you know, bunch of people.

Interesting. Interesting. But I don't think that I think that

the companies, they

they they they throw a lot of data. They store a lot of data. They

analyze a lot of data. But a lot of at the end of the day,

not all decisions, but a lot of decisions are not based on just the direct

decision of the data. They're based on quite frankly a lot

of it's particularly the higher the, higher the

level. Sometimes it's based on what's good for the person, not

necessarily the organization or the business, let alone the customer.

Do you think what are your thoughts on that? I'm familiar with the saying,

if you touch your data long enough, it will confess. That's

right. It goes exactly to the domain.

So I guess you can you can massage the results

right? But, secondhandly, when an

employee comes to me with suggestion with a business plan with,

you know some project I always ask like what's the ROI like what's

it going to be to spend and what's the impact on on you know

other activities and and what it's going to be on expense of

so having numbers having data to you

know to the basic decision or to bring to your boss is

always has been a struggle and it's still struggle today so I

think it overweights maybe some you know,

reluctance to have open data for all just for the

sake of of being able to to have specific context on it.

Interesting. That that is very interesting. And, you know, that I

think that's been the the purpose of a lot of

data driven activities in in corporations globally

is, you know, and for a very long time is how do you convert

data in its raw natural form into

information? Mhmm. And, you know, and and

defining information as, something I

can glance at and know, you know,

almost instantly how my enterprise is performing.

And that was kind of my opening line 20 years ago when I

started in data warehousing is to go talk

to a decision maker, CIO, CEO,

and, you know, try and do a very small, project,

a phase 0. And just ask them that, how do you know?

And the surprising answer, yeah, even then it was surprising,

was, you know, something along the lines of, well,

people email, information to

a lady out front or a secretary assistant guy out front,

and he or she compiles it and puts it into this summary,

and then they tell me. And so, you know, 1 PM

every day or, you know, Monday on 1 PM. I know how we

did last week. Something like that. It's very

manual processes. So does

does Illumix, address that? The

manual part? Yeah. Yeah. Totally. So

I don't think reports will go anywhere, but I think we'll

have, you know, at least 3 types of

experience with data. So I do I do believe in

application free future where you have a

question or a task and then you have a launcher and you

just, you know, articulate whatever request you have.

And in the background whatever applications, workloads, and data have

been engaged with each other to to basically come up with the

results. Right? So I do believe in this future. Right? So this is

the ultimate. Right? But I think we will have this intermediate

stage where we'll have a lot of copilots or

assisted insights in, in the context of

applications you're already using. So using your CRM systems, you will have

all kind of insights, suggestions, you know, data driven,

actions which which might come up with the system in your

workflow inside your context. Right? And you might have to have

this pure experience when you do go to analytic systems like BI

or something else where you do have your static dashboards,

day after day, same way that I go to, you know, to to my

CRM dashboards and see how pipeline is going and all of that. So I do

not them need to them to change. Right? I don't want to go to some

chatbot and and ask again and again the same question, like, what's the pipeline

conversion today? Right? I do want to have those static dashboards where I just,

you know, sneak peek and see if everything in line and

we we in the benchmark. So those three types of experiences, I

do not think they're going to to evaporate in

the future. Right now, we are mostly bound to the last type of

experience of being in the closed garden of our BI tools,

like this 3 modeled analytic experience and then we'll have this

phase where we do have embedded experience. Majority of the companies are

already suggesting some kind of improvements in the

space, some better, some halfway, let's

say. And and the ultimate goal is to to have this

launcher when for for majority of ad hoc

task of questions, you will have this improvised experience.

So a follow-up on that. You mentioned Copilot, and,

Microsoft has been the company that I've heard using that term most

often for some sort of digital assistance. It

to me, outsider looking in, although I I use the

tools, it it seems to have been a quantum leap,

this year in that technology. It just seems like last year, they were

talking about things that it might help with, and I've seen

all sorts of examples of this. But have you seen that? Has that been

your experience that in the last 12 months, these type of

assistants have just, you know, taken a giant step forward?

Mhmm. I will address this question together with the previous one, like, how

Illumax is is positioned in in this context. So I

do see many projects in the companies

which, and mainly, they're providing

copilots, for call centers or support centers

and mainly based on document summarization.

Right? So document summary is more,

lightweight and and risk averse use

of LLM technology where I can actually go and check the document

itself based on the resource. Right? So it's kind of and documents

are already articulated with lots of context in

business language. So it's kind of low hanging fruit and majority

of the companies go to the direction including, Microsoft.

Where Elamax goes Elamax actually,

tackles the market which is less,

less digested, the market of structured data. So you mentioned you

started your career in warehouse and, so warehouses,

databases, data lakes, business applications such as supply

chain, ARP, CRM, and all of that. All of that

con defined as structured data space. And despite the

name, it couldn't be less structured than it is at the

moment. Right? So you have If it is structured, it's not structured

the way you need it. Yeah. Exactly. So the nay namings are not meaningful, like

abbreviations, frank table, or for like abbreviations,

the, frank table or and this

transformation or alias. Right? So all those weird names especially under

SAP systems. I love that and and no

single source of truth. Right? In documents, you might have versions, but you do

still have some alignment to single source of truth. In data, you

can have many definitions even in the same

data source. And the thing is, if you put semantic

models like semantic search on top of them and it works by proximity,

you might have hallucinations and random answers every time you engage

with the tool. So this this is where we chose with

Illumix to to tackle the problem as,

basically, defining as a 3 step approach.

Right? The first step is getting data AI

ready. So there is no yeah. There is

no way of using generative I or AI analytics in general

if you do not have other data. But for analytics, which is

served to you as BI dashboard, it's actually feasible to do

manual data massaging. Right? Well, fun. Yeah.

Yeah. That's fun. That's near and dear to my heart as a as a data

engineer, data quality. Because

you can have the, you know, the fastest, best presentation, the

slickest graphics, and it could be totally lying to

you. And back, you know, even from the days of of

data warehousing all the way through today's semantic models and

dashboards, it's a the the quality

of the data store you're reporting against,

That that data quality, if you were to measure it, you know, there's a number

of ways to do it. But it's well north of

99% of that. And people see that, and they go, wow.

That that's super good. And it's like, no. No. It didn't. You can't do

predictive analytics off of something that's 99%

because that that 1% of bad data or

incorrect data or duplicate data will skew your results.

And what often, you know, the the layperson doesn't understand

is that if it lies to you and tells you you're gonna make a $1,000,000,000,

that's just as bad as it telling you you're only gonna make a

$1,000,000 if the if the truth is you're gonna you're at about 25,000,000.

That's your real projection if you were to follow that line out and do the

extrapolation, you know, properly. And you can make

bad decisions with an overestimation just as easily,

maybe more so than if it's an underestimation. Yeah.

Exactly. So this goes to to, to the ground truth of

your results as good as your data is. And you cannot

trust, simple semantic search

to solve all these problems for you. And

so for us, the baseline, the first use

case is to get data AI ready or generative AI ready And we

do use generative AI for that from day 1. We actually generated company

to explain to our investors back then what it actually means.

Yeah. You know, I I get it. I mean, if you build on a crooked

foundation, you you can't get anything straight, you know,

out of that. So that makes perfect sense to me. And it and,

please correct me if I'm mischaracterizing, the work that Illumix

does. But is it automated,

AI automated, data quality? Is that really what you're

after? So, basically, we automated full

stack of LLM deployment for structured data, and it takes the

AI readiness part. AI readiness, which means we have automated

reconciliation, labeling, sensitivity tagging Okay.

Like lots of lots of data preparation which is automated.

Gartner actually named us as a call vendor for that lately. We have

this layer of a context automation. Right? So so any

LLM, any semantic model needs context and this context and reasoning

usually rebuild by data scientists. To me, it's controversial

because, you know we had data modelers which didn't

understand business logic and now we have data scientists who do not necessarily

fully understand business logic and the model into black

box experience of context. Right? So ElamX

reverses process. We actually automate context and we wrap it

up in augmented governance workflow so business people or

governance folks can actually certify it. So it's auto generated

context for LLMs but certifiable by humans. We do

believe that we need to bring human to the loop, right, to to certify

it. Yeah. And the last I love I'm sorry. I have

interrupted you, like, 3 times now, and I apologize. I haven't met 2. I

thought you paused. So finish please finish your thought.

No. No. I'm saying, like, 3 parts. So you already did data governance and the

actual alarm deployment because you need to interact with the whole thing, and the interaction

to have to has to be explainable and transparent. You need to understand

how, especially on structured data, you need to understand how

the question was calculated based, sorry, how answer was

calculated based on questions and how, data was

actually sourced, what's the lineage, what is the governance and access

control through search your clients. So all of that should be on the interaction layer.

So AI readiness, governance, and the interaction layer explainability to

the end user. Absolutely. Okay.

Thanks. And I do apologize again for the

interruption. So my my characterization of it as something that's just

data quality is is way low. There's a little bit of overlap between

data quality and what you're describing. You're talking about taking this into

that next level that is specific to, generative

AI and perhaps other, you know, AI related,

AI adjacent technologies, machine learning leaps to mind and stuff like

that. But your the tagging, the categorizing,

and all of the things you're describing there, that is next level.

And it's very interesting to me that you're

using AI to get data ready for AI.

That's an interesting combination. Mhmm. It makes sense, though. Right?

You can kinda scale out human capability with AI. I

think that's you you kind of alluded that with Newman in the loop. Right? Like,

I think I think where you were kinda going with that, again, don't wanna speak

for you, but it's like the idea that AI isn't gonna replace

humans. It's just gonna make humans more productive. Yeah.

For sure. Augment us because frankly speaking, no one

wants to to model data, you know, as their

career. We want to solve problems. Right? And to solve

problems, we we have to to understand what the problems are

And letting AI to surface the problems as alerts and for us

to to resolve them as conflicts takes, you

know, 1% to 10% of the time that it should take,

where we are busy, you know, wrangling data still. And, you know,

it's sad to some extent because data is growing and we cannot keep up.

No. That's a good point. Even if even if there are people out there and

some of our listeners may really do like modeling data. Right? But, you

know, Dow, they can model about 10 times the amount of data or maybe

a 100 times more. Right? And then ultimately, the expectation of

what a, you know, what a person

can do in a set period of time is gonna go up just

because I I I think I think you're on to something there. Plus,

I also I would also, like, double click on the idea that you said earlier,

which I think was very intriguing, was this notion of

a lot of the apps that you use would kind of fade away. You just

have this virtual assistant. You know, I I think back to

there's a number of scenes in, you know, Star Trek The Next Generation where they

have a conversation with the computer. Right? Mhmm. You know, you they

don't they use the computer. They get stuff done. There's no

Microsoft Word. There's no PowerPoint. Right? Like, there's no, like, it's

just the the there is no application. The application is kind of invisible. It

becomes the computer. And I think that's a very

intriguing kind of way. And if you had told me that a year ago, I

would have been very skeptical. Now I look at it, I'm like, I

mean, it's it's it's almost inevitable.

Yeah. Yeah. I agree with you. Futures here,

it's not evenly distributed as people say. So I

guess, you know, when you're attending conferences in Bay Area,

it's already it's already here. It happens. Right

and when you go to let's say Europe we

even just say you know just say a EU act in

Europe is is ramping up so it's all about

controls and and this is great So I do not think that regulation and

innovation, actually, jeopardize each other. I think

they should go hand by hand and, that's where I see

industry is going. So so East Coast approach, majority of our customers

are coming from East Coast US, Pharma,

financial services, insurance, highly regulated data

intensive companies. They have now,

sometimes even inventing standards for generative AI

implementations because everything is so new but companies

want to go fast. Right? So no one wants

to to downplay risks on one hand. On the other

hand, everyone want to, you know, to implement generative AI

and see the productivity cuts. It's, you know, it's evident productivity

cuts are already here with all those co pilots summarization,

what have you and this is where we are today. So I

think like again Bay Area running fast

and east is coming up with regulation. We will meet somewhere

in between. I believe in both. Well, if you kind of,

like, look at, like, historically, you know, when .coms first

started, right, there were a number of, hey. Look. You know, we're gonna sell pet

food online. Right? Like, and then it was

like, back in the dial up days, it didn't really make a lot of

sense. So it would just be easier for me to go to the store.

Whereas now, I mean, if you think about ecommerce, obviously,

Amazon is the £2,000,000,000 gorilla in the

room. I like, do I really

wanna think about, you know, dealing particularly as we get into the holiday season, do

I really wanna deal with the traffic at the mall or the store when I

can just click on something, either have, you know, groceries delivered

or, you know, I'm I'm okay waiting 2 days for

something to come up if I don't have to deal with them all.

Yeah. Totally. What's what's the difference between Black Friday

and Cyber Monday? No. It's not. Right? Like not really. Yeah.

Yeah. So it's like Not anymore. I remember Yeah. You

know? So we're recording this just before Black Friday. And,

you know, this whole idea of, you know, going to the store, get

the best deals, it's like, do I really wanna deal with the

crowd? No. Yeah. Although ironically, the name for the

podcast came on a Black Friday, while I was

at a Dunkin' Donuts, drinking coffee, waiting waiting

in line actually to get so there's a I'm a Krispy Kreme

person. So I'm Ah, okay. Yeah. So With you and

I, right, definitely. Right here. This is before we had a Krispy Kreme

near us. So it's I I have split sides, but yeah. Yeah.

Jeff's JT. He's a mess. From up north. So they are

they're Dunkin' Donuts. I've noticed this. They're Dunkin' Donuts, like, north of

Virginia. And he's in Maryland. I'm in Virginia. Then down

south, you rarely see a Dunkin' Donuts. I see more Dunkin' Donuts down

south than Krispy Kreme's up north, though, for sure. Yeah. But

I They're they're from Boston. That's why. Yeah. Oh, that's why. And then So at

Krispy Kreme's from Atlanta. And plus, it's funny. Right? Like, so I live in

Maryland Mhmm. Which depending on who whom you ask is either

north or south. So that's right. That's true.

Interesting. Interesting. We're a quarter state for sure. Yeah. That that's

that goes safe for Virginia. But I wanted to follow-up on, you know, you've

been we've been talking about all the cool stuff. I'm

gonna try and say this correctly. Illumix. Is that correct? Am I getting it

right? So Illumix name

from Illuminating the Dark Side of Organizational Data.

Illuminate like illuminate. Illuminate. I like that. And x x

for the x factor. Excellent. X for the x

factor. Yeah. What? And I'm not asking you to I'll

just ask a question. What are the risks in in what you're doing?

And, you know, what are the risks you're aware of and how are you addressing

those? Yeah.

is going to be, a TCO, total cost of

ownership. So already today,

it's, it's very hard for organizations to to

monitor where the generative AI tokens are spent.

And the benchmark say that 80%

of LLM tokens actually spend on customization

of off the shelf models. And that's not a good news because

which means ROI is is pretty low on on the actual

production use of generative AI in in enterprise.

And I think it doesn't get any better because the

customizations techniques which are used today gains a black box

performed by super expensive data scientists and

they're not very scalable for data that you don't want to, you know,

to schmooze around. I think it's cost prohibitive actually to bring data

to AI. You need to bring AI to data. So so putting

data in some graph structures for graph, frog, and all of that, it's to me,

it's cost prohibitive. So this is why I think that, the Telumex

transparency. We do create this, a virtual,

a semantic knowledge graph, which is transparent to certify, which is

created for business people. Based on business

logic. We do use extensively industry ontologies and so on so forth.

And I think the the most interesting part about generative AI is

we do not necessarily going to mimic processes that

the humans performed. Mhmm. We're going to invent

those processes. Right? So new new processes and new workflows. So

right now, a generative AI is deployed like like

analytics is deployed, which means you you have to

label your data, check the quality, usually manually, and then

you have to to prepare the test set which is fed

into customization of the model and then you actually provide the

context to on every question. So this is

very old fashioned or, you know, 40 years old

machine learning technique to to actually train generative

vi. So this is why why I'm saying that, many companies are

probably going to to mimic what Equinox does in the sense

that you have to you have to be focused on domain

specific knowledge, reason, ontologies, and knowledge graphs. You have

to onboard your customers automatically via metadata because

metadata has the factor all

activities in organization documented for us. We're

just under utilizing them, right? And then you bring your

business people, your domain experts, your governance teams to the

loop because you can simply cannot bring this business acumen,

to, you know, to data. You have to bring data to to those people.

That's an interesting thing because I've seen the the particularly is this this this

statistic around 80% of the tokens are being used to

manipulate the data. I have a microcosm example of that

where I use AI to augment my blog post, my blog

that I create, and I finally took

a closer look at this because I was spending a lot more on

the OpenAI API than I really wanted to. And I'm like, well,

what exactly am I I'm using a product called Fabric.

And I'm like, wait, what exactly is the source of this prompt? And I look

at it, and I'm like, I can't. It's a lot. It's a long prompt. And

I'm like, I really don't need that. Right? So we are gonna do a deep

dive in a show on Fabric at some point. Not not the Fabric Andy

works with, but there's an open source thing called fabric. There's

a I'm sure there are lawyers right now that are doing their

holiday shopping based on how much money they're gonna make off of this

dispute. But, the the short of it is, like,

I realized, like, well, no wonder why I spent so much money. I'm sending all

of this in my prompt plus the content. So I

actually in the verse before you joined in, Andy and I were talking, and I

was like, I actually got a really good result based on a more optimized

prompt. You know? And, you know, strictly speaking, it's

not I I like your approach of bringing the AI to the data rather than

bringing the data to the AI because that is expensive.

You know, I I think that bringing the AI to the data will be less

expensive. How less, I think, remains to be seen. But I like that approach,

right? Because that's typically what we've done, you know, and we've seen

huge upsides to that, whether it's from Hadoop bringing the

compute to the data rather than vice versa. I like that

approach. And it's backed by historical precedent. Right? So it's not

completely gonna be this crazy idea. It's just a very sensible

idea. Yeah. Yeah. I believe the future was already

invented. Right? So it's just the inclination of technologies we already have.

It's been healthy about it. So, we had

machine learning practices which are very healthy like feature

exploration, feature definitions and then we had neural net brute

force and then majority of companies used combination of both,

right, to to to be optimized. This is what I think what's happening with

generative AI. So this, you know, wild west of brute

force or great spend is going to be replaced by methods

which have, like, this automated context filtering or pre

processing and then use like fraction of your budget to to actually

run the query. Yeah. I remember hearing about a lot

of this in the late nineties. And, I worked for a company who

was a big SAP shop. I see you have a history with SAP. Yeah. And

this lady and and and so we were an we were the IT department. So

we were in the basement, but the analytics team back then was in

a closed in space inside the basement. So it was

like even more like, you know, I was the web developer, so I didn't

have a window, but I could see the window about 50 feet away.

But, like, when you when when you went

into this, like, you know, further enclosed space deeper into

the the the the the depths of the IT department,

there was the database team. And and and and in the back of that area

was the analytics group. And I remember this lady telling me

that she was working with these things called OLAP cubes. Oh, wow.

Yeah. And I was like, what is that? And then she went on this thing

and, you know, I'm remembering a conversation, oh my god,

almost 30 years ago. But I just remember walking away with,

like, that sounds either crazy because she's talking about,

like, you know, figuring out patterns. Right? So, you know, will

rainfall patterns in Australia affect not just the agricultural

side of the chemical business, but also the plastics purchasing

versus rainfall in the Amazon versus this and all of

that? And I just remember walking away from that conversation as I as I

as I as I leave the depths of the IT department back to my normal

kinda, basement. Back to the regular basement from

the sub basement. I remember thinking that is either the craziest thing I

ever heard or the most profound thing I ever heard, which

now with the, hindsight of time, it turns out it was the most profound

thing. Yeah. You you can think about it as

semantic layers of, you know, that era. Right?

Mhmm. Right. And I think You know go ahead.

I'm sorry. Sorry. I think it's delayed between the

between the connection. So I think around the same time I was

doing my bachelor and my project was about multi dimensional

theory. So multi dimensional geometry,

of these neural nets. So basically, you model neural nets as multi

dimensional graph and it does operational research calculations.

So it's exactly the same. You you model your universe in a

graph. Back then it wasn't MATLAB. We didn't have any, you

know, neural nets Right. Structures or graph structures and so you're

modeling in MATLAB in this weird language,

a graph which has a neural nets on there. And

this is exactly like modeling all of cubes. Right? A

multidimensional representation of your reality. Now,

unfortunately, we have a new technologies which,

which are semantic and context. Right? Large language

models and graphs, which do the same thing but much

more efficiently. Yeah. So this is amazing. Like, I

think it goes back to what you said. You know, The future's already here. It's

just not widely distributed yet, which I think is a William Gibson

quote, or is it a Esther Dyson quote? I forgot.

But it's one of those 2 kinda luminaries. Yep.

You you said what I was going to say, you know, and it

was, you know, more of what off of what Frank

said is it turns out that we're just

doing more nodal analysis and vector

geometry as a result of that. That's it did all start

with multidimensional and and grow from there. And

that's where these algorithms, like nearest neighbor

originated, was in that math. So

Yeah. Yeah. Great minds. Exactly. Exactly.

Alike. Exactly. Now you're

complimenting me. Thank you. I I feel I feel better

when smart people in the room agree with me.

No. I'm on the right path. You know, I employ

millennials. So so having people with experience in multidimensional

geometry and all of cubes, it's just a miracle to me to to start

with. You know? People now like Python, neural

nets, we do actually, the average age in in in

Lumex is around 35, 37, something like that. So we do

have like also pretty experienced folks, you know, but new talent,

they, they they're not familiar with all all of that.

And I think it's actually a disadvantage because,

when when you do know different patterns in architecture Yeah.

You can model them with new technology. Right? Make them more

efficient, but you already know what works and what doesn't, and it

helps. That yeah. That's a great point. The old

experience, you know, the experience that we have from doing this for

decades is that we see the patterns that have

repeated over time, architectural patterns and design patterns. And,

you know, and we know that they've

I I love that how you said that. The, you know, the future's already been

invented. We we realize that if we reapply some of these

patterns, that there are use cases for them, not just now, but

also in the future. So totally get you.

Too, you know, like,

you know, it it is painful to think that, you know, we've been in this

industry for decades. Right? It is a little hurts a little bit. But,

like, also, if you're listening to this, you've not been in the industry for

decades, and you're thinking like, woah. You know, what are these what are these

old geezers now? I would point out when I was

a young kid in the industry and, you know,

client server was like the new hotness. Right?

And, you know, the whole notion of going back to,

you know, cloud and and and and and, you know, terminal

and an old mainframe geezer basically said to me, like, this is just

this industry has a cycles. Right? It's like the fashion industry. This goes in

style. This goes out style. And it was like, I had that moment

of, like, wait. I think he's on to something, but he's just an old geezer,

so I won't listen. So, you know, so so

if you are a young buck, like, or,

buck is a male deer, right? What would be a Yes. A doe. A young

doe. So if you're a young buck or a young doe, I grew up

in New York City. So all of this wildlife thing is brand new. I'm here

for you. I'm here for you, Frank. So, you

know, listen to, like, some of the things that these, you know, more

experienced colleagues will say. Yeah. You know,

if you don't believe it right away, just put it on the shelf in your

mind because you're gonna need it later. It'll come up at some point.

And it's like, if you look at kind of, you know, everybody ran to the

cloud. Right? And cloud is effectively like a

mainframe effectively. Right? The same philosophy. Right? Centralized

computing somewhere else. Right? And then your browsers become

the terminals, terminals with fancy graphics, but terminals nonetheless.

Now I think you're gonna start seeing it kind of we're about due for a

seismic shift backwards, right, as people kinda move

repatriate data and things like that. Particularly, I think driven by AI

because of the cost of some of this. You know, I had this debate,

you know, the other day. It was like, you know, if if one of these

super clusters with, you know, a 100, 8 100,

all of this, if it costs, say, $500,000,

right, I could probably do the math, and that probably means

about, you know, there's a certain break even point,

and it's probably after about 7 or 8 fine tunings or full

on trainings where it's just cheaper to have it. Just buy it.

Yeah. Yeah. Yeah. Totally on that. And also, you

know, salary skills are the most expensive part. So you

want to spend it on your business specific problems and

not generic problems you can solve with software. Right? So

it's always like that. Yeah. Yeah. So,

I do think that, basically capacity to process data

is is going to be a challenge. Right? And this is why we

see that, that majority of,

of I would even say countries not

only specific enterprises, kind of gear

up with, with GPUs, FPGAs,

whatever hardware you have. Right? So do you see it in

middle east, in emirates? They they have national generative

vi grid and they're building it for, you know, not only government companies

but also private companies. We see the same in Europe

and I would assume, you know, US based telcos

are going to to provide those data centers with GPU soon

enough, right, for, you know, for everyone to purchase as an

alternative to the public cloud. Yes. And we'll

see it. So this is for starters. And second one, the second part where

you don't need, this, you know, heavy machinery,

you might just have your variables processing

parts of whatever generated AI on your end before sending to the cloud

because you do not necessarily need to to process everything in a central

manner. We basically have pretty powerful machines on

our hands or in our hand, you know, as

glasses as well. We can see that, and it's

going to be part of the processing. So the processing is going to be distributed.

You bring AI to your data, where your data is. You do

not shift your data all the time. It's not, it's not

cheap anymore. And we'll have this, as you mentioned,

those central repositories of mass processing

and those distributed powerhouses which are

small enough to to process data on on edge.

I think you're right. I think you're gonna see a set of data being processed

in one place. I think it's gonna be everywhere. There's gonna be some

and and I think that that introduces some interesting, consequences. Right?

So my wife works in IT security, and I can immediately hear her voice in

the back of my head. Contrary to what you think, ladies, we do

listen. We just don't always pay attention. But

I can hear her like, well, if compute's happening everywhere,

gee, couldn't like that be poisoned anywhere.

Right? I think I think that's going to be the next kind of thing. Right?

It's and it's again, it's a pattern. Right? Advancement.

Bad actors take advantage for that. Problem happens. And

then then that's the new thing. Right? So it's almost like you're you're building like

a, like a like a like a layer cake. Right? Like, you know, the cake

goes down then the frosting. The cake is the innovation. The frosting is

security, and then so on and so on. So Yeah. Yeah. Yeah.

So it basically back to the semantics. What we started is

semantic ontology as a baseline for generative AI.

It has multiple benefits. Single source of truth, of course, has the

benefits for accuracy. But also, if you're passing every

question to this semantic ontology context,

it's almost impossible to poison it because we're going to either

match to part of your logic or Right. Right. We're going to

miss. So it's it's another layer of security if you think about

it. So, so yeah.

That's an interesting point. All new. Yeah. All new ontology, all new

semantics have governance meaning, it has

accuracy meaning, it has also security meaning.

And also if you want to have single source of truth you have to to

have means to distribute it to those edge devices or

to to bring it back to central location and without ontologies, without

semantic layers, simply it's impossible to do that. I was gonna

say, like, the the the infrastructure, not just the computer infrastructure, but the

logical infrastructure to distribute this stuff,

it's probably not a trivial problem. That's the first thing that popped in my mind.

I was like, you know, like, oh, yeah. You're right about the distributed

activity on this data, but, wow, what does that

look like? What do updates look like? Like, the whole like, it's a it sounds

like a growth industry to me.

Definitely. Yeah. Yeah. I don't it's, it's

what we call, engineering problem. Right? So

creating ontology is data science or generative AI problem, but

distributing it, maintaining it, thinking it's its engineering problem.

Engineering problems tend to to have engineering solutions. Oh, Oh,

that's a good point. That's a good way to look at it. I like that.

I like that. So did you wanna do the, premade questions?

Because we haven't we've gone a few shows without them. If you're okay with those,

Ina, we can we can ask them. If not, that's fine

too. Of course. Yeah. Sure. Mhmm. So they're not they're not complicated.

They're more kinda just general questions. I pasted them in the chat.

But the first question and and you've had a a pretty

significant career with SAP and and before that. How'd you

find your way into this space? Did you find data or did

data find you? I

found my way to data by being frustrated

user. Right? So I started in engineering

and it was evident to me that

using data as engineer is not enough. You have to go to

data management. You have to fix those things because otherwise

I will I will going to be frustrated for the end of my life. Right?

So I went to data management analytics to to solve the problem

and I discovered that, as you mentioned, every experience

has a footprint. So my experience with graphs and with

operational research and multidimensional geometry and all of that is so

useful for data management. And it was actually exhilarating.

That's true. Like and I like that because, like, every experience does leave

a footprint. Like, you know, that that's cool. I'm gonna I'm gonna pull that out

as a special quote for the episode. That's a great quote. Yeah. So

our next question why we do these? Yeah. Is what's your favorite part of your

current gig? My favorite part of being a

founder is is

unlimited ability of experimentation,

right? So majority of my day actually say no

to things, not to experiment, which is which is hard, which is not fun part,

right? But, still, we can

make decisions and we can do

new stuff every day. So as a founder,

it's been very, very different than enterprise setting. And don't don't take

me wrong. Like, SAP is a huge place of growth and had

very, fulfilling career at SAP, you know, building

stuff, founding p and l's, running big organizations,

but but been able to to actually, you know,

start anything new. And, like, right now, we have this customer

and they want to to try Illumax on in

parallel on the newest, you know, newest BI

tool with semantic layer or and on the oldest

warehouse on premise at once. I'm like, okay. Challenge accepted.

Yeah. And next Wow. Yeah. And next day, you know, engineer

comes with we have this academic data set and they have these benchmarks.

Let's beat them. I'm like, yeah, let's do it. It could be cool stuff.

Right? Lovely. So, you know, you know, it's to some extent,

so we don't need to justify it, you know, business wise and but but in

majority of cases, we can. Cool.

We have a couple of complete the sentences. When I'm not working, I

enjoy blank. I used to

enjoy doing jogging and yoga when I'm not working.

Right? So right now when I'm not working which means when I'm not

traveling I just spend time with my family. Whatever

is the plan for the weekend if it's just you know Netflixing,

or cooking or hiking whatever is the plan I just

join So sometimes just, you know, plan it. But spending time with my

family has become, indulgence and I'm

very focused on that. Cool. Very cool. Our

next is I think the coolest thing in technology today

is blank. I think the coolest tech is

thing right now is not in tech. It's actually the

pull from CEOs of companies

for technology. This is something which didn't experience for decades.

So we were pushing cloud and big data and machine learning and deep learning. We

were explaining to business stakeholders why do they need that. Mhmm.

And now, so you're all coming and saying, okay, I want to have

chatbot experience for x y that, so just

build it. This is actually I think this is the coolest

part because it's kind of a removes majority of the friction that

we had to to deploy technology in the past.

Interesting. On our 3rd and final complete the sentence,

I look forward to the day when I can use technology to blank.

So many things. You know, travel has

been so frustrating lately, and, I

don't think what happened because it's like kind of technology goes

forward but airline, you know, travel technology,

hospitality technology in general, I don't feel it bridges a

gap. So I really look forward to the

future where I can just have this comment, this prompt

of plan, this conference in Dallas on

x and the system already knows all by preferences and

just done. Oh, boy. It would be it would be fantastic.

Yeah. That that the travel experience as I I've had to

travel quite a bit, like, for the past, like,

couple months, and it's just like, oh my god. Like, it never was

great, but awful is not a word I remember. But it's post

pandemic, I think it's gotten way worse. It's like there's just so many small things

that you could be done a lot better. I'm I'm a 100% with you on

that one. So true. So our our next

question is to, ask you to share something

different about yourself.

Sharing something different about myself. I think I'm a controversial

person in general. So, so some people,

so some people agree with, you know, with the degree

of, of living in the future. So I,

I, you know take myself as person who is very much in the future so

all this seed happening and I might be a little bit you know ahead because

I see the technology being developed in my mind is already there, it's already

used right? So and so where this is

where I see myself controversial because you know in majority of the

cases, then you sit over family dinner

and say, you know, we're still paying our bills

online when we have this notification. Right?

So everyday technology has

developed a lot. And when I'm speaking about this application

free future and, you know,

automated, x y zed. Sometimes or many

oftentimes on everyday level, we are still not there and

this is where people think that I'm too visionary or too

too dreamer on that. Interesting.

No. I'm with you on that one.

Growing up, I was the technical person in the family. So

Yeah. They don't they don't know what you're talking about. Right? I I I love

how the, you know, or, you know, they all they

all get confused until the printer breaks and then suddenly

But you're the smartest people in the room. That's why you're the smartest person in

the world. Alright. So where can people find out more about you and

Illumix? I love socializing on

LinkedIn. I don't know that many people think LinkedIn became a

marketing tool. I still see tons of valuable

discussions and I just absolutely love keeping in touch

on LinkedIn and and see the latest and greatest and I also share quite a

bit. So LinkedIn would be the the most

straightforward way in Atokaropsala on LinkedIn.

We do have blogs and I actually write many of

them. So if you go to illumeg.ai/blocks,

you will see lots of materials written on semantics,

on ontologies, on generative AI governance. So those

topics which are close to my heart, and we communicate quite

frequently on that. Very cool. Very cool. Very cool. So

so Audible is a sponsor. And if you

would, like to take advantage of a free month of

Audible on us, you can go to the datadrivenbook.com.

I just tested the link. That's why I was looking over here for anyone watching

the video. And it works. Sometimes it doesn't. And

we ask, our guests, do you have, do first, do you

listen to audio books? And if so, can you recommend 1? If

you don't listen to audio books, just a a good book.

I do listen to audiobooks. I also podcast, more

frequently recently. I I'm not sure this book is

already on Audible, but, if not, it's going to be

in Audible soon enough. So it's Nexus by Yuval Noah

Harari. It is audible. I have it in the library already.

Yeah. Amazing. So it speaks about the truth

in the age of generative AI. Right? Interesting.

What's the truth? What's the ground truth? And I was

actually in the lunch party in SoHo, New York, you know when Yuval

was speaking about you know how how technology

and what we see right now is not very different from what we experience

in you know middle age like when when Gothenburg

and printing was was a new thing and like what was

printed actually was you know rumors

and juicy stuff rather than scientific books and this

is where what we see right now in, you know, in chatbots and internet, on

social overall. So it's it's interesting parallels that he's

taking about what's what truth is in generative

AI age where what truth were was, like, 20 years

ago or even, like, 500 years ago. Yeah.

We're the we're the same species with the same problems and the same drama

and the same drivers. Like, it's just our tools have changed, whether

it's a printing press or, you

know, celebrity gossip or whatever or fake news

or anything like that. Plus, I also think the, you know, there's an old phrase

like who watches the watchers. Right? Like Mhmm. Who decides what's

misinformation and who decides what's true? I think. I think

because misinformation could be, you know, there there's

a image of me robbing a bank. Right? Like, you know?

Mhmm. Mhmm. I thought, Frank, I thought when the US

Marshals put you into the witness protection program, they said

we couldn't bring up you robbing a bank any any longer.

Misinformation. You gotta be careful because, like, one of the things I I wanted the

flow was so good. I didn't wanna interrupt it. But, like, one of the things

was I was experimenting with fine tuning an LLM locally.

Mhmm. And I'm basically trained it on information about my blog. My blog's

One of them hallucinated this really great origin story for my

website. It was awesome. It was awesome. I'm like, I like that

better. So basically, it said that Always. Always.

It was really good. It was basically that Frank's World started as a

show, a kids TV show in the nineties on

the BBC or channel 4. I forget. Like one

of the big British channels. And it was about a talking

trash can named Frank that would teach kids about the importance

of, recycling. That's my favorite part.

And it was and it was the best part was that it was it was

the first professional project of the guys who did Sean the sheep and Wallace and

Gromit. Yeah. And I'm like so I

I I pinged the guy I worked with. Has this ever been a show?

Because no. Not that I ever heard of. And I looked over it. I couldn't

find it. But and then what I did was as an experiment, I fed

that that whole paragraph that it came up with into

notebook l m. Mhmm. Notebook l m

took that and ran with it. There's, like, a 20

minute audio, and it is the funniest thing because it basically

talks about the early environmental movement. They said it was the Britain's

answer to, Captain Planet. Like, they made up all the

stuff. And now it's documented. So now someone is going

to pulling to pull some information. And if you have Right now it's out there.

Right. And I guess to your point earlier about Lumix, like, if you start

building a crooked foundation, right, like, that eventually as

it moves on, it's gonna so, I mean, who knows, like, couple of years

from now, like, Wikipedia may say, like, there might be a

Wikipedia article about this TV show didn't exist. We're talking about it. We're feeding

the machine. That's fascinating.

Yeah. And it was a so a little bit on the books. I have to

mention it, like, in a couple of sentences. So, in US

a legal entity actually is a citizen. It

has social number. Right. So, technically machines

can create legal entities. They can vote, they can,

you know, they can create information and this information is,

you know, created with social number, with identifiers. So it's actually real

information. It's not fake news. It's created by social number.

And so this is how you create, like, this new truth. Right?

And, and how do you control that? So it's an interesting aspect of what's,

what even is defined as ground truth.

That's true. Everybody needs to define it. I think that's gonna be the question of

the 20 That's a big deal. Mhmm. Yeah. Well,

awesome. It's been great. We wanna be respectful of your time. This has been an

awesome show. Yeah. We'll let Bailey finish the show. And

that's a wrap for today's episode of data driven. A massive

thank you to Ina Tokarev Saleh for joining us and sharing her

fascinating insights into the world of generative AI, semantic

fabrics, and the ever evolving relationship between humans,

data, and decision making. If you're as inspired as we

are, be sure to check out IllumiX and follow INA on LinkedIn for

more thought leadership in the AI space. As always, thank

you, our brilliant listeners, for tuning in. Don't forget

to subscribe, leave a review, and share this episode with your data

loving friends or that one colleague who insists they don't trust

AI. We'll convert them eventually. Until next

time, stay curious, stay caffeinated, and remember,

in a world driven by data there's no such thing as a trivial

question, just fascinating answers waiting to be found. Catch

you next time on Data Driven.