Name: Kern AI at FinovateEurope 2024 - The importance of data-centric RAG
Uploaded: 2024-03-07T15:29:31.587Z
Duration: 639.08
Description: FinovateEurope did not fail to facilitate a host of compelling conversations between industry leaders at this year's event in London, UK, chief among them being FinTech Futures' fascinating discussion with Johannes Hötter, CEO and co-founder of Kern AI.

FF.INT.008_1.mp4

FF.INT.008.vtt

ENGLISH

{{00:00:04}}
Hello everybody and
welcome to Fintech Futures.

{{00:00:08}}
We are here today at Finovate
Europe in London, UK.

{{00:00:13}}
My name is Tyler Pathe, I'm a reporter
for Fintech Futures and today I'm very

{{00:00:20}}
happy to be joined by the CEO and Co
Founder of Kern AI, Johannes Hötter.

{{00:00:26}}
Johannes, how are you?

{{00:00:28}}
I'm fine, thanks.

{{00:00:29}}
Thank you.

{{00:00:29}}
Good to be here.

{{00:00:30}}
Yeah, thank you for joining me today.

{{00:00:32}}
You've you've come from
Germany, I believe.

{{00:00:34}}
Yeah, exactly, from Bonn.

{{00:00:36}}
Yes, and it was a good flight, you said?

{{00:00:37}}
Everything perfect, yes.

{{00:00:38}}
Fantastic.

{{00:00:39}}
We're very happy to have you here
at this event here in London, and

{{00:00:43}}
to be speaking with you today.

{{00:00:47}}
You've just done a demo
this afternoon of Kern AI.

{{00:00:51}}
Can you tell us a little
bit about how that went?

{{00:00:53}}
Yeah, the demo was exciting.

{{00:00:54}}
It was a live demo, which is
always something, yeah, special.

{{00:00:57}}
Yeah.

{{00:00:59}}
Because you always want to make sure
that everything works exactly as you

{{00:01:01}}
hope, but it went perfectly fine.

{{00:01:03}}
It was amazing.

{{00:01:04}}
Fantastic.

{{00:01:04}}
It's very from what I viewed, I was very
thankful not to be on the stage because

{{00:01:08}}
it was very much like go, wasn't it?

{{00:01:12}}
But lots of really interesting
technologies and concepts

{{00:01:15}}
being presented for sure.

{{00:01:16}}
There was not one where I was like,
oh, it's boring like everybody

{{00:01:21}}
had something to say and a
reason to say it, which I loved.

{{00:01:25}}
Before we get into our conversation
about Client AI and your solution

{{00:01:29}}
and everything, why don't you give
us a little bit of color into What

{{00:01:34}}
Kern AI is, who you are, and more
about the origin story, really.

{{00:01:39}}
Yeah, of course, happy to do that.

{{00:01:42}}
We initially started Kern
as a student consultancy.

{{00:01:46}}
So I've been in the field of
AI for roughly eight years now.

{{00:01:50}}
And we started very, broadly.

{{00:01:51}}
So we did projects during
our master's degree.

{{00:01:54}}
We as my co founder and me.

{{00:01:57}}
Ranging from weather prediction up to
database chatbots and e commerce AI.

{{00:02:02}}
So very broad ranges.

{{00:02:05}}
And back then, roughly in 2020, we fell
in love with natural language processing.

{{00:02:09}}
We saw GPT 2 back then,
which was mind blowing.

{{00:02:13}}
You could tell, okay, this is quickly
becoming much and much better.

{{00:02:18}}
And so what we were doing a lot is we
were working with insurers and banks.

{{00:02:23}}
And what we could tell is the teams
that put use cases actually into

{{00:02:28}}
production were very successful.

{{00:02:30}}
They never worried about
the algorithm, right?

{{00:02:33}}
So they, when you think about
AI software, typically consists

{{00:02:36}}
of an algorithm and data.

{{00:02:38}}
They always worried more about, okay,
how can I be sure that the data that

{{00:02:42}}
I have really suits my use case?

{{00:02:44}}
And so we saw this trend and we thought,
okay, if the algorithms themselves

{{00:02:48}}
become better and better we should
focus on Giving tools to companies

{{00:02:53}}
to make the data better and better.

{{00:02:55}}
That's how Kern got started.

{{00:02:56}}
Initially a student consultancy and then
transitioning into a software company.

{{00:03:01}}
That's that's really interesting because
I agree with you wholeheartedly there

{{00:03:06}}
that these models are only as good
as the data that they're fed with.

{{00:03:11}}
Do you know what I mean?

{{00:03:12}}
Like you're only as
fast as you're slowest.

{{00:03:18}}
We're delving into that.

{{00:03:20}}
We're here really to talk more about data
centric retrieval augmented generation.

{{00:03:28}}
Very big term very big
term, very complex term.

{{00:03:31}}
What does that term mean for Kern AI?

{{00:03:35}}
Yeah, I agree.

{{00:03:36}}
It's a difficult term.

{{00:03:37}}
So retrieval augmented generation,
it sounds complex, but it's.

{{00:03:41}}
Quite simple.

{{00:03:42}}
It just means if you think of
ChatGPT, you know that depending

{{00:03:46}}
on how you formulate a question,
you get different outputs, right?

{{00:03:50}}
So in Retrieval Augmented Generation,
the idea is ChatGPT itself, and

{{00:03:55}}
with ChatGPT I just mean large
language models in general, they

{{00:03:59}}
are already extremely capable.

{{00:04:01}}
They have a very broad world knowledge
and they can answer a lot of complex

{{00:04:04}}
questions, but they need context.

{{00:04:08}}
So for instance, if I asked something
like, I'm a consultant and would like

{{00:04:13}}
to work as a lecturer on the site.

{{00:04:15}}
Is that covered?

{{00:04:17}}
The large language model can only answer
this question if it has the context of my

{{00:04:20}}
policies of my tariff side, and that's the
idea of Retrieval Augmented Generation.

{{00:04:24}}
You don't start completely from scratch
building your own large language

{{00:04:27}}
model, but you take something that
already exists, that is extremely

{{00:04:31}}
smart in a way, you just try to
understand how can I improve the input.

{{00:04:36}}
into this model.

{{00:04:37}}
That's Retrieval Augmented Generation.

{{00:04:39}}
That's really interesting in that
it expands the scope of these large

{{00:04:44}}
language models, that's, I think from
ones that I've played around with,

{{00:04:50}}
they've been limited with what they
know and what they're, able to say.

{{00:04:55}}
And this actually makes complete sense.

{{00:04:57}}
And it's, really difficult to argue
against data centric RAG because

{{00:05:03}}
it's like why would you want
something that was less capable?

{{00:05:06}}
Yeah, do you know what I mean?

{{00:05:07}}
And it's it needs to be fed with data
in order to produce this outcome.

{{00:05:14}}
What kind of challenges does this
data centric RIG approach solve?

{{00:05:20}}
Yeah, so generally the idea of data
centric REG, Equivalent Augmented

{{00:05:25}}
Generation, means To put it in
simple terms, let's say we have

{{00:05:29}}
an hour time to build a solution.

{{00:05:32}}
Data centric means you spend 55 minutes
just looking at the data, trying to

{{00:05:37}}
understand how is my data connected?

{{00:05:39}}
So I have a, do I have tariffs
that are related to other

{{00:05:41}}
tariffs tariff generations?

{{00:05:44}}
Like just how is my data connected?

{{00:05:46}}
And I just spend five
minutes on the algorithm.

{{00:05:48}}
That's like the overall methodology,
the idea of data center, Greg.

{{00:05:53}}
And the idea is.

{{00:05:55}}
That's typically the bottleneck
in the overall system.

{{00:05:57}}
So let's say, for instance, we are
an insurance company just for the

{{00:06:03}}
sake of simplicity, and we want to
answer the question of a customer.

{{00:06:08}}
We need to understand which contract
does this customer have, right?

{{00:06:12}}
If we by accident pick another contract
and try to answer the question with that

{{00:06:17}}
contract, we get something completely
wrong, even if the model is super smart.

{{00:06:21}}
That's data centric work.

{{00:06:23}}
And so that's what kind of challenges
that you're trying to overcome.

{{00:06:26}}
Leveraging this technology, but making
sure it's reliable, trustworthy, and

{{00:06:31}}
actually works on your own use cases.

{{00:06:33}}
That's data centric work.

{{00:06:34}}
And how soon are we going to see this
approach become an industry standard?

{{00:06:41}}
I do believe really, believe it
will be this year and next year.

{{00:06:45}}
In our lifetime, hopefully.

{{00:06:46}}
Yeah, no, absolutely, because What
you were able to see is the last

{{00:06:51}}
year was the year of demos, right?

{{00:06:54}}
Because you had a lot of retrieval
augmented generation demos that

{{00:06:58}}
were, that felt like magic.

{{00:07:00}}
You would say, here's my chat GPT.

{{00:07:02}}
I have a document.

{{00:07:03}}
I connect them.

{{00:07:04}}
Wow.

{{00:07:05}}
And all of a sudden know
so much about my company.

{{00:07:07}}
But you were able to
tell, okay, if you scale.

{{00:07:10}}
The amount of data that the
large language model has access

{{00:07:13}}
to, it's scaled the complexity.

{{00:07:16}}
And the only way to overcome
this is with data centric REG.

{{00:07:18}}
So any kind of data team that wants to
put a GenAI use case into production,

{{00:07:24}}
and that's all of the data teams,
because a lot of budget is being

{{00:07:28}}
spent on GenAI at the moment, the
only way is via data centric REG.

{{00:07:33}}
So I'm 100 percent sure that it will
very soon get a lot of attention.

{{00:07:37}}
And I think it will
already happen this year.

{{00:07:39}}
Oh, fantastic.

{{00:07:40}}
Current AI is ahead of, the
curve then in that sense.

{{00:07:45}}
How do you think that this approach,
as with anything in any industry,

{{00:07:51}}
how do you think it can be adopted
both safely and effectively?

{{00:07:56}}
I think what we always tell our customers
is to think in steps, you don't need

{{00:08:02}}
to go same with autonomous driving, you
have level 0, only the humans driving,

{{00:08:06}}
level 5, only the cars driving, don't
go from level 0 to level 5, right?

{{00:08:11}}
That would be too much.

{{00:08:13}}
But there's a level 1, like the
parking assistant in a way, and you can

{{00:08:16}}
easily build a parking assistant now.

{{00:08:19}}
We try to tell our customers, Do
it in an iterative approach, make

{{00:08:23}}
small steps, and learn a lot.

{{00:08:25}}
That seems to be a way that works very
well with data centric work as well.

{{00:08:29}}
Because you can, for instance,
carve out the data domain

{{00:08:32}}
that you want to start with.

{{00:08:34}}
For instance, one product
line or one general problem.

{{00:08:39}}
And if you start with that, you can
quickly get something up and running.

{{00:08:42}}
You see the first value, you can
get people very excited about

{{00:08:47}}
this in your own company, which
I think is also very crucial.

{{00:08:51}}
And so it's something.

{{00:08:52}}
To build trust, to get quick
results, and from their plan,

{{00:08:55}}
okay, what are my next steps?

{{00:08:57}}
Seems to be a good approach.

{{00:08:59}}
I like that, very like phased,
very controlled approach.

{{00:09:04}}
I think that's, I think
that's a good answer.

{{00:09:07}}
In terms of CAN AI, you've had
quite a day today, already.

{{00:09:12}}
You're here again tomorrow,
as we were in, in day one.

{{00:09:17}}
You were founded in 2020.

{{00:09:18}}
You've actually made quite a
lot of development within your

{{00:09:22}}
company in the last four years.

{{00:09:24}}
You have offices in Berlin and
Bonn, if I'm right in saying that.

{{00:09:30}}
You're going at light speed.

{{00:09:31}}
What can we expect you to do next?

{{00:09:33}}
Yeah, so we want to be pioneers in
the data center Greg space, right?

{{00:09:37}}
We will spend a lot of effort and
time trying to educate people on how

{{00:09:41}}
they can implement this approach.

{{00:09:43}}
And we see that it's actually the way that
brings those use cases into production.

{{00:09:48}}
So we at Kern, we just want to do as much
as possible to boost this kind of approach

{{00:09:53}}
and also grow a lot with this approach.

{{00:09:56}}
And we see it's basically the approach
that's necessary to bring GenAI with

{{00:10:01}}
industries like insurance, finance.

{{00:10:04}}
to production.

{{00:10:04}}
So that's where we want to head and do
a lot of pioneering work in that way.

{{00:10:08}}
That's really fantastic.

{{00:10:10}}
Johannes Hötter, the co
founder and CEO of Kern AI.

{{00:10:15}}
It's been a pleasure to meet you today.

{{00:10:17}}
Pleasure to meet you.

{{00:10:18}}
so much.

{{00:10:18}}
And thank you to everybody
who's joined this conversation.

{{00:10:22}}
Online.

{{00:10:23}}
I hope you enjoyed it
just as much as I did.

{{00:10:25}}
For more information, please
follow us along at fintechfutures.

{{00:10:30}}
com.

{{00:10:30}}
And until then, I shall see you next time.

{{00:10:33}}
Thank you.

Hello everybody and
welcome to Fintech Futures.

We are here today at Finovate
Europe in London, UK.

My name is Tyler Pathe, I'm a reporter
for Fintech Futures and today I'm very

happy to be joined by the CEO and Co
Founder of Kern AI, Johannes Hötter.

You've you've come from
Germany, I believe.

We're very happy to have you here
at this event here in London, and

You've just done a demo
this afternoon of Kern AI.

Can you tell us a little
bit about how that went?

It was a live demo, which is
always something, yeah, special.

Because you always want to make sure
that everything works exactly as you

It's very from what I viewed, I was very
thankful not to be on the stage because

But lots of really interesting
technologies and concepts

There was not one where I was like,
oh, it's boring like everybody

had something to say and a
reason to say it, which I loved.

Before we get into our conversation
about Client AI and your solution

and everything, why don't you give
us a little bit of color into What

Kern AI is, who you are, and more
about the origin story, really.

We initially started Kern
as a student consultancy.

So I've been in the field of
AI for roughly eight years now.

So we did projects during
our master's degree.

Ranging from weather prediction up to
database chatbots and e commerce AI.

And back then, roughly in 2020, we fell
in love with natural language processing.

We saw GPT 2 back then,
which was mind blowing.

You could tell, okay, this is quickly
becoming much and much better.

And so what we were doing a lot is we
were working with insurers and banks.

And what we could tell is the teams
that put use cases actually into

They never worried about
the algorithm, right?

So they, when you think about
AI software, typically consists

They always worried more about, okay,
how can I be sure that the data that

And so we saw this trend and we thought,
okay, if the algorithms themselves

become better and better we should
focus on Giving tools to companies

Initially a student consultancy and then
transitioning into a software company.

That's that's really interesting because
I agree with you wholeheartedly there

that these models are only as good
as the data that they're fed with.

Like you're only as
fast as you're slowest.

We're here really to talk more about data
centric retrieval augmented generation.

Very big term very big
term, very complex term.

So retrieval augmented generation,
it sounds complex, but it's.

It just means if you think of
ChatGPT, you know that depending

on how you formulate a question,
you get different outputs, right?

So in Retrieval Augmented Generation,
the idea is ChatGPT itself, and

with ChatGPT I just mean large
language models in general, they

They have a very broad world knowledge
and they can answer a lot of complex

So for instance, if I asked something
like, I'm a consultant and would like

The large language model can only answer
this question if it has the context of my

policies of my tariff side, and that's the
idea of Retrieval Augmented Generation.

You don't start completely from scratch
building your own large language

model, but you take something that
already exists, that is extremely

smart in a way, you just try to
understand how can I improve the input.

That's really interesting in that
it expands the scope of these large

language models, that's, I think from
ones that I've played around with,

they've been limited with what they
know and what they're, able to say.

And it's, really difficult to argue
against data centric RAG because

it's like why would you want
something that was less capable?

And it's it needs to be fed with data
in order to produce this outcome.

What kind of challenges does this
data centric RIG approach solve?

Yeah, so generally the idea of data
centric REG, Equivalent Augmented

Generation, means To put it in
simple terms, let's say we have

Data centric means you spend 55 minutes
just looking at the data, trying to

FinTech Futures

Expert Insight

FinovateEurope did not fail to facilitate a host of compelling conversations between industry leaders at this year's event in London, UK, chief among them being FinTech Futures' fascinating discussion with Johannes Hötter, CEO and co-founder of Kern AI.