In this episode, we're joined by Eric Daimler, CEO & co-founder of Conexus AI, Inc, an MIT spin out. We discuss the Conexus software platform, which is built on top of breakthroughs in the mathematics of Category Theory, and how it guarantees the integrity of universal data models. Eric shares real-world examples of applying this approach to various complex industries, such as transportation and logistics, avionics, and energy. Listen to this episode wherever you listen to podcasts. Eric Daimler: https://www.linkedin.com/in/ericdaimler/ Joey Dodds: https://www.linkedin.com/in/joey-dodds-4b462a41/ Rob Dockins: https://galois.com/team/robert-dockins/ Galois, Inc.: https://galois.com/ Contact us: podcast@galois.com
In this episode, we're joined by Eric Daimler, CEO & co-founder of Conexus AI, Inc, an MIT spin out. We discuss the Conexus software platform, which is built on top of breakthroughs in the mathematics of Category Theory, and how it guarantees the integrity of universal data models. Eric shares real-world examples of applying this approach to various complex industries, such as transportation and logistics, avionics, and energy.
Listen to this episode wherever you listen to podcasts.
Eric Daimler: https://www.linkedin.com/in/ericdaimler/
Joey Dodds: https://www.linkedin.com/in/joey-dodds-4b462a41/
Rob Dockins: https://galois.com/team/robert-dockins/
Galois, Inc.: https://galois.com/
Contact us: podcast@galois.com
0:00:00.550,0:00:03.470
[Joey] Welcome to another episode of
building better systems podcast,
0:00:03.760,0:00:07.470
where we chat with people in industry
and academia that work on hard problems
0:00:07.471,0:00:10.590
around building safer and more
reliable software and hardware.
0:00:10.930,0:00:12.110
My name is Joey Dodds.
0:00:12.250,0:00:12.750
[Rob] And I'm Rob Dockins.
0:00:12.750,0:00:16.950
[Joey] Rob and I work at Galois. Today,
we're talking with Eric Daimler,
0:00:17.210,0:00:18.630
CEO and founder of Conexus,
0:00:18.950,0:00:22.070
a company that helps people ensure the
correctness of their data integrations.
0:00:22.810,0:00:23.610
In this episode,
0:00:23.610,0:00:26.950
we talk about what Conexus is trying
to achieve and give some examples of
0:00:26.951,0:00:28.310
challenges that their
customers are facing.
0:00:28.810,0:00:32.430
We also discuss how they've taken
concepts from math and use those to ensure
0:00:32.431,0:00:35.070
that their clients are combining
their data in ways that are logically
0:00:35.071,0:00:38.750
consistent with respect to their
business rules. Let's start the episode.
0:00:38.979,0:00:40.470
[Intro] Designing manufacturing,
0:00:40.720,0:00:44.150
installing and maintaining the
highest speed electronic computer,
0:00:44.550,0:00:47.030
the largest and most complex
computers ever built.
0:00:58.490,0:00:59.630
[Joey] Thanks for joining us, Eric.
0:01:00.110,0:01:00.943
[Eric] Good to be here.
0:01:01.210,0:01:05.950
[Joey] So I wanted to start by just asking you
to sort of give us a high level overview
0:01:05.951,0:01:10.069
of what you and your company are really
interested in doing and what you're
0:01:10.070,0:01:10.903
really trying to do.
0:01:11.340,0:01:12.190
[Eric] Sure, sure.
0:01:12.250,0:01:16.550
So Conexus is a software
as a service company,
0:01:17.080,0:01:19.870
developing a product based
on a discovery in math,
0:01:19.920,0:01:23.950
which is probably the first and only
company you're gonna hear about this year
0:01:23.951,0:01:24.790
that has done that. I mean,
0:01:24.791,0:01:29.790
to have MIT say it's the
first spin out of MIT's math
0:01:29.791,0:01:33.950
department in the history of the
Institute. So, Conexus is in that way,
0:01:34.591,0:01:38.670
literally unique in developing
software based on math.
0:01:38.730,0:01:42.750
The domain of math is in category
theory and the application for it,
0:01:42.751,0:01:46.270
we are using in applications
on databases. So,
0:01:46.330,0:01:48.630
everybody's kind of
gotten the memo, we'd say,
0:01:48.631,0:01:51.190
about data being the new oil and all that.
0:01:51.490,0:01:56.270
What's less known is
that the data sources are
0:01:56.300,0:01:59.230
also increasing quadratically
or exponentially. And
0:02:00.991,0:02:05.550
then therefore what the tough
part is in the data relationships
0:02:06.020,0:02:09.830
that is expanding in a
unfathomably large rate,
0:02:10.110,0:02:11.870
bringing that knowledge together,
0:02:11.871,0:02:16.030
which is represented in
those data relationships and
capturing it throughout an
0:02:16.031,0:02:20.430
enterprise that guarantees the
integrity of that meaning is what
0:02:21.090,0:02:22.750
Conexus does. So I can
give you an example.
0:02:23.370,0:02:27.950
So we'll take a very sophisticated
company, Uber who had a Greenfield,
0:02:27.951,0:02:32.110
they've got a open plan about how to
develop their IT infrastructure with an
0:02:33.791,0:02:38.310
effectively infinite balance sheet to
fund it. And some very, very smart people.
0:02:38.419,0:02:39.910
That that's all true.
0:02:40.260,0:02:44.070
What we found is that
like every other company,
0:02:44.500,0:02:49.470
they focus on the business, not
optimizing for an ideal IT infrastructure.
0:02:49.970,0:02:54.910
So Uber then had the problem
of this large complex
0:02:54.930,0:02:59.840
system around the world that
was responsible to business
0:02:59.841,0:03:03.520
owners by jurisdiction or
by city then had a problem.
0:03:03.740,0:03:07.480
So Uber has a problem of, uh,
business intelligence questions.
0:03:07.870,0:03:12.480
What is the driver supply
given its championship
0:03:12.590,0:03:13.720
weekend? Uh,
0:03:13.721,0:03:18.400
or what is the privacy lattice
for Massachusetts versus
0:03:19.780,0:03:23.040
France for drivers licenses
versus license plates.
0:03:23.370,0:03:24.880
Those sort of every day,
0:03:25.020,0:03:30.000
day to day business questions were
really difficult for them to execute
0:03:30.060,0:03:34.680
on when they then had 300,000 databases.
0:03:35.020,0:03:39.400
So this is a very rich company with
very smart people with a Greenfield
0:03:39.470,0:03:44.360
opportunity, but still had a
non-optimal IT infrastructure.
0:03:44.830,0:03:46.600
They then looked about how to solve this,
0:03:46.601,0:03:51.320
what sort of commercial solutions
involved or existed to try to
0:03:51.321,0:03:54.360
integrate the data or bring
together these data models.
0:03:54.361,0:03:59.280
They had Stanford in their
backyard and explored
0:03:59.540,0:04:03.080
the landscape of solutions
to solve this problem.
0:04:03.790,0:04:08.240
They like many of Conexus
customers found the solution lies
0:04:08.580,0:04:13.560
deeper than the computer science. What
me and my co-founders PhDs are in,
0:04:13.980,0:04:15.200
it lies in the math.
0:04:16.020,0:04:19.000
So they looked at this domain
of math called category theory,
0:04:19.001,0:04:23.920
which is a type of meta math, and that
meta math helps solve the problem.
0:04:24.060,0:04:26.520
So we happen to be 40
miles north of them. And
0:04:28.041,0:04:31.839
Conexus is the recognized leader in the
software expression of category theory.
0:04:32.020,0:04:36.600
So we worked with Uber over a number
of months to develop a solution
0:04:37.020,0:04:41.839
to bring together in one, we'll
say, universal data warehouse,
0:04:42.520,0:04:47.160
you know, it's a universal
data model for those 300,000
0:04:47.520,0:04:49.960
databases, you know, without
moving the bits, right?
0:04:49.961,0:04:53.320
We're not a data lake or any
of those sort of funny names.
0:04:53.940,0:04:57.160
We worked together with Uber
to bring all that together.
0:04:57.800,0:05:01.360
So Uber then could answer these
ordinary business questions.
0:05:01.700,0:05:02.680
To have them tell it,
0:05:03.070,0:05:07.839
they then save on the order
of 10 plus million USD a year
0:05:08.110,0:05:12.200
with the new alacrity with which
they're able to answer these ordinary
0:05:13.120,0:05:15.720
business questions and respect
privacy lattice. So that,
0:05:15.721,0:05:19.120
that's what Conexus has
done for our customers.
0:05:19.620,0:05:22.760
[Joey] So just to try to restate
what you're saying,
0:05:22.779,0:05:27.440
the problem you're really trying
to solve: all the data is stored,
0:05:27.441,0:05:31.080
you're not dealing with how they're
getting the data where the data is,
0:05:32.500,0:05:36.120
but you're trying to make the connection
between what people want to know from
0:05:36.121,0:05:39.080
the data and the data. That's
kind of where you all are living.
0:05:39.620,0:05:44.279
[Eric] We are guaranteeing the
meaning of the data is
0:05:44.310,0:05:48.480
preserved regardless of the
context under which it's queried.
0:05:48.860,0:05:52.480
So we guarantee the semantics
as the data is transformed.
0:05:53.160,0:05:56.720
[Rob] I know just enough category theory
to be dangerous. So, you know,
0:05:57.260,0:05:59.279
for someone who's in
my boots, like what's,
0:05:59.500,0:06:03.040
what's the interesting connection here
between this business intelligence
0:06:03.041,0:06:07.240
problem that you're solving
and the meta math, you know,
0:06:07.240,0:06:08.880
the ivory tower of mathematics.
0:06:09.660,0:06:13.600
[Eric] So category theory is not new.
It's been around for a while.
0:06:14.140,0:06:17.720
It was invented or discovered,
you know, the nature of math,
0:06:18.480,0:06:21.560
discovered to solve a problem in math,
0:06:21.900,0:06:26.520
in guaranteeing that the
translations of problems
0:06:26.589,0:06:30.440
between domains, you know,
between algebra and geometry,
0:06:31.081,0:06:35.200
doesn't have four become
approximately four
0:06:35.380,0:06:36.839
You have to have it be exact.
0:06:37.380,0:06:41.480
So we can describe a circle, for example,
0:06:41.860,0:06:45.720
on a Cartesian coordinate X, Y
coordinates, you know, X squared,
0:06:45.721,0:06:49.040
plus Y squared equals one that
perfectly describes a circle. You know,
0:06:49.110,0:06:53.160
that is a representation in geometry
and a representation in algebra that
0:06:53.850,0:06:56.200
needs to be maintained regardless of how.
0:06:56.279,0:06:58.240
Whether you're gonna use an abstract math,
0:06:58.241,0:07:00.600
like type theory or set
theory or graph theory,
0:07:00.601,0:07:02.920
which is where my
academic research was in,
0:07:03.339,0:07:07.880
the discovery was that categories or
0:07:07.881,0:07:12.680
categorical algebra category theory
could be applied to databases.
0:07:13.350,0:07:18.160
This concept could be applied to
databases so that the integrity
0:07:18.690,0:07:22.720
guaranteed by the logic
of math would translate
0:07:23.590,0:07:28.120
between different domains.
That's the epiphany. You know,
0:07:28.340,0:07:33.040
the current solutions that Conexus
has run into with some clients
0:07:34.100,0:07:37.400
is that you might use legacy software.
0:07:37.420,0:07:41.920
So we work with a
manufacturer of airplanes that
0:07:42.700,0:07:46.720
had these old IBM S four
hundreds running on Cobal, or
0:07:48.240,0:07:49.360
Cobal running has four hundreds.
0:07:49.361,0:07:54.240
And these old systems you'd have
migration technologies to be sure,
0:07:54.900,0:07:58.640
but those would have to rely on testing
0:07:59.630,0:08:04.000
when you are in many modern
digital environments,
0:08:04.340,0:08:07.720
the testing is fine. You know, you
can use a Monte Carlo simulation,
0:08:07.740,0:08:12.440
and if you get a Ddigital ad served,
that's not perfect, it's fine.
0:08:12.980,0:08:16.920
But if you're flying a 100
million dollar jet plane,
0:08:17.270,0:08:22.240
that is in a high consequence
environment with a pilot or two
0:08:22.480,0:08:26.360
whose life is at stake, you don't
really, you don't wanna have errors.
0:08:26.360,0:08:31.120
You can't afford errors. That's
where this math is brought to bear.
0:08:31.780,0:08:36.200
It proves the integration
of the underlying logic
0:08:36.740,0:08:40.600
so that you can depend
upon it with your life.
0:08:41.220,0:08:43.280
That's what category theory provides.
0:08:43.580,0:08:48.320
It brings together that
logic in approvable way.
So in this particular case,
0:08:48.321,0:08:53.240
this airplane manufacturer has
formal methods deployed that we've
0:08:53.241,0:08:56.080
all learned in computer
science school. You know,
0:08:56.150,0:09:01.080
that help prove a
subsystem of an airplane.
0:09:01.420,0:09:06.160
And we would learn it when we're
doing code that we need to prove an
0:09:06.161,0:09:10.360
algorithm is complete, or can prove
a set of algorithms are complete.
0:09:10.420,0:09:11.559
We can do that, and that's fine.
0:09:12.059,0:09:16.600
And so Boeing or Airbus
or Desole or Bombardier,
0:09:16.910,0:09:19.320
they can do that for
these individual systems,
0:09:19.820,0:09:24.640
but there is literally no
method by which any of those
0:09:24.641,0:09:29.270
manufacturers or any other company
can bring together those systems
0:09:30.390,0:09:33.030
in an equally robust way.
0:09:33.420,0:09:37.670
There's no provable system
for integrating those
0:09:38.230,0:09:41.520
multiple validated formal methods.
0:09:42.050,0:09:46.360
Those then get resolved to the
extent that they are resolved,
0:09:46.750,0:09:51.200
just through iterations of
simulations, iterations of testing,
0:09:51.660,0:09:56.360
and to be fair, some rather robust
physical testing. But that's,
0:09:56.540,0:10:00.200
that's what we see in the world. You
know, I was just talking to a NASA,
0:10:01.600,0:10:05.800
engineer a couple weeks ago in
Houston. And this guy was saying,
0:10:06.230,0:10:07.080
they don't know,
0:10:07.790,0:10:11.400
because they don't use category theory
at the foundations of some of their
0:10:11.470,0:10:13.360
rockets yet, but they're looking to,
0:10:14.030,0:10:17.120
they don't know what's going
on inside of their systems.
0:10:17.540,0:10:22.480
The consequences are both that
they are left with some level
0:10:22.620,0:10:26.080
of uncertainty that they don't
like for obvious reasons,
0:10:26.580,0:10:30.880
but it's also resulting
in an overengineering
0:10:31.460,0:10:34.160
of their systems. So they will put
0:10:36.610,0:10:38.640
extra thickness, extra connectors,
0:10:39.550,0:10:43.679
because they just don't know about their
system. That obviously adds weight.
0:10:43.700,0:10:46.120
It adds complexity, it
adds cost, it adds time,
0:10:46.870,0:10:50.920
because they just don't know about these
systems, despite everything that we,
0:10:51.059,0:10:55.160
we have learned about formal
methods in subsystems.
0:10:55.620,0:10:57.360
So that's what Conexus provides.
0:10:57.360,0:11:01.679
That's a sort of solution
that Conexus brings to
0:11:02.390,0:11:05.400
various industries, from
transportation and logistics toavionics
0:11:07.380,0:11:08.240
to energy.
0:11:08.620,0:11:12.160
[Rob] So let me see if I can rephrase
that a little bit from my biases,
0:11:12.161,0:11:15.280
I'm I come from a formal methods
background. So, you know,
0:11:15.521,0:11:17.360
that's the native language that I speak.
0:11:17.361,0:11:20.480
So it sounds to me like what you're
saying is that you've developed some
0:11:20.510,0:11:21.343
clearinghouse
0:11:23.250,0:11:27.640
logic slash mathematical language in
which you can bring these other artifacts
0:11:27.641,0:11:28.920
that other people have generated.
0:11:28.940,0:11:32.000
You can bring them together and have
them sort of speak the same language,
0:11:32.240,0:11:36.480
in some way. You can integrate these
different formal and informal methods,
0:11:37.110,0:11:38.240
have I got the right idea?
0:11:38.780,0:11:40.440
[Eric] We are complimentary to
formal methods. I mean,
0:11:40.441,0:11:44.200
I come from formal methods as well.
That's what I learned in school. Now,
0:11:44.201,0:11:45.720
this is complimentary to formal methods.
0:11:45.721,0:11:50.559
What this does in complimenting
formal methods is it provides
0:11:50.620,0:11:53.830
the optimal path for these relations.
0:11:54.340,0:11:58.790
Formal methods really doesn't
have anything to say about
what is the ideal path
0:11:59.490,0:12:01.830
in trillions of possible relationships.
0:12:02.660,0:12:05.170
There is a computationally infeasible,
0:12:05.171,0:12:09.970
meaning infinite
that would have to be solved to try
0:12:10.030,0:12:14.610
to do what is done here without this
0:12:16.250,0:12:20.290
abstract math. You know, it's done in
quantum theory in quantum computers,
0:12:20.400,0:12:24.050
it's done in smart contracts on the
blockchain. You know, this is done today.
0:12:24.360,0:12:27.490
It's just done in different contexts.
Those contexts might be a little sexier.
0:12:27.750,0:12:30.090
So, you know, quantum computers, you know,
0:12:30.330,0:12:33.410
we would not understand the
output of quantum computers,
0:12:33.710,0:12:38.130
if not for category theory, being
applied to quantum compilers.
0:12:38.510,0:12:41.010
So that's category theory in action today.
0:12:41.470,0:12:43.809
And similarly in smart contracts,
0:12:44.080,0:12:48.770
that the sophistication of smart
contracts wouldn't be enabled if not for
0:12:49.970,0:12:52.890
category theory. It's just that. And
type theory is kind of related to this.
0:12:53.160,0:12:56.770
It's just that it's a lot more fun to
talk about qubits and get the physics of
0:12:56.771,0:13:00.570
quantum computers. It's a lot more
fun to talk about smart contracts.
0:13:00.760,0:13:03.809
It's just that the underlying
language of category theory,
0:13:03.920,0:13:08.850
type theory gets a little
bit lost in those narratives,
0:13:09.130,0:13:11.929
but it's already applied to other domains.
0:13:12.230,0:13:15.170
So I think I'm still struggling a
little to understand, I guess, what
0:13:17.171,0:13:19.010
exactly you're combining.
Because I think I heard,
0:13:19.491,0:13:24.050
we're kind of combining ways of thinking
about data and maybe a semantics for
0:13:24.051,0:13:27.690
data. I think I heard we're combining
people's understanding of subsystems,
0:13:28.070,0:13:32.090
maybe that's represented
as data, I'm not sure. Can
0:13:33.731,0:13:35.050
you try to clarify a little, I guess,
0:13:37.691,0:13:41.530
what exactly are we
composing and combining,
0:13:41.570,0:13:44.690
I guess in the use that
you all are applying.
0:13:45.250,0:13:50.050
[Eric] Whatever companies or organizations
or use cases demand of the
0:13:50.051,0:13:52.050
composition is what will be composed.
0:13:52.429,0:13:56.929
We have nothing to say on
what anyone wants to compose.
0:13:57.260,0:13:59.490
Where Conexus works,
0:14:00.220,0:14:02.730
where category theory is required,
0:14:03.350,0:14:08.130
is if you want to prove
the robustness of that
0:14:08.131,0:14:11.250
composition. So in Uber's particular case,
0:14:11.400,0:14:16.010
they could already in some loose
way, bring together their 300,000
0:14:18.210,0:14:21.890
databases. It's just that they can't
prove that the result is accurate.
0:14:22.510,0:14:25.410
And that's similar to NASA,
similar to the avionics company.
0:14:25.760,0:14:27.210
They can do a lot of different things,
0:14:27.230,0:14:31.210
but is it become a sort of data
model tower of Babel. You know,
0:14:31.211,0:14:34.570
the composition is in the rules to,
0:14:34.810,0:14:39.810
is a short answer to your question. But
how those are composed is up to the SME.
0:14:40.280,0:14:45.090
It's not up to Conexus how the
composition happens in a larger system
0:14:45.630,0:14:49.930
is up to the demands of the system being
constructed. It's not up to Conexus.
0:14:50.530,0:14:54.930
What a Conexus instance proves
is a guaranteed integrity
0:14:55.710,0:15:00.370
of that composition of
models or composition of
0:15:00.460,0:15:01.293
rules.
0:15:01.630,0:15:06.410
So rather than focusing on
what comes into these, I
0:15:07.971,0:15:10.930
guess, to software that's
performing this active composition,
0:15:10.931,0:15:11.770
you're focusing on,
0:15:11.990,0:15:16.970
on that composition itself
and checking that composition
0:15:16.971,0:15:17.890
is happening correctly.
0:15:18.230,0:15:22.130
[Eric] We prove that the integrity
of the semantics is preserved.
0:15:22.450,0:15:26.170
[Joey] Is it possible to do that without
understanding what's coming in? Can
0:15:28.251,0:15:32.130
I know that I'm composing data
correctly without sort of intimately
0:15:32.131,0:15:33.930
understanding the data itself?
0:15:34.190,0:15:37.210
[Eric] So it's not about the data,
it's about data rules,
0:15:38.600,0:15:41.930
data models. So there
is a logical data model.
0:15:42.480,0:15:47.090
What Conexus provides is
a guaranteed integration
0:15:47.630,0:15:51.010
of those models, guaranteed
integration of the rules,
0:15:51.550,0:15:55.450
so that there exists then
a universal warehouse,
0:15:55.890,0:16:00.090
a universal data model, a
universal knowledge graph.
0:16:00.750,0:16:04.090
[Joey] So you don't have to worry about the
specifics of the data itself, but
0:16:06.110,0:16:07.850
you do have a representation of, I guess,
0:16:08.331,0:16:11.970
shape of the data to some extent or
some expectations about the data.
0:16:11.971,0:16:16.850
Are those coming from you? Or, a
company like Uber, as you mentioned,
0:16:16.850,0:16:19.130
has those descriptions
available in general?
0:16:19.560,0:16:24.130
[Eric] Yeah, we wouldn't define
anything about the rules,
0:16:24.730,0:16:26.170
anything about the
characteristics of the rules.
0:16:26.600,0:16:31.370
It's definitely part of the process
that you have to do some degree of
0:16:31.530,0:16:32.650
entity resolution and disambiguation,
0:16:33.630,0:16:38.530
but that's not any sort of secret sauce
for Conexus. That's just part of theflow
0:16:40.050,0:16:43.570
chart of work in any of
these exercises. You know,
0:16:43.571,0:16:48.210
foundational for Conexus is
using a chase engine to bring
0:16:48.530,0:16:53.450
together all the possible relationships
in the exercise that one would go
0:16:53.451,0:16:58.370
through doing formal methods and look
for the optimal path that then defines
0:16:58.710,0:17:01.130
the totality of the universal warehouse.
0:17:01.560,0:17:05.690
Another way of thinking about it is
this is a sort of deductive database,
0:17:06.190,0:17:10.970
and this deductive database then
allows for a database of viewpoints,
0:17:11.850,0:17:14.010
or a database of perspectives.
0:17:14.230,0:17:18.170
So instead of a database
of just data in a table,
0:17:18.760,0:17:23.530
it's how is the data being
used for a particular
0:17:23.880,0:17:24.710
user,
0:17:24.710,0:17:29.050
or how are these data models then
represented in one world view
0:17:29.720,0:17:32.250
without requiring consensus,
0:17:33.060,0:17:37.210
which is just a complete
mind shift from how most
0:17:38.131,0:17:42.010
people have to operate in
requiring consensus. You know,
0:17:42.011,0:17:46.210
blockchains work that way.
They require consensus. I mean,
0:17:46.211,0:17:47.410
you know,
0:17:47.411,0:17:52.119
many of these require the energy
consumption for the entire country of
0:17:52.360,0:17:55.480
Ecuador. It's just resource inefficient
because it requires consensus
0:17:57.630,0:18:02.119
this level of math upon which
Conexus is expressing for our
0:18:02.670,0:18:06.359
roughly fortune 500 clients
doesn't require consensus,
0:18:06.900,0:18:11.880
but nonetheless gets to this
universal data model that then
0:18:11.881,0:18:16.240
can be queried any which way a user wants.
0:18:17.630,0:18:20.650
[Rob] So I wonder if I could dial in
a little bit more into this.
0:18:20.650,0:18:23.369
You've mentioned this notion
of rules a couple of times,
0:18:23.370,0:18:27.250
and I feel like I'm still struggling
to have a good mental model of what you
0:18:27.251,0:18:30.690
mean by that. To me, when I talk
about rules and databases, you know,
0:18:30.691,0:18:35.010
I go back to my old database 201
course, or whatever. And you know,
0:18:35.090,0:18:37.090
it talks about data invariance
that you have to satisfy.
0:18:37.650,0:18:39.290
I get the feeling you're
talking about something else,
0:18:39.291,0:18:42.970
but I'm not quite sure
what yet. Are those views,
0:18:43.070,0:18:46.609
or are they in variance on the data
or are they ways that you combine data
0:18:46.890,0:18:47.359
together?
0:18:47.359,0:18:52.330
[Eric] A helpful way to think about that maybe
is a logical data model in the database.
0:18:52.630,0:18:54.930
So whatever your logical data model is,
0:18:57.160,0:19:01.090
however it's represented, whether it's
represented in Excel or some other way,
0:19:01.880,0:19:06.570
it's that logical data model
representation that then is put into the
0:19:06.640,0:19:09.609
semantics of a Conexus instance,
0:19:09.610,0:19:14.369
which is essentially like SQL.
And it's just that representation
0:19:16.040,0:19:20.290
with a Conexus instance software
that then is able to deduce
0:19:20.830,0:19:25.690
what is the optimal connection between
0:19:27.091,0:19:29.930
those models or those
relationships. They call rules,
0:19:30.250,0:19:32.530
they call 'em business
requirements if that's better.
0:19:33.230,0:19:36.330
[Joey] And so you've called, I think,
a couple of times you've said
0:19:38.171,0:19:42.770
Conexus instance. So you're going
in, you have a piece of software --
0:19:42.800,0:19:47.770
is this usually installed
in your company's
0:19:47.771,0:19:50.930
cloud or is this is something that you
all run locally and they access as a
0:19:50.931,0:19:51.119
service?
0:19:51.119,0:19:55.730
[Eric] Yeah, we are a software as a service
company in that this is just a license,
0:19:56.190,0:20:00.810
but so it's cloud native, but it can
certainly also be run on prem, you know,
0:20:00.811,0:20:04.730
with our financial services clients and
with some of the clients in governments,
0:20:04.731,0:20:06.730
you know, they want it to be
run on prem, which is also fine.
0:20:08.350,0:20:09.730
[Joey] And so, I guess,
0:20:11.570,0:20:14.010
as a user of one of these instances,
0:20:14.310,0:20:18.810
am I asking it questions directly or is
it kind of keeping an eye on things and
0:20:18.811,0:20:22.090
telling me whether things went well or
not? How do I interact with this thing?
0:20:22.350,0:20:26.770
[Eric] So you're querying this
universal data warehouse call it.
0:20:26.990,0:20:31.410
And what it is gonna tell you in
response is if there are any logical
0:20:31.590,0:20:36.250
contradictions. So I can give you a
use case , that might appear. You know,
0:20:36.251,0:20:38.210
we work with this big
engineering firm and this
0:20:39.970,0:20:42.609
engineering firm happens to
do oil and gas exploration.
0:20:43.109,0:20:45.490
So their workflow goes like this,
0:20:46.090,0:20:50.520
that they have have exploration that
then says, "Hey, can you do an explore?
0:20:50.619,0:20:52.800
Can you do a well drill here?
0:20:52.980,0:20:57.680
We think there is some resources
to extract." So the well people
0:20:58.270,0:21:02.520
they look at the situation and find
that the ground is a little softer than
0:21:02.521,0:21:06.880
normal and have to modify
part of their well then
0:21:07.841,0:21:11.680
it goes down to approval and then
fabrication and then distribution.
0:21:11.681,0:21:13.440
And then it gets down to the
people that actually dig the hole.
0:21:13.660,0:21:18.160
The people that dig the hole also had
to modify their approach because the
0:21:18.161,0:21:20.480
ground was a little softer
than they had expected. Well,
0:21:20.481,0:21:24.160
that approach then broke the flange
that was modified back at the first,
0:21:24.680,0:21:25.560
at the first sequence.
0:21:26.580,0:21:29.400
And that's a problem because then
the flange falls in the hole.
0:21:29.401,0:21:31.640
They have to fill it up with
cement. They have to move their rig.
0:21:31.641,0:21:34.400
And apparently this thing
costs a lot of money.
0:21:34.460,0:21:39.320
The story we hear is $50 million, which
is just a mindboggling amount. And,
0:21:39.321,0:21:41.040
you know, nobody wants that. You know,
0:21:41.041,0:21:42.600
we've heard amounts of about
a hundred thousand dollars.
0:21:42.660,0:21:45.840
We even heard of a half a
billion dollar error like this,
0:21:45.980,0:21:49.880
no lives are lost but a lot of money
and time. You know, that's a bad day.
0:21:50.340,0:21:54.320
Sso these companies spend a lot of time
0:21:55.109,0:21:59.200
then to prevent these things
iterating through data models.
0:21:59.230,0:22:02.600
Before actually fabricating the flange
before we're getting approval for the
0:22:02.601,0:22:05.640
flange, they'll send that model down the
0:22:07.641,0:22:08.520
down the flow chart.
0:22:08.980,0:22:12.600
And then it'll iterate back the first
time and then down the flow chart.
0:22:12.619,0:22:15.240
And this happened in this
particular instance as well.
0:22:16.220,0:22:20.520
But if you rely on just
doing Monte Carlo simulations
0:22:21.140,0:22:23.200
and test and fail or test and pass,
0:22:23.940,0:22:28.810
you're gonna have these errors
in some number of times.
0:22:29.730,0:22:30.230
I mean,
0:22:30.230,0:22:35.130
one error we had heard
was a mistranslation of
0:22:35.400,0:22:39.490
this footnote from Mandarin to
Spanish. And actually it was,
0:22:39.630,0:22:42.530
it was the leaving out of a footnote.
0:22:42.531,0:22:44.970
So the footnote got forgot
to be translated. These are,
0:22:44.971,0:22:49.810
these are things to be avoided
and as data relationships,
0:22:49.811,0:22:52.050
that's how we started out.
We're saying data's growing,
0:22:52.051,0:22:56.330
quadratic data sources are growing,
quadratic data relationships are just,
0:22:56.331,0:23:00.010
unfathomably large. So as data
relationships get to be so big,
0:23:00.011,0:23:02.130
or if you just think
you're ordinary database,
0:23:02.390,0:23:05.609
as you think your row count
is getting to be so large,
0:23:06.030,0:23:09.010
you have to think in
abstractions. Because you just,
0:23:09.030,0:23:12.890
you can't be thinking of trillions of
data points the same way you'd had before.
0:23:12.891,0:23:15.850
And you know, the column names
kind of, don't quite, you know,
0:23:16.340,0:23:20.850
speak to all the ways in which your
data can be represented. So this,
0:23:20.920,0:23:22.609
this particular client of Conexus,
0:23:23.119,0:23:28.090
they've come to the realization that the
different approach that is represented
0:23:28.350,0:23:32.930
by the math of category
theory is the requirement
0:23:33.200,0:23:34.810
that we bottoms up,
0:23:35.190,0:23:40.090
foundationally agree on the logical
data model for each of the many
0:23:40.091,0:23:44.010
different contexts. So one
engineer has a logical data model.
0:23:44.011,0:23:46.440
Another engineer has a
logical data model that,
0:23:46.441,0:23:51.280
so the logical data model here for the
person that designs the well then another
0:23:51.281,0:23:53.680
logical data model for the person
that's gonna drill the hole,
0:23:54.010,0:23:57.920
those get combined into this
universal data warehouse,
0:23:59.520,0:24:04.520
whenever all of us would then add or
subtract from our own universal data
0:24:04.521,0:24:06.920
model. What we then just find is,
0:24:06.980,0:24:11.880
as it gets integrated that
logical contradictions would
0:24:11.881,0:24:16.200
get exposed. That's it it's really,
as simple as that. It gets uploaded,
0:24:16.460,0:24:17.680
you know, it's definitively proved,
0:24:18.260,0:24:21.720
and then you get to see whether there's
a contradiction or whether the integrity
0:24:21.820,0:24:22.653
was maintained.
0:24:23.000,0:24:25.640
Yeah. So it sounds to me like,
if I've understood right,
0:24:26.380,0:24:30.800
the main goal here is to sort
of encode the important rules
0:24:31.510,0:24:35.640
that the business cares about in such a
way that you can monitor and make sure
0:24:35.641,0:24:37.840
that they're all in some
sense, consistent, you know,
0:24:37.841,0:24:39.840
like the guy drilling the hole,
0:24:40.010,0:24:42.560
isn't making a different assumption
in the guy building the flange that's
0:24:42.880,0:24:45.280
supposed to fit on it. And
by bringing those together,
0:24:45.441,0:24:49.040
you can expose when those things don't
match up so that they can be fixed before
0:24:49.060,0:24:51.160
you spend 50 million
building the wrong flange.
0:24:51.740,0:24:54.520
[Eric] You got it. Exactly. Okay.
That's it. That's exactly it.
0:24:54.521,0:24:59.040
And I'll even go further with
how you perfectly represented it,
0:24:59.130,0:25:03.480
which is this isn't magic. And so we
don't read anybody's mind
0:25:03.480,0:25:07.480
So the important part you said at
the beginning is worth repeating,
0:25:07.481,0:25:11.880
which is you're encoding and logic
being the rules you care about. Well,
0:25:12.160,0:25:15.160
you know, yes, it has to be really
all the rules. Right. You know,
0:25:16.440,0:25:20.560
whatever portion you leave out is the
portion that can't be considered for
0:25:20.561,0:25:21.640
logical contradictions.
0:25:22.100,0:25:27.080
[Joey] And who's doing this in encoding, because
obviously the rules exist, you know,
0:25:27.140,0:25:29.520
for each company in a range
of different forms. Right.
0:25:29.660,0:25:34.359
But you need to have some representation
of these things that probably relates
0:25:35.140,0:25:39.680
to the data models and also
is something your tools can
0:25:39.681,0:25:40.514
consume, right?
0:25:41.109,0:25:45.880
[Eric] Yeah. Now, this is just a brilliant
sequence of questions here.
0:25:45.881,0:25:47.880
You guys got it exactly. And,
0:25:48.660,0:25:52.520
you just defined our engineering
roadmap
0:25:53.150,0:25:57.760
what you have to do is define
by customer or really by,
0:25:57.900,0:26:02.160
for these very large companies,
even the divisions of the customers,
0:26:02.460,0:26:06.320
how they input their own logical data
model. So we can import anything,
0:26:06.580,0:26:06.961
you know,
0:26:06.961,0:26:11.840
we haven't experienced yet any
sort of technical risk in taking
0:26:13.310,0:26:17.720
what we receive into a
universal data model.
0:26:18.100,0:26:20.600
But we have to allocate
resources to do that,
0:26:20.601,0:26:23.960
whether it's Excel or whether it's in,
like I said, or whether it's Oracle or,
0:26:23.961,0:26:27.760
you know, some other form, you
know, as long as it's not in PDFs,
0:26:27.761,0:26:28.594
which would be the worst.
0:26:28.890,0:26:31.920
[Joey] Based more on kind of the
formal methods experience,
0:26:32.380,0:26:35.359
one of the things I would expect, and
I'm curious if you've experienced,
0:26:35.430,0:26:40.280
this is even that act of, well,
we're gonna take this kind of,
0:26:40.300,0:26:42.760
you know, maybe shaky notion or this
0:26:45.201,0:26:49.800
less consistent kind of
notion of our business logic.
0:26:49.800,0:26:54.080
And we're gonna encode that into a really
concrete, meaningful form, you know,
0:26:54.280,0:26:55.113
probably with the semantics.
0:26:56.060,0:26:59.280
My expectation based on formal methods
experience is that you'd actually find
0:26:59.281,0:27:00.200
problems even in,
0:27:00.300,0:27:04.400
in that step before you kind of take
the next step of actually, you know,
0:27:04.401,0:27:07.200
sort of thoroughly checking
for logical inconsistencies.
0:27:07.619,0:27:08.240
[Eric] You know,
0:27:08.240,0:27:13.160
I think that the logical data
model as it's represented
0:27:14.140,0:27:14.973
by
0:27:16.120,0:27:21.119
the business is sort of
manifestly their best casefor how
0:27:21.120,0:27:24.280
they run the business. We're
not gonna second guess that. You
0:27:26.680,0:27:31.640
actually point out another really great
point about implementing any part of
0:27:31.641,0:27:32.450
these systems,
0:27:32.450,0:27:37.320
which is that taking what
they've already done and just
0:27:37.321,0:27:42.200
proving that is so much more effective
than giving them yet another new tool
0:27:42.230,0:27:47.160
that will inevitably have some gaps and
0:27:48.000,0:27:49.840
that that'll take time to work out.
0:27:49.859,0:27:53.800
So we wanna just make this as
easy to our customer's workflowas
0:27:54.960,0:27:57.240
possible. And that's
what we're working to do.
0:27:57.660,0:27:59.440
[Joey] And then I guess my other question is,
0:28:04.750,0:28:09.560
I'll imagine I'm a customer, what am
I supposed to do when I query your
0:28:09.760,0:28:13.359
instance? And it's like, well,
there was an inconsistency. Um,
0:28:13.360,0:28:17.080
obviously I don't act on that data in the
way that I would maybe have otherwise,
0:28:17.080,0:28:19.480
but where do I go from there?
0:28:19.500,0:28:22.960
Do I just keep querying and only
act on the data where, you know,
0:28:22.960,0:28:25.320
where we don't see problems
or can I do better?
0:28:26.080,0:28:26.913
[Eric]
0:28:26.960,0:28:30.080
I think the logical consistency
is something you'd wanna address.
0:28:30.081,0:28:33.000
That becomes immediately apparent.
It's not like it's gonna be,
0:28:33.000,0:28:37.560
stuck upon you saying suddenly there's
a logical consistency appearing,
0:28:37.561,0:28:42.520
you know, this Tuesday. I don't know
when it happened or where it exists,
0:28:42.520,0:28:45.400
but you know, good luck. You know, these,
0:28:45.530,0:28:47.040
these sort of things get integrated very
0:28:48.601,0:28:52.680
quickly and you can see where the logical
contradiction will be taking place so
0:28:52.681,0:28:57.560
that you can bring the
appropriate expertise to bear to
0:28:57.561,0:28:58.394
correct that.
0:28:58.600,0:29:00.720
[Joey] Gotcha. So I should be able to
sort of see the inconsistency.
0:29:00.720,0:29:04.640
Will the tool guide me into figuring
0:29:05.500,0:29:08.920
out what went wrong and point me
in the right direction at all.
0:29:09.260,0:29:12.120
[Eric] I'm gonna say that's really
dependent on the use case.
0:29:12.660,0:29:17.040
It really depends on whether we
are talking about our clients in
0:29:17.041,0:29:21.160
pharmaceutical research or our
clients in risk analysis and finance
0:29:21.801,0:29:26.720
about the degree to which they're gonna
feel that they need to work on the
0:29:26.721,0:29:31.560
model themselves or do something else
to resolve the logical contradiction.
0:29:31.940,0:29:35.600
[Joey] But the good news is if I believe I've
resolved the logical contradiction,
0:29:35.640,0:29:39.840
I assume I should be able to check that
by rerunning it and sort of right away.
0:29:40.190,0:29:43.350
[Eric] That is the promise of using
math. You know, it's foundational.
0:29:44.980,0:29:46.510
This proves, you know,
0:29:46.511,0:29:51.350
just like you don't need to use
a calculator and every day run
0:29:51.750,0:29:54.390
infinite queries on your calculator.
You just kind of trust that it,
0:29:54.391,0:29:58.070
that it's proven that, you know,
nine times nine is gonna be accurate.
0:29:58.130,0:30:00.870
And then 12 times 12 is
gonna be accurate. You know,
0:30:00.871,0:30:03.990
this is a proven logical data model.
0:30:04.500,0:30:08.630
What the challenge of
this exercise becomes
0:30:09.530,0:30:11.990
is that people, organizations,
0:30:12.300,0:30:17.150
will be confronted with the degree to
which they keep knowledge in their head.
0:30:17.570,0:30:21.990
So implicit knowledge
doesn't scale well. You know,
0:30:21.991,0:30:22.830
that is
0:30:24.870,0:30:29.230
a known failure point and something
that needs to eventually be made
0:30:29.830,0:30:31.830
explicit. That's really the overhead.
0:30:32.090,0:30:36.670
And that's why the clients for Conexus
are generally larger organizations
0:30:37.140,0:30:39.830
operating in higher
consequence environments,
0:30:41.020,0:30:44.230
smaller companies, the
proxy for this, not exactly,
0:30:44.370,0:30:48.470
but the proxy is databases
under five databases
0:30:48.980,0:30:53.590
generally suggests as simple
enough infrastructure that overhead
0:30:53.930,0:30:58.270
of making the implicit explicit
is maybe not yet worth it.
0:30:58.810,0:30:59.260
Uh,
0:30:59.260,0:31:03.910
another place where this is a
little tougher is in areas where
0:31:04.711,0:31:08.350
you're the king or the queen,
and you can dictate, you know,
0:31:08.351,0:31:09.950
what your universal data model is.
0:31:09.970,0:31:13.790
So I don't need the engineers to
tell me or not tell me consensus,
0:31:13.900,0:31:14.750
I'll just tell 'em.
0:31:14.850,0:31:19.710
And that's the solution you'll
find in places like Amazon
0:31:19.890,0:31:20.723
or Apple,
0:31:21.831,0:31:26.790
where they don't have this complex
engineering system in the same way of an
0:31:26.870,0:31:28.950
avionics company or an energy company.
0:31:29.540,0:31:30.373
[Rob] Yeah, I think that's one
0:31:31.951,0:31:35.950
topic that strikes out to me when I
look at your resume is that you have
0:31:36.940,0:31:41.030
some public policy experience,
as well as business experience.
0:31:41.710,0:31:46.070
I wonder if any of the lessons that
you've learned here in this business are
0:31:46.890,0:31:48.030
things that impact
0:31:49.740,0:31:54.630
your public policy opinions or
things that you think ought to be
0:31:54.631,0:31:57.630
done or should be done. Is there
any interplay between those things?
0:31:58.530,0:32:03.150
[Eric] Yes. It's interesting that you
bring that up in several ways.
0:32:03.270,0:32:06.310
I can first say that I'm really grateful
for the time that I had serving in
0:32:07.591,0:32:09.270
policy, serving in the US government.
0:32:09.730,0:32:14.350
I came into that job with many of
the biases others might have had
0:32:14.911,0:32:17.310
going to work for the US
government, but I immediately was
0:32:19.751,0:32:24.070
impressed with the drive and
intelligence of the people with whom,
0:32:24.790,0:32:27.830
I worked in the Obama
White House, in my case.
0:32:28.710,0:32:31.470
I hope to do that again in some way.
0:32:31.470,0:32:35.230
The public services I found
to be really enlivening in
0:32:37.231,0:32:40.160
the difference one could
make. The output, I guess,
0:32:40.161,0:32:44.320
is it comes in at least two ways that
I can think of based on your question.
0:32:44.740,0:32:47.600
One is it had me see this,
0:32:48.190,0:32:50.960
Conexus's value as an opportunity.
0:32:50.961,0:32:55.880
Because I've spent my career in and
around various expressions of AI from an
0:32:56.120,0:33:00.280
academic researcher at various
schools, to a venture capitalist,
0:33:00.281,0:33:03.120
to an entrepreneur. And at that high level
0:33:04.641,0:33:07.320
of implementation, you know,
just at very, very large scale,
0:33:07.800,0:33:12.360
I quickly got to see that investors
and organizations would be very
0:33:12.361,0:33:17.200
disappointed with the returns they're
gonna get on their AI investments. Uh,
0:33:17.220,0:33:21.800
and I got to see where the
blockage was. It wasn't in, uh,
0:33:21.830,0:33:26.800
some of the ways we're looking at
it such as data cleaning or bringing
0:33:26.960,0:33:30.720
together every, all the books in the
library, throwing 'em all in and saying,
0:33:30.880,0:33:34.080
Hey, it's integrated, or then sorting
them by height and saying, great,
0:33:34.300,0:33:35.920
now it's now it's structured.
0:33:36.360,0:33:41.160
difficulty of fulfilling the promise
0:33:41.260,0:33:42.720
of AI was elsewhere.
0:33:43.140,0:33:48.120
And that's how I got to
find this research that
0:33:48.121,0:33:51.520
then led to the expression
that is Conexus.
0:33:51.900,0:33:54.000
On the other side, I guess,
0:33:54.020,0:33:58.480
or another way in which my public
policy experience informs my
0:33:59.921,0:34:04.760
current view is in encouraging
people to be a part of the
0:34:04.761,0:34:09.320
conversation around AI or around
these digital technologies so
0:34:10.601,0:34:13.680
that they can be comfortable in
how these things get expressed.
0:34:14.360,0:34:15.600
I came to find that many
0:34:17.580,0:34:22.000
people would not really understand
the definition of what AI
0:34:22.620,0:34:27.080
is, let alone how they expected it to
be put to use in their organizations.
0:34:27.340,0:34:30.360
And this is from any level of employee,
0:34:31.190,0:34:36.080
from the most junior level employees,
right up to the boards of directors.
0:34:36.610,0:34:40.920
Often the terms would get
mashed around, you know,
0:34:40.930,0:34:45.840
often the understanding of what they
intended to have happen from some of
0:34:45.841,0:34:50.160
these experiments, I'll call
them in AI, was misunderstood.
0:34:50.780,0:34:55.160
And I came to appreciate both
a way that I could represent AI
0:34:57.700,0:35:02.560
to members of Congress, which I feel
like I had to do multiple times a week,
0:35:03.060,0:35:06.480
and how I could work
with non technologists,
0:35:07.081,0:35:11.840
citizens in any capacity to
feel comfortable engaging in the
0:35:11.841,0:35:14.960
conversations, even if they
were non experts, you know,
0:35:14.961,0:35:16.760
even if they were non technologists.
0:35:17.600,0:35:22.450
An example of this is that
I think that all of us can
0:35:22.451,0:35:26.650
benefit by thinking through where
we want circuit breakers to occur.
0:35:26.950,0:35:31.770
So as technologists, you know, we think
automation's good, automation continues
0:35:33.331,0:35:34.164
to expand,
0:35:34.550,0:35:39.200
and that linking automated
systems is also just nothing
0:35:39.380,0:35:43.840
but good. But, you know, the
canonical example is an automated car.
0:35:44.300,0:35:46.640
You know, as an automated
car rolls down the road,
0:35:47.060,0:35:49.920
it senses something that it
thinks may be a crosswalk,
0:35:50.260,0:35:53.160
and then it senses something,
maybe aside the crosswalk. well,
0:35:53.720,0:35:58.080
what level of confidence? Is it a person?
Is it a tumbleweed? Is it a shadow?
0:35:58.660,0:36:03.480
And then does the car
slow down? Does it stop?
0:36:03.510,0:36:05.520
Does it keep going?
Does it ask the driver?
0:36:05.850,0:36:09.000
Those are all a way of thinking
about circuit breakers. You know,
0:36:09.001,0:36:10.400
we have it right now where, you know,
0:36:10.401,0:36:12.880
my car might ask me to
jiggle the steering wheel.
0:36:13.460,0:36:17.520
We need to think through that as
technologists, in addition to,
0:36:17.540,0:36:20.560
as citizens about where we
want these to take place,
0:36:20.561,0:36:22.920
instead of just linking
automation and then,
0:36:23.160,0:36:27.000
and letting people then
revolt the danger about that.
0:36:27.001,0:36:30.280
And this is gonna be worth us all
being conscious of, as technologists.
0:36:30.420,0:36:34.760
We want people to be embracing
what we develop, right? Our
0:36:36.950,0:36:37.561
companies,
0:36:37.561,0:36:42.000
and I might even say our civilization
Western society is going to benefit
0:36:42.750,0:36:47.160
from us embracing this in many
cases, life changing technology.
0:36:47.700,0:36:49.920
And if we have people
somehow resisting it,
0:36:50.190,0:36:54.640
because we and this technology elite
somehow developed technology that killed
0:36:54.640,0:36:56.840
people, then it won't be trusted.
0:36:56.960,0:37:01.239
It won't be adopted and we
will be missing out perhaps,
0:37:01.580,0:37:04.640
relative to enemies of
western civilization.
0:37:05.040,0:37:08.520
We'll be missing out on the promise
that technology could bring.
0:37:08.870,0:37:11.520
[Rob] Yeah, I think it's a really interesting
question, how you can get people
0:37:13.041,0:37:17.200
to develop trust in a technology that
they maybe don't fully understand.
0:37:17.500,0:37:20.080
You know, I think that's
one of the, you know,
0:37:20.081,0:37:22.600
the enduring difficulties
of this kind of work.
0:37:22.900,0:37:23.733
[Eric] Indeed.
0:37:23.940,0:37:26.790
[Joey] Well, I think this is about
all the time we have. Eric,
0:37:26.860,0:37:30.190
it's been wonderful talking with you.
Thanks for sharing so much about your
0:37:32.311,0:37:33.830
company and the way you
all are approaching things.
0:37:34.210,0:37:36.390
[Eric] It's been a great conversation.
Thanks for having me.
0:37:36.739,0:37:39.190
[Joey] This has been another episode
of building better systems.