Latest / Agentic Conversations (formally mlops.community) / Building Better Data Teams // Leanne Fitzpatrick // Coffee Sessions #113

Transcript

Building Better Data Teams // Leanne Fitzpatrick // Coffee Sessions #113 · Agentic Conversations (formally mlops.community)

0:00
Hi, I'm Leon. I work at the financial times.
0:03
If I was gonna drink a coffee, it would be an oat milk flat
0:07
white. My man mihael.
0:11
How you doing dude? I'm doing well.
0:13
How are you? Demetrius, I am great.
0:16
I got a new shirt. It says Ohm Shanti, I am Ohm
0:19
Shanti in out of things right now.
0:22
Nice, nice. We might have to bleep that out
0:26
because Leann has been working so hard, not to cry.
0:30
This on this fellas bleep, we cannot have one podcast without
0:38
you mentioning them and we're just have to bleep that I
0:41
wouldn't be. It was just gonna be I did it in
0:43
the chat, okay I was being conscientious, it's hilarious.
0:49
Anyway, Leanne was awesome. I'm so like some of the answers
0:54
that she gave I got goosebumps because she was nailing it so
0:58
hard like the amount of You that she was bringing with these
1:01
answers. What did you think do?
1:03
And I just super biased what you're going to say right now.
1:06
But I thought it was incredible. What about you?
1:09
I thought that her points were just so on the Mark, I mean,
1:13
around, when she just basically started the entire conversation,
1:17
discussing structuring of a data science, organ, terms of make
1:22
starting to limitations instead of aspirations, right?
1:25
Constraints, instead of, like Northstar goals, I thought that
1:29
was such a A realistic take on the state of like we spent a lot
1:33
of time talking about data tooling, but we maybe spend not
1:36
as much time talking about just like data organizations and
1:39
efficiently running and organizing data teams.
1:42
I thought that was such a realistic take on what the state
1:46
of building data organizations look like today, right?
1:49
That while we being in the community, right?
1:55
Having these sessions were incredibly bullish on data and
1:58
ml. As a next Wave for the future
2:01
that for a lot of organizations globally, they still are trying
2:05
to justify their investments in these these ideas and these
2:08
tools in these projects and we're still at that nascent
2:13
stage. We're getting that buy-in is
2:16
something that's going to take time, right?
2:17
That it's going to take having some of these winds that come in
2:20
and trickle in. And and this these are the
2:23
actual friction points, the data teams deal with.
2:25
And I think she just did a great job of codifying that struggle
2:28
right at places She's worked just really good General pieces
2:33
of wisdom for any data team. Yeah, I loved the metaphors.
2:37
There were so many good metaphors and following up on
2:40
that idea. She was talking about how it's
2:42
like data is looked at as the little sibling in the
2:47
organization because one thing that hit home, so hard to me was
2:52
how Leon said data teams aren't trusted with how they use their
2:59
resources. Is sometimes you can get like
3:02
this. It just feels like they aren't
3:04
being trusted with the resources that they're given or the tasks
3:09
that they're given and how they go about meeting these goals.
3:13
And that was brilliant. The other piece that I wrote
3:18
down here was like, it's not clear how you articulate the
3:22
value or why you need a data scientist instead of in ml
3:26
engineer, or why you need a data engineer instead of machine.
3:29
Learning engineer or now, as they're called ml Ops engineers.
3:33
And so, whereas with like software Engineers, it's very
3:38
clear to the business what they do and how their value is
3:42
articulated. But in the data Spectrum, it's
3:45
not as clear yet. And she boiled it down to were
3:49
really new still like you said, and it's just that maturity
3:52
needs to happen. And we need to have this time
3:57
where things settle into place. So that's it.
4:02
Let's get into the actual conversation and hopefully
4:05
everyone out. There will learn a new vocab
4:07
word waffling. Mihael.
4:10
Learned it today. That was really, really good,
4:14
and the learn everything on this podcast.
4:17
Everything it's not just from you, learn.
4:19
You learn vocabulary to, she's laying British slang.
4:23
There you go. The other thing that I want to
4:25
mention is that if you have any ideas on how we can make this,
4:30
Show better and how we can bring more value to you the listener.
4:34
Please just write us. Let us know, reach out to us on
4:37
slack. What would you like to see us
4:40
doing more of less of better if you want less of the the
4:46
bullshitting around in the beginning of the episodes?
4:49
You want more of it. I don't know.
4:50
You tell us we're doing this for you all.
4:53
Hopefully bringing you some value at the end of the day or a
4:57
good way to pass a little bit of your time while learning.
4:59
Thing, that's it. Let's jump into this episode
5:02
with in the end, and I hope you all enjoy.
5:04
We also prioritize any email, and not any actual handwritten
5:10
letters over overwritten emails. Those are the ones.
5:13
So, anyone that goes to the trouble of writing us, an actual
5:16
piece of mail will get will get priority in terms of, we are
5:22
definitely going to start doing whatever you suggest and if it
5:25
is literally, if it's calligraphic by Nature.
5:29
ER it's like you can buy come on and be the next co-host.
5:33
Yeah you're it. You're it.
5:35
There we go. Let's get into this actual
5:37
conversation. Now, see y'all if you want we
5:44
can start it off with with either seaweed or the guitars
5:51
your choice. So we can start with seaweed
5:55
because she's gone for her for a bath because she had surgery.
6:01
So who is seaweed tell us about? Yeah, that's a really weird way
6:05
for me to kick off a series, right?
6:06
Seaweed is my tree like a dog that I rescued me, my husband.
6:10
Rescued in Austin Texas. She is a she's a mutt dog.
6:15
So she looks like a dingo but she's technically a Texas.
6:18
He'll across with an Australian Stumpy cattle.
6:20
So she is sadly missing a leg a womb and a tail.
6:25
So you know she's having a bit of a tough life but gives a lot
6:28
of love for that. So she's a champ and I heard she
6:32
has flown more times in business class than I have because she
6:39
flew from what was it from the u.s. to the UK in business class
6:44
if I get really good. Yeah, which I've even what I
6:47
mean. So height of the pandemic, my
6:50
husband unfortunately got stuck in the UK.
6:53
We were both on visas working in the u.s.
6:55
It's a really long story. Now, I'm going to try and keep
6:57
it short. I was stuck in the the u.s. in
7:01
Austin having a whale of a Time by myself by the way.
7:03
And seaweed. We were chilling out.
7:05
Yeah. Yeah.
7:06
So we were stuck with seaweed over in Austin, but my husband
7:10
was unfortunate stuck over here in the UK mr.
7:12
The pandemic and we needed to fly back after five months by
7:17
myself living in Austin as a will.
7:19
At the time I had it was time to call it quits.
7:23
And so we had to get back shipping, a dog into the UK is a
7:28
real Multi not going to go into all the details here because,
7:32
you know, I might have some people coming back on my back
7:36
but this short, long story short, is that series got to fly
7:39
in the business class cabin. It's my one luxury treat to
7:42
myself. Was that instead of flying
7:44
economy was like five hundred dollars and business was 650.
7:48
Because of because they obviously and because obviously
7:52
flight to a complete disaster at that time, between the UK and
7:55
the US. And so, yeah, we bumped
7:58
ourselves and we had a lovely time.
7:59
I drank lots of champagne, see we didn't, but she did sleep
8:03
hallway and she checked out, she checked all the boxes.
8:07
I imagine you got all her papers in and nobody's going to come
8:11
asking you for any official papers.
8:14
It's dog. She's totally legally allowed.
8:19
I had heard about that because a friend of mine just moved to the
8:21
UK, like, a few months before the pandemic, and he brought his
8:24
two cats. And so he was like, who, Who Was
8:27
Bitten of a nightmare to get him over so much, so that.
8:29
That he doesn't want to go back to the US just in case he ever
8:32
has to go back and have to like, go through it again.
8:35
So anyway, we're not here to talk about pets as cute and
8:38
lovable as they are. We've got a lot to jump into.
8:42
I would love to start with this idea.
8:45
I mean, There's so many places that we can go but I think the
8:50
most pressing on my mind is something that comes up time and
8:54
time again. And I want to hear your wisdom
8:56
on. How can we architect data teams?
8:59
What are the structures that you have seen that have been winning
9:04
formulas? And how does that look like,
9:07
that's a great question. I can shed some wisdom whether
9:11
or not it's good. Wisdom is going to be a
9:12
different question, but let's go for it.
9:15
So I think it's such a difficult question because It starts with
9:18
the fundamental fact that you've got to build for the businesses
9:21
in front of you. That's the first issue, right?
9:24
And, you know, as much as we like in certain commercial
9:28
focused areas, we can turn around and say, right.
9:31
I now need to increase my head count by 50 and, you know, the
9:34
whole business is for scale size.
9:37
There is maybe a thousand people. 50 is a heck of a lot of
9:40
new headcount for one department, so you can have all
9:44
these kind of architectures and these designs and these ideals
9:47
that you Want. But you're always going to be up
9:49
against your commerciality. You resourcing your op X from
9:52
your head count. Or maybe if you're doing it
9:54
through contractors, your capex budget.
9:57
And so that's the first thing that you have to have in mind as
10:00
a kind of as a data leader, that's thinking about.
10:02
Okay, where do I want to put my bets?
10:04
Where am I going to hedge my bets?
10:05
When I architect my team. I rather would start with my
10:08
constraints rather than my aspirations and that sounds very
10:12
sad. But I think it's because that's
10:14
all I'm more of a realist with a little bit of optimism throw.
10:17
And, and if I know those constraints and those caveats,
10:22
then I can start designing for the right team that I have.
10:26
So practical example, at the Ft, we have 11 data, scientists,
10:32
we're roughly Financial Times, is roughly 2,000 people across
10:36
with, for those that don't read the financial times, really huge
10:39
journalism, media business. I'm not here to push the Ft, by
10:41
the way. I'm just giving some contacts
10:43
and so, you know, data science is a large part of what we do,
10:47
but the Purpose of data science was more to drive a lot in the
10:50
beginning, a lot of our subscriptions and kind of the
10:53
our subscription strategy and optimizing a lot of our
10:57
decision-making need to see and B2B decision driving behaviors,
11:02
right? So the hence why the size of our
11:04
team makes sense? So when I came in I didn't
11:07
really want to grow with just that was just 11 days to
11:10
scientist. I didn't really want to grow my
11:12
data science team. What I want to grow as my
11:15
engineering resource that's actually going to make that
11:17
happen. Right?
11:18
And so we've been doing a great job working collaborating with
11:21
our bi business intelligence Engineers to be able to do a lot
11:26
of batch modeling through our database layer in bigquery.
11:30
But we're, we're getting really where we will have massive
11:34
issues in terms of our growth strategies.
11:36
The fact that we can't do anything in real time.
11:38
A lot of what we do is in batch and we've done some stuff in
11:42
kind of low latency near real-time deployments.
11:45
And when I say real time, sorry, I mean dear.
11:47
Real-time. So for example you land on our
11:50
barrier page where ask you to make a payment to stop reading
11:54
and we can do some personalization there in low
11:57
latency that's what stuff with Million Miles Away.
12:01
So a new we need to talk get some engineering resource to the
12:03
approach. He took their was going to data
12:06
platform where we have a huge data engineering team.
12:10
And we work really closely, they're in different functions
12:13
of business. We've got what underneath the
12:14
CTO and underneath the chief Analytics.
12:19
Luckily, those two of kind of data platform, team always
12:22
worked very closely with out analytics team, so it wasn't a
12:26
stretch for me to ask for to two new heads, essentially or two
12:29
new resource to then be embedded in data science as soon as I
12:35
joined and luckily, there was a good opportunity.
12:38
We had lots of hiring and and you headcount availability so
12:42
they were happy to put two new head count towards that.
12:45
So, that's how I've done that here, right, I've only been at
12:47
this EFT for 12 months, I'll swear it's been a real pain.
12:51
Like, you know, again, I find that leading data science teams,
12:55
the initial reaction for the data science in, from the
12:59
outside businesses, you just need data scientists.
13:02
And it was like, well, there's that whole piece around as we
13:05
talk about all the time, the whole data acquisition your
13:08
whole data layer, and how you management team that you're like
13:12
the and the impacts of your architectural stack and your
13:15
choices that that will have on the data science team.
13:17
Long run. So for practical example, again,
13:19
at the Ft, we're doing, we've got a little bit of a magpie
13:22
architecture. We've got some stuff in a bit
13:24
of. U.s., some stuff in gcp, again,
13:27
comes when we get into kind of low latency deployment.
13:30
I'm trying to wrap my head around the strategy around that
13:32
to ease some of that friction. So if you've got your data
13:36
engineering resources that in a different bit and waffling, I'm
13:39
sorry. I'm gonna, I'm gonna stop in a
13:41
second, but if you got your data Engineering in a different
13:45
technical, like an architectural decisions.
13:47
Place to your analytical and data science resource.
13:50
You're going to have these architectural playoffs, that a
13:53
really difficult so that that's another constraint thing and
13:57
then I'll talk about it more. I don't think Miguel knows what
14:01
waffling means. I don't think he is your Pian
14:05
thing. Yeah it's going on and on and
14:08
I'm just gibbering along and I'm doing a monotone and I'm really
14:12
sorry. I'm gonna I'm gonna try and like
14:16
well that's that's That's a really cool tidbit.
14:21
Yeah, I thought you might not know that one because I learned
14:23
that one fairly recently and I thought it was amazing.
14:26
I just always think about Waffle machines.
14:28
Whenever I hear someone say that and like that was, like,
14:30
literally the thing that popped into my head, I was like, what
14:33
it waffling. Like, are you making something
14:35
like, what is making a monotone? That's awesome.
14:42
That's awesome. So there's something that you
14:44
said that we in that I thought was particularly interesting.
14:47
And which is this idea that maybe even a bit
14:51
counterintuitive. This idea that instead of
14:54
designing or thinking about teams data science, teams in
14:57
terms of aspirational goals, you think, maybe you start with the
15:00
constraints, right? The things that are restricting,
15:02
you maybe is another way to put it and this is interesting too
15:06
because a lot of the times one thing that we're seeing, right?
15:11
A large explosion in recent years of like data tools and
15:15
different companies and organizations trying to put out
15:17
different Was that resolved at least purportedly resolved data
15:21
problems for practitioners. But one issue that a lot of
15:25
people claim is that there's just like organizational
15:28
deficiencies, right? That actually the biggest
15:30
problems aren't always, there were missing another another
15:35
tool to deploy to kubernetes but we're really missing is just we
15:39
have these organizational friction points that make it
15:42
difficult to get things done and data teams.
15:45
Right. And so when you mentioned And
15:47
that you've kind of design with constraints, starting with
15:50
constraints, rather than, like Northstar Vision goals, right?
15:53
Like, Hey, we're going to do the latest and greatest things and
15:55
data science, is that a function of some of these like
15:59
organizational constraints or, you know, because this is
16:02
something that that's I've also worked in engineering teams
16:04
where we don't start with constraints.
16:05
Like we rarely think just purely in terms of constraints.
16:08
I wonder how much of this is just like a function of Of
16:12
being, you know, a data organization of being a data
16:14
team that has to deal with this kind of stuff.
16:17
Think it's a really good point that you taught him.
16:19
I had a whole like I hadn't thought of it that way before
16:22
that actually date has always kind of been the the the little
16:27
the little brother or a little sister to the rest of
16:29
Technology, right? And actually, or maybe the
16:33
middle brother or a middle sister, I'm an only child.
16:36
So I really liked having to make some guesses here.
16:39
You know, always drew the Short Straw Orbit.
16:42
And I think it definitely is that that I think it's really
16:45
easy to articulate in the business why you need an
16:47
engineer or a product manager or project manager?
16:51
And it's still today is really difficult to explain.
16:54
Why you need an ml engineer versus what we're seeing.
16:58
Now, in ml Ops engineer versus a data scientist, versus a date to
17:04
translation layer, maybe with some, like, change management
17:07
and delivery management within that all with data expertise.
17:11
He's though that is really still not very transparent for many
17:16
businesses and even with businesses that have like the
17:20
money supermarkets in the UK, they've had really fantastic,
17:23
cdo's, Chief day after officers alongside Chief analytics,
17:26
officers, you still see there's those underneath the hood.
17:30
I think there's still a lot like friction pulling of like what to
17:34
these roles were these roles doing, but an engineer is still
17:38
an engineer that they're not any new don't and answer.
17:41
What we see at the moment in my world.
17:43
It's like when we talk about, can we have some data
17:45
engineering resources? Not that data.
17:49
The thought of layer like oh we can't trust the data team
17:52
though, to determine what they need to do with that engineering
17:54
resource technology. Need to determine that and I
17:58
think it's just a function of out in a sense like that is
18:00
literally by definition where we are in our journey is that we
18:04
can talk all we want about how data science has been happening
18:07
for you know 50 years and actually you know it's just just
18:11
all Statistics in Disguise. But the reality, the reality is
18:14
in businesses. At most this thing has been
18:18
happening for 10 and if you're in, if you're lucky, maybe it's
18:22
been happening in your organization 45 and I would say
18:25
that that puts you quite mature if it has.
18:27
So I think I think you're completely right.
18:29
That is it it's a maturity piece coupled with a real lack of
18:33
still real like at the board level of understanding of what
18:37
it takes to be successful and that I in my experience, I found
18:43
that the board or the key decision maker help decision
18:47
makers that kind of Seaboard or except board level are willing
18:52
to compromise on the data function more than they're
18:56
willing to compromise on the technology function and I mean
18:59
that in no harm to any of those decision makers.
19:01
But more it's just a reality of where we are as a field and that
19:06
were real growth part to your strategy rather than I kind of a
19:12
both keeping the lights on and keeping the business safe and
19:16
gross data. Still not seen in that way.
19:19
I was having a really good way to articulate that.
19:22
I mean, Wow, there's there's so many bulbs that I can think of
19:26
to this. I mean for starters.
19:29
Yeah, I agree with this right. I mean, that it's there seems to
19:32
be this, this tension here, that this conflict and a lot of times
19:36
data is a little bit on the fringes, at least maybe because
19:39
of its naissance and the fact that has been around for too
19:41
long. But we as you know, we're data
19:45
practitioners, we're here on the MLS Community podcast like we
19:47
clearly believe this is like the future right?
19:49
That this has to be the way the world works that we're not just
19:52
this add-on. That makes something a little
19:55
bit better. That actually we will be the
19:56
core of what makes future organizations modern and able to
20:01
do things, right? So clearly we buy into this
20:04
hypothesis that it's not just a little thing.
20:07
You add on as a bottom line on your, you know, expenditures or
20:11
whatever. We really believe that there's
20:13
And fundamentally different about what makes data science
20:17
and data data scientist and ml, really what it is.
20:21
So how do we expedite this process?
20:24
I mean, how do we get to get to the point where people, these
20:27
stakeholders except that, this isn't something that you just
20:32
you do like a little trial project and then if that goes,
20:35
well, great, otherwise Cole, get rid of the data team or like
20:39
give less resources. How do you just how do we get to
20:41
the point where there's more buy-in and And were able to move
20:44
forward. So that people stop.
20:45
They like trust us. You know what I mean?
20:47
That this is what we're doing. So for me it comes, good
20:51
question and I don't think I have all the answers that are
20:53
completely extensible. But for me, it comes down to two
20:56
things on that specific question.
20:58
The first is like, you're either want to me in that context.
21:02
How do you get that buy-in around?
21:04
A to your either, the type of organization where your Roi is
21:08
data itself, white by what I mean by that is you either
21:11
selling data as a product, or you are selling a capability
21:15
that supports data people to do their jobs.
21:18
So, You know, you all aw Ice Age Maker or data bricks or answer
21:24
your doom, you're some kind of data platform service provider
21:28
or your Roi itself is like paid for hit.
21:32
Essentially, a version of data is a product but paid for hits
21:35
per machine learning model that's run, right?
21:38
So that you're that type of organization where your Roi,
21:41
your bread and butter is the data services you offer or your
21:46
the other type of organization which is that Each was a part of
21:50
your organization that combines with many other things to help
21:54
your strategy. So in my situation, you know,
21:58
fewer organized, I've worked for a credit bureau.
22:00
I've been telcos and now I'm in journalism that I will media.
22:05
That was, that's the kind of areas that I would say data, is
22:08
part of it. I've also built a start-up
22:10
that's been more on that detail as a product side as well.
22:14
I think, when you're on that date was a product Side by very
22:18
nature. ER, if you suddenly no longer
22:21
support your data capability, you can't keep your lights on
22:25
which goes back to that three thing about like keeping the
22:28
lights on making sure your organization is safe and
22:31
growing. So you can't physically do that,
22:34
unless you invest properly and you trust the person that you've
22:38
put in charge of the data strategy, which is almost, it's
22:42
got to be so close to your CEO, that they're like, their hand in
22:45
hand, because that's the whole nature of your business.
22:48
I think in other organizations now you've got a different Focus
22:53
which is more around. How do I operationalize?
22:56
The business change. I want to see through the data
22:59
services. I'm enabling.
23:01
So they're taking something simple like a churn or a
23:04
propensity like propensity to churn model.
23:06
Look for example, always a good one to rely on.
23:09
Right. But, you know, people still were
23:12
like, oh, we can turn that thing off because most of the times,
23:15
it might like, drive one or two chemo.
23:18
Think campaigns, but it's not actually inherently changed the
23:22
whole of the business. The way the business thinks
23:26
about actually it's whole commercial marketing strategy
23:30
and I think that's where we're, that's the next opportunity.
23:33
That's where I'm very passionate about.
23:35
It's all well and good. Put in like these but in a churn
23:38
propensity model and production, but it's then about how am I not
23:42
going to just get that to feed that one or two or ten different
23:46
marketing campaigns if that's what it needs to do?
23:48
Do but how am I going to actually change my business?
23:51
My business behavior in those teams to rely more on the data
23:56
services that I can enable them to the they're freed up to do,
23:59
things smarter to help things and there's all that typical
24:02
like data organization change. That's really hard in the second
24:05
type of business and you feel like that the lot of
24:11
misunderstanding comes from let's say the marketing maybe
24:14
it's the demand gen lead who doesn't understand.
24:19
The machine learning can be leveraged to help them do their
24:21
job. And I don't think that's on them
24:25
to necessarily understand that, right?
24:27
I definitely think that's that in-between layer.
24:30
So something, we definitely got passionate about the ft's.
24:34
We ran rolled out a whole day to democratization program.
24:37
And this is wider in the data science, sits within analytics
24:40
at the financial times. And we have these different
24:43
embedded folks, who are there. To essentially, literally be
24:48
embedded by the to, with the tooling and the technology.
24:51
So I'll customer data platform. We Individual who is there to be
24:55
an expert, in how that customer data, platform works for the b2c
24:59
marketing team, and then to be that point of contact and
25:03
liaison between say, data science.
25:06
And, and that commercial and we hear this a lot.
25:09
But the differences is that, I think in this scenario that
25:12
person is fully embedded into that kind of CR G and the the
25:17
toolkit at the same time. Whilst also having been the
25:22
physically like Reporting line is in within the data and
25:26
analytics world. So it's a it's difficult.
25:29
I don't think there's I think it's up again about going for
25:31
the business, but you do, I guess my answer to that is,
25:34
there's got to be something in between the team that's
25:38
responsible for building and servicing the capabilities of
25:42
the model and the team that's actually consuming it.
25:46
And often times, we rely on the data side of the Builder and
25:50
then also the servicer. But there's also a middle layer.
25:53
She's like the embedded. This is the how we're going to
25:56
use and champion this capability.
25:59
What type of profile is that middle layer?
26:02
What what kind of person is that are they very technical?
26:05
Are they very business? Do they have a lot of business
26:09
Acumen? I guess.
26:11
What does that look like? Because I imagine they have to.
26:15
Yeah, they have to understand the translation which they
26:20
probably have to be technical, but they also have to understand
26:23
the business side. So it's like almost a unicorn in
26:26
a way that's gonna say you gonna say that these are on Union
26:29
columns or just create a new row.
26:33
I don't think they are unicorns because I see enough of them.
26:35
And I see enough of these people like my, you know, different
26:40
people live, either coach torment, or even managed.
26:44
Like, I see so many of the of these people that come from, not
26:50
a data science background. As I say, I coach tormented
26:52
them, because a lot of the people, I've always managed been
26:55
data scientists, teacher, engineers, and data
26:58
infrastructure folks, and my passion is very much an ml or
27:02
data science space but more widely in the analytics
27:05
Community. You see these people who have
27:07
come from that quite technical background, maybe did a lot of
27:09
like data visualization and have this gift.
27:13
To liaise with stakeholders. And I think a lot of data
27:15
scientists can do that as well, but they have a natural
27:19
storytelling ability and they get to that point in their
27:23
careers are, you know, I definitely feel like sometimes
27:25
I've got that point where I don't want to be that
27:28
technically hands on any more, I want to do something else.
27:30
I like working with people and it's definitely, that way you
27:33
can play between that, like a really good product manager,
27:36
who's come from, maybe an engineering discipline or a
27:39
data. Discipline who then is always
27:41
had that gift. It to listen to user stories or
27:45
to translate them into user stories and then figure out that
27:49
that that kind of mediation. That's the type of person and I
27:52
think again it comes back to the immaturity I think in our field
27:56
that we just haven't seen enough of those folks explode yet
27:59
because a lot of them don't see those career paths for them yet
28:02
and they're like, they're neat. She got to kind of develop it
28:04
yourself. It's not like I'm going to go
28:06
out and advertise for a new. I don't know like embedded
28:09
consultation, conduit between technical Oil and commercial
28:14
stuff. Please maybe that's what we can
28:17
call it. That that sounds like a really
28:19
good description. That's I'm sure a lot of you get
28:21
a lot of job job applications with that one.
28:25
Pretty simple title, right? Yeah.
28:27
Very that's all yeah. Was that the description was
28:30
that the title name? I don't know what.
28:32
This definitely wasn't the tile. No, I'm not.
28:37
I'm not good at describing hence, why, you know?
28:39
And the Lots never even came to me when I was like this.
28:42
Finally, in terms of the That we've been through the last
28:44
picture. Hey, I'm Vishnu.
28:49
I'm a data scientist at first hand, and I definitely think
28:53
that you should subscribe to the ml Ops coffee sessions podcast.
28:56
It's the best podcast out there to stay on top of what ml Ops
29:00
actually is to talk to the truth, thought leaders in the
29:02
space. And oh, by the way, dimitrios is
29:05
absolutely hilarious. What a weird guy, you should
29:08
definitely subscribe to the podcast.
29:10
I'd love to maybe take a step back.
29:13
And I'm fascinated because you know you're you lead data
29:17
science at a publication right? Of infantile times.
29:20
It's like a huge. Huge will publication.
29:23
What is data science work? Look like on a day-to-day there
29:27
and what kind of problems do you guys deal with and what are the
29:29
main challenges that you guys face?
29:31
What is what do you what do you do actually?
29:34
Yes. Oh I'll team was spun out more
29:37
to support our product and Tech capability.
29:40
So what I mean by that is our digital first I guess.
29:43
Gee. I'm so we're very digital
29:46
centered business. So whilst we still have the
29:50
newspaper, which we call like the pink.
29:51
We also have a number of different, not just the app, the
29:55
website itself and our digital subscription model, but also a
29:59
variety of other Fu products under that.
30:01
So we have events ft live. We have Lefty specialist, so
30:06
data science, is there to support a whole Realm of the
30:09
like the group capabilities. What that means in principle Is
30:13
that we can be doing anything from doing a customer
30:17
segmentation on how b2c user base in terms of how to actually
30:23
upsell cross-sell, whether or not, we want to just kind of
30:27
leave customers alone as well, right?
30:29
Through to something like this quarter, we're building an
30:32
article vectorization or improving our article
30:35
vectorization, that then feeds in to I'm assuming that most
30:40
people know what article vectorization would mean, but
30:42
effectively. Deriving.
30:44
Okay, good. Okay.
30:45
I'm not gonna go into the details and so and then that
30:48
feeds into sort of our recommendation models that we
30:50
drive. So right through from a B2B like
30:54
topic recommendation models, that might be an org level or
30:57
right through to say, when somebody lands on our barrier
31:00
page, being able to determine which article they see next, for
31:03
example, right through to newsletters as well.
31:06
So, those are the types of things that kind of gives you a
31:09
flavor. We're also responsible for a lot
31:11
of, like, metric building here at the f.
31:13
Which is a bit of a new flavor for me in terms of data science.
31:16
So we responsible for things like our engagement modeling.
31:19
So how engaged readers with us, as well as like lifetime value
31:22
modeling. So we've got a very broad
31:26
spectrum for a quite what. I would consider a very small
31:30
team in that purple Spectrum. So I'm wondering about the point
31:38
that you made earlier when it came to real time and how you
31:41
were talking about, you're trying to get to To real-time it
31:44
is something. I mean Michael and I have talked
31:46
about this various times on this very podcast about real-time and
31:50
the Mystique that is around it and how it seems like many
31:55
companies want to get there. But whether or not it is useful
32:00
or the reasons for getting there.
32:01
My sometimes not be so clear. Can you dive into why you feel
32:06
like you need to get to real time and what that pushes for?
32:10
Yes. Oh I love this question because
32:13
it's Something I'm very passionate about us again,
32:15
building for the business. You see in front of you and, you
32:18
know, don't if you if you're in a heavily regulated feel don't
32:22
feel like you have to push something out the door within 2
32:24
hours, you know? Take the take the week door,
32:26
however long, it takes to make sure that you've actually
32:28
verified that model. And So within that context where
32:33
we are at the moment is majority of our models, are running
32:37
batch. But the way that we I can go in
32:41
a second, I'm going to do another one.
32:43
Flea market monologue because I can go into some details about
32:47
that in a second. If you like from a technical
32:49
architecture for him but where we have drive things that need
32:53
to run in low latency. So for example, if you as a user
32:57
sign up for a newsletter, we don't want to be sending you a
33:01
newsletter, you already signed up for, right?
33:03
So that's a simple filtering that we can do, but the
33:07
architecture of doing that filtering is really complex in
33:10
our current state of play. And what it means is that
33:14
actually wear those data assets, it's not transparent at all to
33:18
the data scientists or the people curating and looking
33:21
after the model. So like for me if I was then
33:24
told tomorrow go and you know, pull this repo down, do some
33:30
fixes to this newsletter recommendation there's a huge
33:33
chunk of that model. That's not exposed to me as a
33:37
model builder, which is all this additional layer of filtering
33:40
that's happening. That's a problem for us.
33:42
And that's the bit that It's happening is like the there's
33:45
two different things I want to talk about with near real-time,
33:47
right? There's filtering of data and
33:49
nura time and then there's actually running a model that
33:52
needs to enact in low latency. So I'm just touching on the
33:56
reasons behind the filtering for now that's difficult for us to
34:00
see. It's a little bit and it's also
34:02
a little bit. I mean this in the best way
34:04
possible to my entire of the Ft team, we build brilliant things
34:07
that get me wrong, but it's also a little bit Frankenstein
34:10
monster, and I don't think anyone will kind of hate me for
34:12
that. Saying now so the the push for
34:17
when I talk about near-real-time is more to come up with a
34:20
transparent architecture that it's simple for us to see, like
34:25
have this generic type of way of deploying that we can see all of
34:29
Frankenstein's monster within like essentially within wrong
34:32
rapper. Let's say so we can actually see
34:36
what's happening in in the Box on the other side where I talk
34:38
about with low latency models. So, one example is that is last
34:42
year we built A personalization model to enact on the barrier
34:46
page. So when a user lands on our
34:49
barrier page, we want to extract what data, we they've consent to
34:54
for any personalization at that point and then plug that into
34:58
our, as inputs to the model and then derive, the most
35:01
appropriate barrier page for them to see almost
35:03
instantaneously that we physically can't do with our
35:07
current architecture. So that's warm practical
35:09
example. What we want to do, I wouldn't
35:11
say that. Like, we're there for gunning.
35:15
All, you know, we're hedging all of our bets on making low
35:17
latency happen because that's just one example.
35:20
But, when you get into say we ever wanted to do anything
35:22
around, we have a third, lots of, third parties, running
35:25
things like a fraud detection. If we wanted to do things like
35:28
that, with in data science, that's where we would need to up
35:32
our panting so to speak. So it's more to do with the
35:35
nature of where data science has come from the Ft today and the
35:39
what we've been delivering to the business and kind of the
35:42
Growth stage, we need to untap some capabilities but also just
35:47
for our own transparency, we need to have some capability.
35:51
I've got a little side note or bit of a tangent question, but
35:56
as you mentioned like, there are third parties doing stuff with
35:59
fraud. Do you feel like the that's kind
36:04
of the future of companies, like, breaking up all of these
36:07
little use cases and sending them out to third parties, or do
36:12
you feel like It's moving more towards.
36:15
Okay. Now, we've gotten from 0 to 1,
36:18
we want to go from 1 to 10. We're going to bring that
36:20
in-house and it's going to be our thing that we're going to
36:22
throw a bunch of data scientists and resources at.
36:25
I think it comes back to all reward.
36:27
I surely right. If he specially if you're
36:30
working for a commercial business and not kind of an NGO
36:34
or something like that, if I think it comes back to, yeah, if
36:40
you want to spend something it up quickly and you have the
36:42
resources, In house all ready to test that out.
36:45
So like a quick fraud model. That do that.
36:48
But obviously from a payments perspective because we take
36:51
payments for our digital subscriptions, we needed fraud
36:54
straightaway data science, definitely wouldn't have existed
36:57
at that point. So we went to a fraud vendor
37:00
took them and I will quite happy with their capabilities plus
37:03
they're specialists in that same with someone a lot of our LP
37:06
capabilities on our actual content with those are a lot of
37:09
that is driven by a third party and I honestly I'm happy with
37:15
that because unless I wanted to suddenly spin up her Team of 60
37:18
and LP Engineers, I'm probably not going to be out there
37:22
service. Not I'm not trying to Champion
37:23
not going to also say any names because and no one knows it.
37:26
But no trying to Champion that way, but I think it's again
37:29
organization in front of you. I think, definitely if you're
37:31
doing the data as a product capability, you want to be
37:34
building it in house because what's your IP that point if
37:38
you're, if you're, if you're supporting a business to do to
37:42
put, you know, information and content out there or you're my
37:47
As place of work, talk talk. You're providing, you know,
37:50
internet service to people, your priority isn't to say all but
37:54
this is our now you know this is our secret sauce to make things
37:57
happen. We're going to use the resource
38:00
where we can to make the right decisions.
38:03
Yeah, so good. And I mean, I wonder at some
38:06
point the thing that I think about is Do the data scientists
38:12
or the data practitioners in the company.
38:14
Do they ever feel like we're giving away too much?
38:17
Like they're going to automate me out of this?
38:19
Or is there always something next that they can do and you're
38:23
just augmenting their abilities with these third-party tools.
38:27
It's definitely an augmentation, right.
38:29
It's got to be an augmentation because again, it goes back to
38:32
that business understanding that.
38:34
If you're in that business, you have the domain knowledge to how
38:36
to know how to stitch the things together.
38:39
So again, if we go back to that for Solution, right.
38:42
That fraud Solutions not going to necessarily know the ins and
38:44
outs of exactly how you doing your payments.
38:47
Exactly all the other constituent pieces of kind of,
38:50
you know, how will your, your payment platform physically
38:54
works end-to-end? It's just one part in a small
38:57
Niche, which is the job of that capability is to say it's this
39:02
credit card, a potential fraud or not, and you have the data
39:07
and the capability to actually do that.
39:09
If we wanted to do that, we'd have to to go away and invest a
39:12
lot in terms of accessing that type of data, Etc.
39:17
So it just I think I think it's those things that sit again in
39:21
the in this second type of business where you're not a date
39:24
or as a product type thing where you're actually just building a
39:28
business that's going to grow for some, you know, whatever you
39:34
trying to achieve. I don't want to keep talking
39:36
about the financial services. Not only my brain today don't
39:40
but you know whether or not your, you know, your car
39:44
manufacturer for sample, right? You're not going to be there
39:48
thinking, okay, I've got to have this, got to be my own USP about
39:53
my special secret, maybe if your Tesla but but my secret Source
39:58
about how I like architect at this, the the speedometer and
40:02
with all the leg data capability in that we've got an automatic
40:05
speedometer that can shut off based on you know out underlying
40:09
algorithm. I think you see a lot of car
40:10
manufacturers actually advertising, right that right
40:13
now. Oh golf clubs actually golf
40:14
clubs out AI driven flat face. And I'm like, what does that
40:18
what does that mean? But I don't think any of those
40:21
gold clock manufacturers are going to be like but we had our
40:23
own data scientists working on this.
40:25
Probably have some data scientists who are using some
40:28
kind of third party or additional tool that serving
40:32
other Golf Club manufacturers. That's what I would think but
40:36
maybe not and I think I'm probably doing a disservice
40:39
because my my personal opinion, it should be every business,
40:42
should be building. Yeah, huge.
40:44
Scalable data, science capabilities.
40:46
He's and doing everything in-house, but where's the like,
40:51
where's the value in that for you as an organization?
40:55
And there are there are situations where you're going to
40:58
be better off with a third party.
41:00
Whose purpose in life is to do that in the same within a no
41:04
one's going to build Gmail for you too in your own business,
41:08
right? Well you're not going to build
41:10
juror if that's how you want to do project management.
41:12
There's certain things you're just going to say we're going to
41:15
buy this thing in. Yeah, under certain things that
41:17
you're going to say, okay? And I think we're just seeing
41:19
that same contention in the data space actually have two points
41:24
on this. I think because it raises some
41:26
really interesting thoughts. One is this idea of you don't
41:31
want to just you don't have to reinvent the wheel.
41:33
Basically this is this is kind of it's a common principle
41:37
actually. Write it back in the days.
41:39
Historical but historical lesson thing, you know what, when
41:42
Netflix was getting started? They sort of had the option of
41:46
Do we do we spin up our own cloud systems?
41:49
Our own the data providers or do we just piggyback off this new
41:53
thing called AWS? Right?
41:55
And they very proactively made it this principle of what they
41:59
called avoiding undifferentiated.
42:01
Heavy lifting right which is this idea that you shouldn't be
42:04
spending any time building stuff that doesn't isn't really
42:07
pushing forward your business your core business, right?
42:09
And so for them, having a bunch of Linux boxes in a data center
42:14
somewhere. At that point wasn't core to the
42:16
The business of trying to get people movies, right and get
42:19
people, good efficient movie system that you could rent
42:23
quickly and whatnot. So they opted against that and
42:25
that end up being like good, strategic decision.
42:27
Which I think is exactly what you're getting at here is for a
42:29
lot of data teams. If you can get away with it and
42:32
it's not part of your core business than a, why should you
42:35
even if we're engineers? And we love architecting things
42:38
and we love to have complete control over everything.
42:40
You don't really have to do that if you can get away with it and
42:42
it probably isn't the best business decision even if it'll
42:45
make some of the Engineers or the data scientist less happy,
42:49
you know, because they're not gonna have the full control,
42:52
which I think is one idea in the second thing is, I think for a
42:54
lot of vendors, they recognize their, their, their pitched
42:57
shouldn't be about replacing. The pitch should always be about
43:00
augmenting, right? You're fundamentally trying to
43:02
sell to data science or at least they still be your first
43:05
Champions. So if you come in and you say,
43:08
hey, we're gonna get rid of you entirely, that's not going to
43:10
make anyone happy. And it probably isn't even
43:13
possible this point given the tech, but I think any Business.
43:17
That knows what it's doing. Their strategy will always be
43:20
like, yeah, we're just going to make you better and faster what
43:22
you're already doing which which will you'll never lose people,
43:25
you know. No one will ever complain for
43:27
making their job easier. It's really interesting.
43:30
You ready to love this as well because it's made me question.
43:32
My own mentality quite a lot because when it comes to
43:36
actually like something I would call like a capability.
43:39
IE, only said about that fraud like that fraud detection piece,
43:43
that's just serving our payment and Options team, it does that
43:49
that's fine. But when I get vendors coming in
43:51
to me that are like okay so you're trying to solve this low
43:54
latency and your end-to-end kind of machine learning flow.
43:58
Let's say but not ml flow but their whole end-to-end
44:01
operational pipeline for both development through to
44:03
deployment through to maintenance.
44:06
I'm wanting to pick and choose how we do that in-house.
44:10
Because what I want to do and I use this terrible analogy, but
44:14
hopefully will help some people. See my team were really set in a
44:19
certain way of working and really works for them and it's
44:22
really efficient and a lot of these vendors that are out
44:26
there. Right now mean that I'd have to
44:27
take my team. They would have to have a
44:29
massive step change in terms of build it, using a new
44:33
development environment, the new deployment pipelines everything,
44:37
and maybe also have to do a lot more hacking like the math.
44:41
Do use some see Ally interfaces and things that they're just
44:44
like I've got a lot of like upskill Not just a migration but
44:47
now an upscale time as a big costing for me on a small team.
44:52
And so what I want to do is I want to say right?
44:54
We've probably got a bit of like a Ford Fiesta, that's a tough
44:57
but like, you know, a Ford Fiesta or gets me from A, to B.
45:00
It's really reliable and it's doing what we needed to do in
45:04
the business. What I want to do under the hood
45:06
as I want to put the Ferrari engine under their right.
45:09
And I maybe I want to upgrade my wheels.
45:11
I want to put some Michelin tires on them and like that's
45:15
what I want to do. But my Team are still getting
45:17
into the Ford Fiesta, but then like next time they put their
45:20
foot on the accelerator, the like are like this is you know
45:23
it's just so much easier. So when it comes to vendors that
45:27
are providing the like the actual development and
45:29
deployment stack for specifically for data science, I
45:34
actually feel differently that I want to build that more
45:36
in-house. I want to be more open source.
45:38
I want to check pick and choose because of where we are, and I
45:41
think it's because of that skate again, it goes back to.
45:44
When, would you need It's because like that, you know,
45:48
that digital payments we needed a fraud tool like we didn't come
45:53
build something in House at the time.
45:54
So we just went to third party. What's happened in a lot of the
45:57
data science, machine learning spaces.
45:59
You've already got the team, well established and all of the
46:03
tooling is only kind of starting to catch up to it, in kind of in
46:07
in retrospect and the tooling that we all needed.
46:10
Five years ago, you know, look at the ml observability space.
46:14
We've all been shouting about you know, me, Learning
46:16
monitoring that Lord knows how long scuse my language and, you
46:22
know, we're still only kind of just getting there and actually
46:25
the best things on the market right now actually in the open
46:28
source, like nanny ml, for example.
46:30
Right. And that I think that's the
46:34
weird thing is that we've all got these teams and now you've
46:36
got all these vendors who are like oh you know what you're
46:39
doing? Just change back on live.
46:40
If we were here for you, five years ago, everything would have
46:42
been great but it's really, it's a really tough environment.
46:46
Moment, whether whether your user or a provider, I think huge
46:52
points there and I don't want to delay too much longer because I
46:59
really want to ask you the question about our in
47:02
production, but if I can sum that up and try and distill,
47:06
what you're saying about your Ford, Fiesta, which lovely cars,
47:10
got it? Especially the design.
47:13
I can't help but smile when I think about the Ford Fiestas.
47:16
And throwing the Ferrari engine in it.
47:18
Would it not be also beneficial? If you could just have a
47:22
different abstraction layer that you bring the tools in on and so
47:26
the data scientists nothing changes.
47:28
They still get into their Ford Fiesta but underneath the hood
47:32
you've souped it up. Exactly.
47:35
That's exactly my point. So from their perspective they
47:37
don't know from an infrastructure point that we're
47:40
moving from a single node to kubernetes multi instances and
47:45
the scaling of that is It's all handled under the hood.
47:47
They don't need to worry about that because the way they deploy
47:50
still, the way we, they see our deployment pipeline is exactly
47:55
the same. They still push to GitHub or,
47:58
you know, they still push a commit to master and that gets
48:01
approved and merge and then it's in, they don't need to know
48:04
about that the CI and CD has all changed that under the hood.
48:08
There's a slightly different, there's different service
48:11
between serving and then hosting the model and the capabilities
48:15
that happen. That's where you need.
48:17
Good engineering though. And you also really need data
48:20
and data engineering and Engineering, who really
48:23
understand this world and a willing to nudge the data
48:27
science team. Rather than say, I think we
48:30
should use this capability. I know there's been a lot of
48:33
that Cube flow on this podcast lately, you know, somebody
48:37
coming in and being like we're now going to use Cube flow and
48:39
going to completely change your way of doing everything.
48:43
If it isn't already as you've talked about before like Que
48:46
Bonita shop. For example, that's the
48:48
Practical practicality P, so you're completely right.
48:51
That's that's what I am really passionate about was like, what
48:55
does it look like today and changing underneath the hood?
48:59
And that's, that's not I don't think that's easy and I don't
49:03
think we've got a cracked here. Okay.
49:05
So are in production. You are passionate about this,
49:09
can you explain a little bit more?
49:11
Yeah. So when so I started out my
49:13
career in the terrible software programming Fast.
49:16
Which I don't even think we can call it a software programming
49:19
language. I think it's just a SQL
49:21
interface, right, but it, we don't crush me on that having
49:25
before that done everything in C++, then moved to SAS weird
49:31
trajectory. And so when it came to going to
49:35
a start-up where everything we were doing in open source in
49:39
2014, I was like, oh what do I learn to kind of to build
49:45
regression model? That's what effectively always
49:47
doing in the original days of a lot of decision trees a lot of
49:50
regression models. Don't judge me.
49:52
That was literally how you know this field was back then and it
49:57
seems much more straightforward to go from SAS to our and python
50:02
scikit-learn. Just wasn't there in terms of
50:04
the support that it has nowhere near this voice today.
50:07
But back in 2014, it still felt very shaky was our had a huge
50:11
Community, had a great contribution and a really
50:14
commitment to Especially like carrot package was really well
50:18
maintained. So my journey with our has been
50:21
that because we became a very orientated team with with quite
50:26
a lot of python actually thrown in around that.
50:30
But development and models, the start up our hours are I was
50:33
there for six years, was a lot in our and then eventually in in
50:37
Python because we had to do a lot of very low latency
50:40
deployments. As I said, we were that data as
50:42
a product capability, so deploying NG models to search on
50:48
a user in real time with their consent of data.
50:51
So you had to pull all the data from a different API and compute
50:56
the decisioning model. Let's say up in the same with in
51:00
a matter of milliseconds that's when Payton did come into kind
51:04
of multi-threading and helping with a lot of the
51:07
multi-processing parallel processing.
51:10
Sorry, you can edit that bit out home.
51:13
So yeah, and some multi-threading and Arlo
51:16
processing for pop in Python. But prior to that point, we've
51:19
been building all of our models in our and had to just load
51:23
balance them. And I'm not going to say the
51:25
word I was going to say that but the load balance them a lot and
51:30
too hard to handle that. So all are in production in
51:34
putting our in production since 2014 and built, lots of hacky
51:40
ways before rocker actually was properly supported through the
51:45
docker. T ways to kind of deploy our
51:49
with a lot of very good, engineering infrastructure
51:52
resource behind. It really disappointed that the
51:55
time like in our startup we just didn't have the capacity to
51:58
properly. Commute contribute to the open
52:00
source Community for what we were doing.
52:02
But I think, you know, I was massively advocating for our
52:07
introduction here in Manchester in the UK at the time and I've
52:11
always had a little bit of a soft spot for our even though
52:14
now I've done various Use them in python as well.
52:18
Well, the I feel that the data science take on our is
52:23
interesting or like, where our sits in the ecosystem is very
52:26
interesting because on the one hand clearly python has captured
52:31
a lot of mindshare and the last five years and certainly all the
52:36
big tech companies have pushed forward.
52:39
The all of our tools, tend to fill up high towards.
52:42
I mean, all these things that the Heavy Hitters are and the
52:45
new tools are All usually, the first language you support
52:49
natively is Python and yet, so it's like, so new person coming
52:53
to the fields. Will see this and just assume
52:56
that, that's the only way the data science is done, but if you
52:59
spend any time talking to data scientists that are, you know, a
53:03
little bit got into a little bit before 2020.
53:05
Let's say, our is actually, still is used by a lot, a lot of
53:10
data scientists, right. And a lot of data work is still
53:13
done in our and there's an interesting effect.
53:16
Because of that aspect, there is actually very extensive tooling
53:19
and Library support in the our language.
53:22
I wonder. I mean is this is this something
53:25
where we expect that our to still carry and hold its own
53:29
over the years to come or will this will Python and gradually
53:34
just eat away at this and eventually all the native our
53:37
shops will just have to get with the times and and switch over to
53:42
python. I would be heartbroken if that
53:45
happened. And I really would be just
53:47
because of the grammar of Graphics in are so beautiful
53:50
that. I just I just feel like as well
53:52
for people moving talk about scaling data science.
53:56
I think our is one of those languages that from a
53:58
statistical sense is so easy to get to grips with very quickly.
54:02
It's purpose-built. So that's, you know, without
54:04
going into all of us python debate, I think specifically on
54:07
your question that I wish our studio themselves, is it
54:11
invested a lot bring all the tooling and capability to the
54:15
Forefront. And I've we really significant
54:18
dividends pay for that now. So the whole don't tidy models
54:21
Universe does a lot of what ml flow says on the hood for
54:26
example but for our I think we're as having a hard time it
54:31
specifically like a lot of the open-source tooling, it doesn't
54:36
really hasn't really accommodated for our Upfront,
54:40
for example and it's kind of like an after 4 or it's like oh
54:43
or isn't on our roadmap and I I think that's where it's going to
54:47
suffer. Is that our studio can do so
54:50
much. And they visit invested heavily,
54:52
but they've also had to go co diagnostic, right our studio.
54:55
Now, supporting all everything that they build, that's more of
54:59
an Enterprise scale or paid for license.
55:03
It's supporting both Python and our cause I think they're
55:06
realistic about that. I don't think that I think it
55:10
depends. What happens at the
55:11
universities. I think that's honestly, I think
55:14
that would be a 10 for that because that was a whole thing
55:17
with SAS is that a lot of people sat really got into the
55:20
university institutions and a lot of people learn SAS.
55:24
I think it's what we're universities decide to hedge.
55:27
Their bets in terms of python or are from an open source
55:30
perspective or in fact, potentially Julia and what what
55:38
training comes out of that, I've always been one to say you
55:41
should as a good data scientist, specifically data scientist, you
55:44
should be able to work. Early code agnostic.
55:47
Why? The concept should stay
55:48
especially with the way the packages and libraries are
55:53
today. You don't have to be building a
55:55
limb and logistic regression from scratch anymore.
55:58
Were you understand, do you know the weight of evidence equation
56:02
and log odds? You it's literally all out of
56:04
the box for you, so you should be able to switch around your
56:09
your, your coding, your. Your your the language or
56:13
writing in. Yeah, I don't want to see of
56:19
eaten by Python and I think we'll always have a fairly good
56:22
step. I think the university and the
56:24
education system will have a lot.
56:26
So for that. Hmm I think you're right.
56:29
Julia will also because I've been coming, that's another
56:32
that's another Dark Horse. To keep your eye out for and
56:35
seeing that even at the University level you know,
56:39
computational programs mathematics programs that are
56:42
maybe even starting with Julia before Python and in Python ends
56:45
up being this, Kind of weird player in the whole Space where
56:49
you can do a lot of stuff like ML and data science but you can
56:52
also like right servers and python, right?
56:54
Like Django in the old days but now even more recent tools like
56:58
Fast API so it kind of does a lot of things but I don't know,
57:02
I don't know that I'm super optimistic to be perfectly.
57:05
Honest about our future don't don't.
57:09
Don't make that a hot take. This is this is no I was just
57:12
gonna I was just gonna do that. I think you might be right
57:17
though unfortunately just because pythian he's so
57:19
generalist. But coupled with the fact that
57:22
you have extremely well-versed and well knowledge maintainers
57:27
in the specific libraries like Spacey and And, you know, see
57:32
everything with talking face is all built in Python.
57:34
So, you know, there's so the you've got those and you've got
57:38
a great, like, you've got the T, the T shape right down, it's not
57:43
down. Yeah, of course, the teacher in
57:46
the piping language and it makes me, yeah, it does make me sad
57:51
because we were not, we're at the financial times out.
57:55
Its science, team isn't all shop.
57:57
At the moment, we will go towards more code agnostic with
58:01
Different things, as I said about our maturity curve, but
58:05
that's not in any time soon because we don't need to write.
58:08
So it's about what capabilities. Can we build with our?
58:12
And that's where I found. Oh, there's so much from the
58:15
market now but it's all four pipe them.
58:18
Yeah, yeah, totally. So alright, want to wrap up with
58:24
a few last questions for you? Maybe a little rapid fire.
58:28
First one is, what was the last book?
58:29
You read children of time? Didn't try Koski nice.
58:36
And what was the last bug you smashed?
58:42
I think. Beetle for my bathroom and I put
58:45
in like a physical book or like a computer software by, but it's
58:53
fine. You can do a physical one,
58:55
that's totally. Okay.
58:57
It's a cool one. You're like, actually one of the
59:01
first people that has interpreted that way whenever
59:03
dimitra's ask that question in my head, I'm thinking, like,
59:06
people must know, this can go both ways and then everyone was
59:09
like, no, no, the software the software.
59:10
But you actually interpreted literally, which Thank you for
59:14
someone who's sensible? I know that.
59:17
Okay. I'm like me.
59:20
Yeah. Go with that.
59:21
That was a good answer. No it's funny because what I
59:28
don't know if we have time for this long story so I'm just
59:30
going to leave it right there or great answer last time that you
59:34
saw some machine learning advertisement or marketing that
59:40
made you laugh because of how corny it was or it was just well
59:45
done either or that's a really tough question.
59:50
There's no, there's no, it's just not that good.
59:54
The marketing teams aren't that good?
59:57
No way. Well, that's it.
59:58
Yeah, but that did all the answer that.
1:00:04
There we go. There is none.
1:00:05
There is that it was a trick question.
1:00:07
Yeah. So we're going to keep going
1:00:14
because that was a perfect. That was a perfect answer to sum
1:00:18
up this state that we are in. What is a piece of technology
1:00:22
that you are? Bullish on.
1:00:23
That might surprise people. So you can't say, our we already
1:00:26
know your love for our, what is another one?
1:00:32
And it can be feel like I gotta press his all these questions
1:00:36
now because I don't know where you're going to take them but it
1:00:39
could be even like home automation or something.
1:00:42
You don't it doesn't have to be like software.
1:00:44
I don't want to bias you with like oh kubernetes Wow, I know,
1:00:58
there we go. The the the like plant watering
1:01:03
God and automation space. Oh, I'm alternating doing things
1:01:10
in the garden because that's really hard, but it's also
1:01:13
really time consuming. That's, that's where I'm gonna
1:01:16
go. Brilliant.
1:01:18
I love it because it was worth the wait last one that your
1:01:23
Really Gonna Love if you've been loving all these last question
1:01:27
is, how do you want to be remembered that?
1:01:30
I waffled like, well my Tombstone is this was it just
1:01:37
from this cop car? No, yeah more like when you
1:01:41
leave the planet when I leave the planet, The II finished
1:01:46
loads of different Hobbies. That's how I want to be
1:01:49
remembered. She always finished loads of
1:01:51
different interests. She did a lot of different
1:01:54
things. Did Leon