Latest / Agentic Conversations (formally mlops.community) / Harnessing MLOps in Finance // Michelle Marie Conway // #174

Transcript

Harnessing MLOps in Finance // Michelle Marie Conway // #174 · Agentic Conversations (formally mlops.community)

0:00
Hi guys, I'm Michelle Conway. I'm one of the lead data
0:03
scientists at Lloyds Banking Group, currently based in London
0:06
in the UK. I take my coffee first thing in
0:09
the morning. Welcome to the ML OPS Community
0:12
Podcast everyone. I am your host Dimitri OHS, and
0:15
today I am with my man Stephen, who is phoning in from the
0:23
coolest city in the world, Berlin.
0:25
What's up dude? Yeah.
0:27
Hi there. It's doing good.
0:29
I'm so good. I am so good, and we should
0:33
probably mention folks, that you're on the Berlin local
0:38
charter Organisation team. You and a bunch of other
0:42
incredible people that too many to name, because Berlin's like
0:45
the strongest. No, it's not the strongest, but
0:48
it's one of the strongest chapters that we have and we did
0:52
a hackathon a few. By the time this comes out, it
0:55
will be like a month or two ago. And had a blast doing it.
1:01
We also were judging and saw some funny judging pieces.
1:08
And I figured, you know what, dude?
1:09
You gotta come on and host the podcast with me because we have
1:12
too much fun when we're together.
1:14
You got a little stage fright, though, didn't you?
1:16
Let's not lie. It's like usually, like, you
1:19
know, like I've been recording for podcast and like when
1:21
someone interviews you, it's easy.
1:23
When you interview yourself, what is it harder?
1:25
You're like oh. Do I want to?
1:27
Or something. That's it.
1:29
Because I kept writing you on Slack like you asked the next
1:32
question, you asked the next question.
1:33
But you had everything. Everything.
1:35
Off my phone is away. Everything is like no one could
1:39
bother. Me.
1:40
Yeah. You were absolutely untouchable.
1:44
And so that's fine though. So listeners, if you do not hear
1:49
him that much in this episode, do not worry because you will be
1:53
back. We're going to have your own
1:55
episode. It's just going to be me and you
1:57
shooting the shit. Like we do and when we have.
2:01
The occasional calls. But for this conversation, I
2:05
mean, it was awesome. Michelle is such an incredible
2:08
human. I was laughing.
2:11
The whole way, basically, she was.
2:15
Putting me at ease, just cracking jokes the whole time,
2:19
bringing such a fun and humorous light to such a serious topic.
2:23
I mean, they're managing a lot of money.
2:26
They're also spending a lot of money on their cloud migration.
2:32
Yeah, I was like, yeah, we're going to go back to the Google
2:34
office in Berlin, you know, and be like, hey, all that money.
2:38
Yeah, exactly. So they she mentioned in in this
2:42
episode, I think they're going to allow us to keep this.
2:45
Started in that they are migrating to GCP and it is
2:50
costing them 3 billion and when she first said it I understood
2:54
million with an M and then when she said it the second time I
2:57
was like ohh. No, that is a lot of money that
3:00
we're talking about and over a nice chunk of time and so she.
3:06
Just talked about the the challenges that you face when
3:09
going from fully on premise using everything that is tools
3:14
for the on premise world to now. Going to a cloud scenario and I
3:21
know GCP is the main one, but she also mentioned that they're
3:24
going to be using a few other cloud providers I guess, so they
3:27
they do have that. Feel system in place.
3:32
Yeah, the failover system and. That was an awesome piece.
3:37
What were some of your key takeaways, though?
3:40
I think on my side was all like everything related to security,
3:43
You know, like especially working in tech, usually you
3:46
kind of see security as like, Ohh, it just annoys me, you
3:48
know, I just want to write my code and everything.
3:50
And the way she was saying it was like, yeah, I mean, you
3:53
know, when you're a bank, humans like the founders of so many
3:56
people and it's like you can't make mistakes, you know, so.
3:59
So I think that was my key takeaway.
4:01
It was like, yeah, that security, the monitoring of the
4:04
human models, you know, everyone always talks about, yeah, we
4:07
have monitoring. We have explain Ability is very
4:09
important. But what they have for
4:11
expandability, sorry, is amazing, you know, she was
4:14
saying, yeah, like that feature is wrong in that part of the
4:17
code is wrong. That's what your model is
4:18
failing. I wish I would have that.
4:22
Something like that is a godsend.
4:24
And of course they are highly regulated industry, so they
4:29
can't afford. To play it fast and loose as you
4:33
say. So it makes complete sense.
4:36
I I also think that there was some super fun stuff that she
4:40
talked about when it comes to being a woman in the tech
4:43
industry. We we kind of went off on a
4:45
tangent at the end there and I found it enlightening to hear
4:49
her perspective and and her experience of how she was the
4:54
only woman in her class at university and then how that has
5:00
how she's basically lived through that scenario sometimes.
5:05
Her job and what that is. So it was cool.
5:09
I enjoyed this conversation 100% and hope everyone else does too.
5:14
Let's get into it. Let's just dive right in if.
5:19
You could though. While you're out here listening.
5:24
Hit that like or give us a 5 star review or subscribe any of
5:30
these things that you can do on wherever you are listening to
5:32
your podcast. That would mean the world to me.
5:35
And if you want to go above and beyond because you are one of
5:38
those overachievers, send this episode to a friend and tell
5:42
them they must listen. All right, let's jump into the
5:46
conversation with Michelle. well just get into it.
5:56
Let's let's get started and where else to start besides?
6:02
I mean, Liz, I'm not going to go right to the core and ask you
6:06
about your GCP migration because I feel like we need to soften up
6:11
the blows a little bit before we go deep down into the fun stuff.
6:17
So maybe you can talk to us a little bit about how you came to
6:22
be in tech and what your journey has been so far.
6:25
Yeah, I'd love to. So I have been working in the
6:30
data and tech industry for about 12 years now.
6:32
Not to give away my age and 21 plus VAT in my head, always.
6:36
That's Value Added Tax. So I loved mathematics in school
6:40
and I studied mathematical science for undergrad, really
6:44
enjoyed it, studied it in Dublin, it's really good fun.
6:47
And after that I moved to London and I started working in
6:50
investment banks, went on to work in insurance, Then we went
6:54
and worked with Amazon, BT for a bit, worked for a data
6:58
consultancy. And then five years ago I joined
7:01
Lloyds Banking Group. And when I joined I was actually
7:05
midway through a part time Masters and data science,
7:08
because I really wanted to upscale in that computer science
7:10
theory that I didn't know an undergrad and really understand
7:14
that programming metrics. Because I've just been doing the
7:16
statistical analytical side of things.
7:19
And I had been obviously heavy lifting in big data and we could
7:21
get that and understand that using SQL.
7:24
What's I really wanted to offer things.
7:26
And that's where I got to learn Python for the first time.
7:29
Absolutely love it and move into that more techie space, from
7:33
being just a flat station to being a day's lions and moving
7:36
more so into like the machine learning engineering space and
7:39
understanding those best principles.
7:41
Suppose for me, I like love to solve problems and puzzles and
7:44
learning so-called. Kind of lifted in nicely.
7:48
So that's that's one journey that I've seen.
7:53
Being the harder of the two, and I say the harder of the two,
7:57
going from like software engineering into data science, I
8:01
think is a little it lends itself better if you if you can
8:05
do it because I definitely can't.
8:08
But then I think there's a lot of things that you need to focus
8:13
on when you go from the statistics and math into
8:17
software engineering and more of that engineering role.
8:20
What are some key things that you mentioned how you saw?
8:24
The best practises of being an engineer and better ways of
8:28
doing things. What are some key pieces that
8:31
you remember being like? Ohh yeah that's actually very
8:35
important. If I do that and I didn't
8:36
realise it when I was just writing scripts or yeah I think
8:40
so. For me, I was really used to
8:42
building models and production using fast and fast was very
8:45
like it's built for you little boxes and it's like, I know it's
8:48
very old school like don't judge me, but you would be surprised.
8:51
A lot of banks actually still have fast lie for their credit
8:54
risk couples choosing blocking group.
8:57
It does exist. So when I went to learn Python I
9:00
was like what is this object oriented programming?
9:03
What do you mean? I can construct everything from
9:05
scratch myself and build it if that really one.
9:09
I felt exposed because you're trying to like things just don't
9:13
come packaged and wrapped for you to go build the models, you
9:16
have to build a pipeline and everything from scratch.
9:19
You have to think it through and there's a lot of coding to it in
9:22
a good way. But that means it's more
9:23
flexible and it's more adaptable and it's it's quite creative as
9:27
long as you are not terrified at the start like I was.
9:30
So I was like, Oh my God, programming is really hard.
9:33
Should people do this? So 100% And then what got you
9:39
over the hump? When did you know that you were
9:41
comfortable with it? I think once I got the basics of
9:46
Python like the standard library stuff.
9:49
Once I understood that, I I find they're key because when it
9:53
comes to working on production code bases, yes we do pip
9:56
install libraries from like pandas, numpy like staff models,
10:00
sklearn and there's a rake of other pytorch, et cetera.
10:06
But being able to understand how to like loop through things,
10:10
nested lists, nested dictionaries, be able to read
10:12
that and understand us. And like being able to like
10:14
click through code and writing things in modular code bases and
10:19
navigating around them because you could have repos there like
10:22
30,000 lines of code. How do you get into that Like
10:26
nose dive right into it? Because you need to.
10:28
It's not built into the packages you've just installed.
10:31
It's actually your code base. You need to be able to read that
10:34
and understand that. So having the bare basics of
10:36
Python, that's the fundamentals that help you.
10:40
So we'll put that is such a fun piece of the whole coding
10:46
journey. Is just like, OK, who was the
10:48
last person to do this? Where's the?
10:50
Where's the comments on this code?
10:52
Why? Why is it not well documented?
10:56
I'm not going to lie. My guilty pleasure is looking at
10:59
git annotates to be like who has worked on this and created this
11:03
this way? And it may have been something
11:05
that happened over the last six years, but it's still worth
11:09
looking at because everyone has their different styles.
11:11
So like understanding it like and learning.
11:14
I don't know how people have involved and readability is
11:16
really important, but some people like writing things in
11:18
one line that. Just 50 things, but that's OK.
11:21
Exactly. Well, it's funny you mentioned
11:23
that because there is, there's a few people in the community and
11:27
that we've interviewed over the last couple of months that talk
11:31
about how important it is for data scientists to understand
11:35
clean coding fundamentals. And so it feels like you learned
11:41
that maybe. Just.
11:43
Trial by fire or like baptism by fire, can I say I definitely
11:48
learned the hard way. I remember I was building a
11:50
model in a notebook and I had to go do I had all the data
11:54
processing in this notebook. I had all the different
11:56
versioning that I needed to do. I had like multiple models built
11:59
and I remember I sent it to an MLA and machine learning
12:02
engineer at the time. Check your finger, roll back
12:05
this one line. He's like it goes without saying
12:08
this notebook needs tidying up because I had like massive
12:11
Python scripts. I hadn't followed any best
12:14
practises with modular code and now I'm like Oh my God, they
12:17
were so right. I need things like cleaner, easy
12:19
to use like notebooks should be to like display things like
12:22
plots, graphs, outputs the data frames, not hold like 2 lines of
12:27
code like iPhones. In doing so, I definitely
12:30
learned the hard way and I was not doing perfectly from the
12:32
start. And what do you do now?
12:34
What's the way that you do it? I imagine you're still using
12:37
notebooks to explore. I'm still using notebooks, but I
12:42
am very big on GitHub and version control and writing
12:46
things in modular fashions and following the DRY principles of
12:50
Do not repeat yourself and like not hard coding variables into
12:54
functions. Especially if it's like in are
12:57
like dock strings, like having it in one place and conflicts
12:59
that you can like whip it around the code base and being very
13:03
careful how we structure things. And even now I've got junior
13:06
data scientists on the teens and I'm like month by month training
13:10
them and I'm trying to embed those best practises before just
13:13
give them a notebook to like explore and do whatever they
13:16
want on. Because I feel like it's harder
13:19
to revert it when you have a spiders web in a notebook, but
13:22
if you build it up bit by bit it's a little bit easier.
13:25
I'll let you know how it turns. Out Ohh, that's so classic.
13:28
And, yeah. And how do you go to trade those
13:30
people to try those juniors? Like, you try them one by one,
13:33
or is it? At the moment, it's just in
13:36
small groups of three to five. I'm very big on mob programming,
13:41
which is where, you know, you call between 3:00 to 6:00 people
13:44
is like the optimal kind of thing.
13:46
So having little study groups like that also helps when you're
13:49
like showing someone something new and they can bounce off each
13:51
other. They kind of find if you're just
13:53
teaching one person on their own and they go away to look at
13:56
something, they might get stuck. And having a peer there would be
13:59
like, oh, I'm stuck in the same thing.
14:00
Here we try to solve it because I might not spot the little
14:03
thing that they're on. Some might see the bigger thing
14:06
that's further along down the line, if that makes sense.
14:09
I mean I I obviously want to go right into the GCP migration
14:15
because it's been. I've been thinking about it ever
14:19
since we hit record. And I feel like there are some
14:22
things you left a few bread crumbs in the way that you were
14:28
talking about, the code of what you find at a bank that has been
14:34
around for. 50 plus years, I think you mentioned.
14:39
Yeah. So the group altogether, between
14:42
about 16 brands, there's 70,000 colleagues.
14:45
It's over 300 years old. Wow, OK.
14:49
It'd be like between all the brands because the main brands
14:52
would be like Halifax Bank, Lloyds Bank, you'll have Bank of
14:55
Scotland and then Scottish Widows, the insurance as well as
14:58
those of other brands for insurance, private wealth
15:00
investment that will get into. So because of that size and
15:04
scale, everything joint is joined together under one group
15:09
and a lot of our stuff is very much on Prem when it comes to
15:13
data because we are risk averse. We take data very seriously in
15:17
Lloyds like it's it's customers, financial details, products like
15:22
that has to be well looked after, well catered for.
15:25
It can't just be, hey, we'll leave it over there, it'll be
15:27
fine. Like there's loads of risks and
15:29
security. That is like checking
15:31
everything, which means we do have, yeah.
15:34
I I would want that in my bank. That's kind of what I I would
15:37
air towards the side of that exactly.
15:40
So it's like Fort Knox, everything is well protected but
15:44
it does mean because we've protected everything, we've
15:47
built everything in house and owners ages on Prem.
15:50
We're now trying to migrate that legacy that rich data onto cloud
15:55
and we've picked Google Cloud. We do have other little
15:58
providers, but that's our main agenda and in the last five
16:01
years we've sets like 3 billion to do that.
16:05
With a B would it, would it be, But it's a big journey because
16:09
it's like you can't just do it really nearly.
16:11
So we've started doing it and releases were you like get a
16:14
small versioning, but the environments you get are very
16:16
restricted lockdown. It's not like we're getting GCP
16:19
like being you could like open in a website right now and give
16:22
them our credit cards. We're not getting that person,
16:25
we're getting the one that's been scanned through security.
16:28
That's like fit for purpose. We can't just give colleagues
16:31
anything they can go spin up massive cost.
16:35
We we may have had stuff in the past that was running overnight
16:37
in GCP and hit 100K and something.
16:40
So it's like we can't do that. We need to be sustainable and
16:43
how we do that is very much it is difficult because we're so
16:48
large like we've got 70,000 colleagues across the whole
16:51
group, which means that when we're doing this, you've got
16:55
multiple impact departments of like hundreds of people trying
16:58
to drive this forward. So it's like a gigantic ship.
17:01
I think the Titanic, you're trying to move that with
17:04
everyone in it and like steer it in the right direction.
17:07
It's great when everything goes to flow, when everyone's in the
17:09
wrong line and Burrow like steering the same way.
17:12
But sometimes things veer off and like people are like it's
17:15
like scattered cat but on another level that you're trying
17:18
to get things done. And you'll also have live
17:21
production models that you're trying to protect because the
17:23
business use them and their own. And it's like you need to take
17:26
this model, rebuild it into GCP and have parallel runs and they
17:30
left hunter and it's and then you have to move the data as
17:34
well, not just the code and it's it's entertainment cause it's
17:38
like the stress is real and the intensity is real but it is
17:42
moving. It feels slow, but it is moving
17:45
at a pace that when we get there, it's going to be this
17:48
beautiful thing. What I feel like is going to be
17:50
this beautiful thing. But between then and now, that's
17:54
gonna be the real journey of, like all hands on deck and
17:57
trying to, like, muck in together to get it done.
18:00
But you said upside like. Five years to do this?
18:04
How long? I think it was a couple of
18:06
years. I don't know the exact plan.
18:08
All I'm going is by what we've released publicly and that was
18:12
one of the bits of public information that we put alive.
18:16
But we will see how long the journey goes.
18:18
Right now it's in process and all wheels are in motion.
18:21
You do see good things like environments being created and
18:25
you're like, oh, this actually could work when we're using
18:27
this, which is good to see, but we just want to keep those
18:30
releases coming and seeing the end version that is the usable
18:34
thing that we can now start moving those live models into.
18:38
And is it mostly focused on much learning models that are going
18:40
to work that are going to run on Google or is it everything
18:43
related to Lloyds Bank? They're trying to move as much
18:47
as their data as possible to GCP and because they they want to be
18:52
really sustainable with like let's create data products, have
18:55
the proper risk management, have the proper metadata, have
18:58
everything in one place that it's usable stored on as well.
19:01
It's more manageable when it's rather than just like a little
19:05
bit. Obviously some bits are now in
19:06
GCP for testing and streaming and stuff, but we need to like
19:11
kind of move everything because how are we going to manage to
19:14
different systems. You know.
19:17
And Speaking of two different systems, you mentioned something
19:20
there that I think is worth digging into, if we can about
19:25
how. Effectively.
19:28
Before you're fully migrated. You're running two systems, and
19:34
so how do you know which one is running?
19:37
How do you keep tabs on what you've been able to do?
19:42
What? Like it just seems like you just
19:44
added, you just doubled all of your complexity, definitely.
19:48
And that bit I'm going to enjoy seeing because I've not seen it
19:51
done yet. But we're gonna have to be so
19:54
careful because you need to make sure like the the two systems
19:58
that you have, the two models you have live like what's
20:00
feeding into both of them is the.
20:02
Same. And then checking what's coming
20:04
out of the both of them is the same, because if it's not, it's
20:07
like, well, it's something's gone awry and like, how are you
20:10
going to explain it? Because then essentially you've
20:12
got two different models. Now it's not.
20:15
The same in the mouth. So we are doing a lot of
20:18
preparation on how we do all that checking and how we make
20:21
sure the data is matching when it feeds in to make sure it's
20:24
like the same data going in. Because if there's like
20:27
discrepancies, it's going to like make a difference because
20:29
these models are live training. That's not just like some of
20:32
them are pickled and it's like feed stuff into it.
20:35
It's like a living and breathing thing that operates and moves
20:39
forward with time. So it is.
20:41
It is going to be interesting. Having these two systems,
20:45
hopefully it won't be for long. I think the parallel runs will
20:47
be about like two to three months once it's up in action,
20:50
and then we shall see. We should say goodbye to the
20:54
legacy one that we've had at having the funeral would be
20:57
like. Thank you very much for your
20:58
time. Great.
21:00
The the thing that I I was thinking about too on that is
21:05
any small discrepancy and I imagine upstream you're going to
21:10
have so many different ways that you're dealing with the data
21:14
that it could Causeway big changes downstream and what the
21:20
model is outputting. So debugging that is also going
21:23
to be a lot of fun to figure out.
21:24
OK, well on Prem it's working like this and we are using these
21:28
tools. How much of what you are
21:31
shifting over is just an exact replica and how much is an
21:36
upgrade or new? Because you'll have to tell me.
21:39
I always make jokes about banks running Cobalt.
21:44
Are you Are you all running cobalt in production still?
21:48
Is there some? No, but have you heard of Nexus?
21:52
We've done all Python codes like for security and we have it in
21:55
our own repository that we'll all pick installed front.
21:58
So we don't actually pip install from remote websites.
22:02
It's it's just a security thing to make sure there's nothing
22:04
coming in externally that could be vicious or religious kind of
22:08
thing. But no, we're not using, but I'm
22:09
sure we're using other things. For.
22:12
Yeah, equally as a legacy, Yeah. But actually like talking about
22:18
Nexus, that's a funny one. Because I remember Diego
22:22
Oppenheimer about a year or two ago was talking about how pip
22:26
installing things. Like you pip install these
22:28
packages and you don't know what's in them.
22:31
And also now it feels like there's a lot of the you can
22:36
just download a model from hugging face and you have no
22:39
idea what's in that either. And so there's a lot of things
22:44
that we kind of played a little bit loose and sometimes it comes
22:47
back to bite you in the ass. It's always on, I think.
22:51
People are doing this new thing. Well, it's not new thing, but
22:55
it's where do you know if you misspell a Python package where
22:58
you do a pip install, someone has like intentionally done the
23:00
misspelling version and like may have something like vicious and
23:04
malicious and they're like malware and that you've now
23:06
installed that. So even though for all this data
23:10
scientists going to use those packages, it has to go through
23:13
all their clearance and security.
23:14
And it might fail every now and again because it's like Fort
23:17
Knox. But I'd rather do that because I
23:20
don't think I've ever seen a virus, something laptop, my work
23:23
laptop ever. Because everything is like so
23:26
strict and so secure, we found that out trying to record this.
23:30
We had to move to your phone. We did.
23:33
I I couldn't use Zoom. I couldn't use Teens.
23:37
I sent emails with like Microsoft team links and it.
23:40
But yeah, no, you received them, but it wouldn't let me like
23:43
actually in. So everything's really secure,
23:47
even my personal e-mail, because I have access to data on my
23:50
personal my work laptop. I can't send external
23:52
attachments because it's seen as a security risk.
23:55
And and rightly so, because it's like, what if that attachment?
23:58
Had data or something, It's to minimise the risk or anything
24:02
leaving that's classic. Yeah, but maybe to follow up on
24:06
that, So so you know, when you when you train machine learning
24:09
models, usually you need to access some data and you know
24:13
with other companies, usually some startups can be a bit like
24:17
free flow. You know, you just access
24:18
everything. How does it work on your side as
24:20
a data scientist there? So.
24:24
We presently have data in our global data warehouse and that
24:29
uses Teradata. I'm not sure if you're familiar
24:31
with that. A lot of banks use their data.
24:33
It's quite old, school robust and we are migrating that to Big
24:37
Query in Google. But any time currently someone
24:42
queries any tables. In my database, someone like my
24:47
boss is responsible for them and will receive emails and alerts
24:50
every time a certain table is queried.
24:53
And even more emails and alerts if someone was to query let's
24:56
say a customer table and they just pull back one record
24:59
because it's like. But what's she doing that one
25:01
record? That's private information.
25:03
Obviously to get to those tables is extra security and Terrance,
25:06
it's not like everyone has access to.
25:09
But they are very careful with. Governance of like who has
25:13
access to what and who governs it and how it operates.
25:16
Like it's not a free for all. Everything is well documented,
25:20
monitored and reviewed and access is removed every quarter.
25:23
If someone doesn't physically go in there as a line manager and
25:26
say, yeah, I'm about treating this person, is there a lot of
25:29
that information that gets removed by the data engineers
25:32
before the data scientists can even see it?
25:36
And we would still have access to the same source that data
25:40
engineers have. We do.
25:43
So we kind of have two types of data engineers in the bank will
25:47
have the data engineers that just work on our core stack, so
25:51
purely coding and SQL and then we'll have data engineers that
25:55
are coding in Python And SQL so that we kind of call them more
25:59
machine learning data engineers just because of that Pythonic
26:02
element and that different skill sets between the two.
26:05
So it it, it depends. We can like as data scientists
26:09
we can see like all the levels all the way through, but we
26:11
might have the access. To manipulate them.
26:16
Yeah, yeah, yeah. And you mentioned something
26:20
before about when you're going from one data, what did you call
26:28
it Teradata to then you're you're migrating to Big Query or
26:33
is is there other stuff that you you get to migrate to Like did
26:37
you go on a shopping spree, is that why it's 3 billion?
26:39
I I would love that. Well, it it's been estimated.
26:45
Those books are very carefully monitored to make sure it's not
26:48
an overspending thing or a shopping spree.
26:51
There is so many committees, risk views and forums that
26:55
everything goes through to make sure that everything has been
26:58
done and spent responsibly. And I can say it's been done
27:01
very, very responsibly because I think it's we're understanding
27:04
more so than anything else. Because they're being very
27:07
careful to be like during this, is it fit for purpose?
27:10
Are we bringing something in that's just waste.
27:13
We don't want that. It has to like it has to be what
27:16
we need. It has to pass all the risk that
27:18
has to be protected customer data.
27:20
And I think a lot of banks are moving to GCP.
27:24
I know HSBC are doing it, so it's not just us.
27:28
I can imagine that's on this lovely journey to cloud from on
27:31
Prem, yeah. But but it does feel like you're
27:34
opening up a whole new world, right?
27:36
Because I think there's a lot more vendors and service
27:40
providers that can you can access that have a more robust
27:46
offering for cloud offerings as opposed to on Prem.
27:51
That might just be me, like that might just be my bias.
27:54
But it it's interesting to think about how now you get to you get
28:00
to kind of explore and I imagine you're going to get your inboxes
28:05
also going to get more blown U by all these vendors as before
28:08
you could have just said we're on R.E.M.
28:10
Like, fuck off. Confirm they don't know and it
28:15
was it was interesting recently we had consultancies come in to
28:18
be like ohh we we can build you this like amazing model we can
28:22
do exist and I'm like you do realise our current tech stack
28:25
is like Domino Data Lab. It's on Prem.
28:28
Have you heard of Domino? They had great speciality in
28:32
AWS, the great Specialist and GCP.
28:36
Was a little bit like, ohh, really sorry to burst your
28:39
bubble and I'm sure you're very skilled, but that like this is
28:41
what we have it like it's it's not the best, but we have to
28:44
make do, yeah. Yeah, and maybe it's follow up
28:49
on that actually. It's like you know, Nas and like
28:52
if you were to graduate now, you would probably look for a
28:55
company where you can work on AWS or GCP or something, right.
28:59
So how does it work to actually hire when you look for new
29:01
people, when you want to hire new people, how does it work?
29:05
How can you sell that stack as well?
29:07
So we could like, yeah, that's cool to work here as well.
29:10
Yeah, Sass, sell him. Sass, you.
29:12
You must be an amazing person. Not use Sass and Lloyd.
29:16
Thank God I gave up my soft skills before I joined the
29:19
group. I will say it may seem like we
29:23
have a lot of legacy stuff and security bits and pieces.
29:28
But we also have 30 million customers and we have all of
29:31
their financial data and we have the ability.
29:35
To do good things for them. On all those risks are in there
29:40
and those cheques are gone through to make sure we minimise
29:44
any damage that happens. Because if you think about it,
29:46
imagine you're a developer and you're sitting here and you're
29:49
writing code and you've written some code that then affects
29:53
customers accounts, and let's say they go to their ATMs and it
29:57
spits out they're all the wrong money.
30:01
If you're serious about you, hit subscribe right now.
30:09
Like you don't want someone calculating the wrong things and
30:13
they're not being checked and validations in place to make
30:15
sure it's safe and secure. So there is massive satisfaction
30:20
in knowing what you're doing makes a difference to like
30:22
people's finances and it's done in a really safe, secure way.
30:27
No dig at any of the the tech people, they're very graced.
30:32
But if you want to do loads of marketing, fill your boots.
30:35
Mesh and Google and Apple fill your boots.
30:38
If you want to make a difference in financial services with
30:41
customer products, more than welcome to join.
30:43
There you go. Yeah no Oh my Yeah, but yeah,
30:56
sometimes I I worry about putting too much money in there
31:01
in the account. But because I don't know if they
31:04
take this same mentality and the same type of rigour and it feels
31:09
like you need that. You need those cheques and
31:13
balances and you need that airtight just like air gapped
31:18
situation because of the amount of customers you have, the
31:24
amount of money that you're playing with, the amount of data
31:26
you have. All of that is at a scale.
31:30
It's gigantic and it could Causeway real harm if there is
31:33
anything bad to happen. Exactly.
31:35
And I know it's not the most sexiest thing to be like we're
31:40
regulated by the FCA, but thank God we are like.
31:45
So that's on like Google may get to operate with like different
31:49
things, Mansell will get to operate in different ways.
31:52
And like in, all the fintechs have their flexibility.
31:56
But we are very much making sure that things are right by society
32:01
and one at Lloyds like purpose is to help Britain prosper and
32:05
to do that in the most safe and secure way with putting people
32:08
1st and still busy first. So I am really grateful that
32:12
they do have those cheques in place and that the FCA is there
32:15
giving those guidelines and they are met just to reduce any
32:18
massive catastrophe. That everything is well, like
32:22
the risks of it are constantly being looked at, whereas I
32:25
imagine Google and Facebook are thinking of different risks,
32:29
yes. So this is a perfect segue into
32:34
one of the things that I really was excited to talk to you about
32:37
besides the GCP migration. And that is when things do go
32:41
wrong. Because inevitably I think that
32:44
it's we're all humans and there are sometimes that things do
32:49
fall over. So what it has been your
32:53
experience there and how do you mitigate those problems when
32:58
they do arise? We have.
33:00
Definitely had that and we do have model validations in place,
33:06
so and one example of and what I can give you.
33:09
On certain models it will check to see if features are
33:13
statistically significant, so do those P value cheques, and if
33:17
it's above a certain threshold that we have set, it will fall
33:20
over. It will not produce a forecast.
33:22
It will not build that model. And the way that the pipeline
33:26
has been engineered is that it will pop up in the UI to be
33:30
like, hey, this is crashed halfway through.
33:33
We know we have a tonne of tables and pipelines and
33:35
structures, but I can tell you from the way we've engineered
33:39
it, it's like it is this specific component, it's this
33:42
specific model, it's for this specific reason, so you know
33:45
where exactly to go and look through and like.
33:49
Having enough information that you can then go explain to this
33:51
and go hey these these models have crashed.
33:54
This is the reason why we're checking these features.
33:57
And we actually found during COVID.
33:59
I know COVID changed things for everyone and it was a very
34:03
difficult time during the pandemic, but it also changed
34:06
how people were using their money and how they were
34:09
behaving. So for example December.
34:11
Really expensive Time of the year for everyone.
34:14
Lockdown. No one's leaving their house.
34:15
There's no Christmas parties. Things are stripped back a lot
34:18
more. You don't have to visit friends
34:19
and relatives. You're saving a fortune.
34:22
Which meant like customers are saving more.
34:25
But then in January they might spend more because January sales
34:28
are out there. Why not treat yourselves online
34:30
shopping? We're going to be home anyway.
34:33
So we found that there was sole nanny products and so many
34:35
models that we had, but we had to apply COVID exclusion periods
34:39
just to be like we are going through a time in our lives
34:41
where there is no point including this data because it's
34:44
just off the walls. But you no one could predict
34:47
COVID and our models couldn't predict it either.
34:50
So we had to apply those periods because it meant all our model
34:53
validations were falling over and I'm so glad they were there
34:56
so could help catch it. But obviously it wasn't
34:58
sustainable. Letting that go on every month
35:00
was like, OK, there's a clear passion here and it's the data
35:03
right now. We need to just eat to get
35:05
through it. That is so fascinating to know
35:10
that it's like you. So basically you're saying if
35:15
things are too much to one side or the other.
35:19
Like you, you set the guardrails and if the statistical models go
35:24
or if it hits a can't remember how you said it just there.
35:28
But if if it goes off the guardrails in any way, you not
35:33
only are able to tell where and why, but you can say this is the
35:41
root cause and so whoever's debugging can go and figure that
35:45
out real quick. Which also.
35:49
Explains why we will have some models that might have 30,000
35:52
lines of code because it has that like explain ability
35:55
engineered into it. And when I say explain ability,
35:57
I don't just mean pip installing sharp and sticking it on top of
36:01
the black box model because that seems to be the cure for explain
36:04
ability anytime you talk to anyone.
36:08
Knocked on a production code base.
36:10
I feel like they explain ability of being able to pinpoint the
36:13
function, the data, the model type is so much more important.
36:17
And yes, it may take more lines to write up, but doesn't it save
36:21
you more time in debugging when it does hit production and kills
36:24
out? Yeah, when it comes to the
36:30
producers or yeah, the producers, if you want to use
36:34
that term of the data upstream and things happening outside of
36:41
your model, but then your model gets it and it is tainted
36:46
because of it. Like how?
36:48
How do you deal with that? Do you just pass the buck to the
36:52
data engineer and say, hey, not my problem, you, if you mess
36:54
with my model some wrong over there, go figure it out?
36:58
Or do you also work with those teams to figure out why there's
37:01
a problem? So I suppose I.
37:03
I know what you completely meet. We do try to eliminate that from
37:07
the teams that I have worked with and you will have Daisy
37:10
engineers and. The scientists in strong teams
37:12
working together to reduce that Hey, it's your fault.
37:15
No, it's your fault. You know the finger pointing
37:18
game, yeah. And I think that does help a lot
37:22
because you need both people involved in that, like on a
37:25
daily basis and interacting communication together.
37:27
Like you don't want it to be these two big separate teams
37:30
that have, you know, disputes, You want people to collaborate
37:33
and get on and understand each other.
37:35
So when it comes to ML, we do have data scientists, data
37:38
engineers in parallel. That being said, we do sometimes
37:43
have DBA DATABASE ADMINISTRATORS that may have from time to time
37:48
gone into data and change Things change.
37:53
Controls. And.
37:56
We took. Classic.
37:57
Or it's always the BA's too. I think there's like, they're
38:02
super. Valuable, but they're super like
38:04
they're in their own little. Like not World or Box, but
38:09
they're in their own little nook.
38:12
And you're so far removed from them that it's like, hey,
38:15
something went wrong. If it's time you figure it out,
38:16
it. It's like it all roads lead back
38:19
to the DDA had done something to change things and it I I feel
38:22
like that does come with legacy not legacy but would like would
38:26
really high security stuff that there is those layers that you
38:31
were so far removed away from like the Dbas that like by the
38:34
time you get to them it's like. OK, we've already had a fix to
38:38
this that we reworked. Anything that went screw is null
38:41
values or replaced it with something else.
38:44
So it does happen, and I think that's why it's really
38:46
important, having like date engineering validations to be OK
38:49
Has the distribution of this data changed?
38:51
If so, how has it changed? Why has it changed?
38:54
Do we want to overwrite it in a way that it'll make it more
38:57
usable to go into the model, or do we need to completely unpack
38:59
it? That's funny you say that with
39:03
the Dbas, because our my friend Chiro was talking about how when
39:08
he met, he was like the head of AI for his company and he met
39:12
with the Dbas. And he was saying how for them
39:16
it was so hard to grasp the idea of immutable data and like, huh
39:22
what? Immutable data?
39:24
Like you can't change anything and there's like, no, you don't
39:28
touch data that's just like non. It's off limits and they
39:33
couldn't grasp that concept. Agree.
39:37
And I I feel. Like that's the concept where we
39:39
need like snapshots or something or like.
39:43
And then if there's a meant to be like changed, yeah, you can
39:47
roll back. We do actually do that a lot.
39:49
And one of the models, we snapshot the data every month so
39:51
that if if there's a problem with the data that comes out
39:54
next month, we can be like well, look, we can still keep it
39:56
rolling on like the thing that we have last one, this is not
39:59
like everything falls over especially in the businesses
40:02
using UIS for these models to make sure that's something is
40:05
there for them to work on. But yeah, it it is interesting
40:08
to hear Dbas don't understand the concept of immutable.
40:12
I I get it because it's they're used to working in a different
40:16
way. So the rest of us like we know
40:18
how to like keep data mutable but also build a pipeline that
40:21
you can do loads of preprocessing, you can go back
40:23
through all the steps and you have all that cold and that
40:26
logic and you can see it. Whereas they might work
40:29
something up and be like, OK, cool, I've created this table,
40:32
I'll overwrite it, put it into this, and I'll delete the logic
40:34
in the meantime. And everyone's wondering how we
40:36
ended up the final masterpiece. But it happens.
40:40
That's classic. And.
40:44
The other piece that I wanted to chat with you about was you
40:49
coming into the tech field and uh.
40:55
And being able to talk to people who are starting out in this
40:59
field, are there things that you're thinking about that have
41:04
drastically changed right from when you started and now when
41:07
you have spent time learning? And also just because of the
41:12
whole AI and generative AI boom that's happening, what are
41:16
things that you think about with those junior engineers that you
41:20
relate with? And how do you?
41:24
How do you see their job being different as opposed to when you
41:28
were coming in to the field? Absolutely.
41:31
I feel like when I came in, it was the time when one when I was
41:35
in secretary school, we had no Wi-Fi.
41:38
We had those brick button phones and then when I was in
41:41
university, Wi-Fi was just about coming in, but it was normally
41:45
like the cafe shopping. So when I entered the workplace,
41:49
that's when I had like a smartphone for the first time
41:52
and Wi-Fi for the first time. Whereas you have like Gen Z's
41:56
entering the workplace now. They, you know, so like the back
42:00
of their hand, like they are ninjas on the thing.
42:03
Fair enough. There may be a lot of like
42:04
TikTok videos, but generally like super smart guys like that
42:08
know how to use tech and they're amazing with it.
42:11
And I thought, we get that and I think they're gifts is absorbing
42:16
information really quickly because we've gotten into this
42:18
area now where we need, we're learning more but we have to
42:20
learn faster. I heard like my.
42:23
Feeling like I had to learn loads, but then with the tech
42:26
that's come out and mobile phones it's like everything is
42:29
moving so much quicker like. We are changing constantly the
42:34
different modern tech bits that we are using.
42:37
You can't just learn one method and be like right, I'm done.
42:40
You need to know how to alley it.
42:41
And we called like new methods and tips and tricks are
42:44
constantly coming out and new releases.
42:46
You need to roll with those. You need to like being able to
42:50
get comfortable with things constantly moving.
42:53
Like you can't just stop learning and be like, oh, I've
42:55
done my undergrad. I've done my masters.
42:57
Like for me I'm like, no, I need to constantly learn.
42:59
I need to know how to access the latest documentation for these
43:02
Python packages. I need to understand how things
43:05
are being implemented and maintained and operating.
43:09
I would say for our early careers workforce, it's like,
43:13
yes, do roll with that documentation, but don't just
43:17
roll with what's here and now and what's coming.
43:22
Learn to roll wet. Like what has been so like going
43:26
into all code bases, being able to understand the logic that was
43:29
written. Then there's this migration, I
43:32
think a fellow engineer recently said, and I think the statement
43:35
was right, some people's called that they've written can last
43:39
longer than their careers and they don't realise that.
43:42
And that like really blew my mind.
43:44
I was like, oh God, not that we need to make it be careful with
43:47
every single little thing, but we do need to be mindful that
43:51
what if the called you right now like outlast your career.
43:55
So much pressure. Ohh, thought it just it just
43:59
mean like it's we should take things seriously we should
44:01
really think it out and like do it in a good way.
44:04
But I do feel for the Gen said now.
44:06
Like, it's great how they learn right now instantly and they're
44:08
prepared for the future. But if they're to go through an
44:11
old legacy code base, I imagine they might struggle because
44:14
there's no pretty documentation to guide you through it that you
44:17
have to weave through it and understand and go backwards.
44:20
Like it's it's saying through. It's like if a project is using
44:24
a certain version of Python, if you go back to that version and
44:27
we'll release notes, there's certain methods that don't exist
44:29
back then. It's like, well, how do they get
44:31
around it all? They did this being able to
44:33
understand that logic and actually go through the legacy
44:36
stuff. I think that's still hugely
44:39
important and valuable. It's not always about the new
44:42
stuff and maybe. To follow up on that, at the
44:45
beginning you said that you learn Python by checking out the
44:48
documentation with standard library for example.
44:50
So is it something you would still advise for newcomers?
44:53
Because everyone's going to be like generative AI and stuff,
44:56
but they could be like, you know, we don't really care about
44:58
this standard library, so I am everything.
45:02
I think it is super important and I am trying to teach the new
45:06
joiners on the team to be like, hey this.
45:08
I use the basic Python make strings, lists and loops, for
45:12
loops, etcetera. And here's how you create
45:13
functions. And now OK cool, we're moving
45:16
into installing packages, we're going to learn pandas and numpy.
45:19
We are going to look at the documentation so we can
45:21
understand from source how things were meant to be
45:25
implemented because the developers have included
45:27
examples like that's super useful to be doing that.
45:31
That being said, I do think Jenny, why is like hugely
45:34
valuable. So me being dyslexic, I love
45:36
getting it to write out my LinkedIn posts.
45:38
I'm not going to lie, it does it for me.
45:40
I'll be like I've done this. I've gone to this thing, write
45:42
it out because my spelling grammar is going to be awful.
45:45
But I have been using it to be like hey, right out some
45:48
training material please explaining what pandas is and it
45:51
does right now something and it looks really pretty but it's
45:54
very basic and there's no depth to it and I'm just like as a
45:57
developer I would rather be nose deep into the developer release
46:02
adults than what I'm getting back from ChatGPT because it's
46:06
like it's it's surface level it's not that domain knowledge
46:10
so so yes people are like it will change their lives for the
46:13
better. Just like calculators helped
46:16
versus the number of people with pen and paper, it's gonna be
46:19
that. Situation people like what's
46:20
gonna take away jobs now? It's not it's gonna hand surveys
46:23
for the better, for the gold. We're all going to embrace that
46:26
and use it. And I personally can't wait to
46:28
get in the bank. Might take years, but we'll get.
46:32
I was going to ask. I mean, it's the hippy question,
46:36
right? But how or what are you doing
46:38
with large language models if anything.
46:40
So we are doing a lot. I think there's a lot to besides
46:43
in terms of we have documentation and confluence
46:47
like it's it's one of the things the bank uses for all its
46:50
auditing, all its written stuff, it's all there.
46:53
I think if we we are currently looking at building internal
46:56
large language models to hopefully be able to look at
46:59
that internally so that it's completely safe.
47:01
It's not leaving anywhere. So you could just be like from
47:04
my team's Confluence, hey pull back all the models that used to
47:07
random forest or hey, pull back all the models that had this
47:10
kind of thing for customers. I think that would be super
47:13
valuable and usable because I I do speak to friends that work in
47:16
consultancies and they've installed CHPT on their
47:19
documentation. They're like, it's actually
47:21
really helpful because you're going to client site and you
47:22
need to look at previous use cases.
47:24
It just goes boom boom boom. So we kind of need to embrace
47:28
things like that. I know we need to be careful
47:29
with security, but we also need to be like hey could actually
47:33
save a lot of time and save a lot of value to do it for us.
47:37
I would love those. Well if it wrote the emails.
47:39
Cause me writing an e-mail I think it's one yesterday took
47:41
about 3 1/2 hours. It's not special but it takes a
47:44
long time. That is one of.
47:46
The main use cases I think the that we saw in the LLM in
47:51
production report that we put out and it was one of the cooler
47:55
use cases, was managers getting up to speed really quickly with
48:00
how and what their teams have been up to because they could
48:04
just ask an LLM. What's going on with this?
48:08
What's where we at with the budget or what's the latest
48:11
updates, Right. And so you wouldn't have to.
48:13
ING the employees, you could just ask.
48:17
The LM and then if there was anything that you couldn't get
48:21
from the LM, you could. Graduate U to the actual person.
48:27
Yeah, no, I'm. I'm smiling because I was
48:29
recently thinking about my own team and how like we're in
48:33
relapse team, we live within the business.
48:34
It's really cool. But we will have team members
48:37
that won't be technical but we need them for like the
48:40
governance and the risk and stakeholder management and we
48:44
use GitHub Enterprise quite a lot for all our models.
48:47
It's like it's really useful. Who's trying to get some of our
48:50
non-technical team members to embrace using it.
48:53
It was quite, it was like you're saying like getting the ML.
48:57
So I come back with your forecast is like you know like
48:59
get of enterprise can show you everything that he's doing and
49:01
this is how you could check and you could see all the great work
49:04
And it was it was yeah. I felt like I was explaining to
49:10
my 90 year old grandmother how to use mobile phone.
49:12
It's interesting pulling on that thread, not.
49:15
The 90 year old grandmother one, but the the other piece on your
49:19
ML OPS team and all these different stakeholders that you
49:22
have, I imagine you have to. Be.
49:27
Handstand communication with so many different people.
49:31
Who are those people and what are they?
49:35
Why do you have to talk with them?
49:36
Basically, what are the different stakeholders that
49:39
you're? Dealing with.
49:41
Well, we have to talk with them because if you think about it
49:44
from a bank's perspective, you're going to have
49:46
stakeholders that are business people, but knowing the business
49:49
products and the history inside out, like you're going to have
49:52
someone that knows the mortgages, books and products
49:55
for the last 10 years in their head.
49:57
Everything is there. They are internal database if
50:01
that makes sense and likewise savings spotlights the savings
50:04
stakeholders. They will know every single
50:06
thing. They will know when the
50:07
financial crisis hit, what happened, when and why.
50:09
Again, all internal databases. Credit cards guys the same.
50:14
They will know how things operate, how different banks
50:16
merged at different times. So you have all these really
50:20
skilled individuals who are subject matter experts in their
50:23
area. So that's where we do have like
50:26
non-technical members of the team that do that stakeholder
50:28
management and have those relationships built with them.
50:31
So they know what they want and need and then they'll come that
50:35
obviously tell our team like this sort of business is looking
50:37
for. Can we incorporate that into our
50:40
delivery like they would like to see something next Sprint
50:43
regards to this model and that model and they then they really
50:46
helped because they do all the communications engagement which
50:48
means like the technical team can just get on with the
50:50
technical work and not have to worry about that liaisons
50:54
because we do have quite a few because we are within the
50:58
business. It's like everything looking at
51:00
the consumer lending products and other customer facing bits
51:05
and pieces that we do need to look at.
51:07
So it it can be very time-consuming, but we do have
51:10
the right people in place taking care of that so that the
51:13
technical work it's not taken from those colleagues because
51:16
you're. Dealing, if I understand this
51:19
correctly, you. Have almost like a translator or
51:25
a a buffer layer between maybe not a translator.
51:31
We do have translators as well. We do the central data office
51:34
that has translator that goes out to sell things and translate
51:37
them back. We do business leads that we'll
51:39
do kind of like, hey, what do you work on?
51:42
What your requirements I do feel like it is a set of skills that
51:47
involves a lot of patience and a lot of being a people person and
51:51
understanding and like managing that relationship which she
51:55
needs patience at the same I think that's something that I
51:58
that's on my development something but I I'm, I'm
52:02
definitely much more their tech side of things like black and
52:06
white why we do this why we do that.
52:07
So being able to like manage like human interaction I think
52:10
is a massive skill set that our team would drown without having
52:14
the people on it. They can do it.
52:17
So I wanted to talk to you about being a woman in the tech
52:21
industry. It feels like when I think we
52:25
have four women listeners if the world around the globe.
52:29
And so, yeah, speaking for those. 4 Yes.
52:36
Like what are specific challenges or obstacles that you
52:40
feel like you've gone through and how have you navigated them?
52:43
So for me, I've had a really, umm interesting experience in
52:48
the sunspot for primary and secondary school.
52:50
I went to like an all girls convent school and then for
52:53
university I went and I was the only girl in my class and it was
52:59
a complete it was a complete flip but I was just like ohh wow
53:04
this like why? Why is there not more girls here
53:06
like I know? Like in secondary school that
53:09
was 90 girls. In my class there's only 10
53:11
people doing advanced math. I I didn't realise I went to
53:15
university just be me. So for me it was a bit of a
53:19
shock. But then I kind of realised the
53:22
people that were there, they were all super driven by
53:24
mathematics. She's like me and we're all
53:27
competing with each other in a healthy way.
53:29
Like we all want to be #1 and. I really enjoyed that
53:33
competitive, that like learning from each other.
53:35
And I'll say this much, it was like a class of like 15 people.
53:40
I was the only girl. I'm still really close with
53:41
those guys like. Became really song friends are
53:45
all working in the data industry.
53:47
Keep in touch. It's been super valuable and I
53:49
think because of that really good strong foundations and
53:52
experience when it came to the workplace, yes, I did join
53:56
companies where sometimes the management to be like oh by the
53:59
way be the only lady on your team and I'm like that's OK I've
54:03
I've coped in the past I look coped now and I I do see it and
54:07
but more and more I will say the pipeline of women coming in it
54:11
is improving. Like it's not just one talking
54:15
girl on that team. Sometimes it will happen, but as
54:18
I'm progressing I've gone now into more leadership as a lead
54:22
data scientist. Still doing hands on tech, but
54:24
the other half is the managerial operations behind the scenes.
54:27
I do find I will end up being the only woman in the room in
54:30
meetings. Then you do have that thing
54:33
again. Um.
54:35
And I think for me the the biggest hindrance I've had is
54:38
probably my lack of football knowledge needs to get better.
54:43
It just needs to get better because like understanding
54:47
everyone's culture and embracing that, like I think makes a
54:49
massive difference. And people are passionate about
54:52
different things in life and they don't understand relate to
54:55
those passions. I think is hugely valuable
54:58
regardless of what the passions are and.
55:01
Even if it is football, it is awesome.
55:04
I I'm one day I will perfect defining the offside roads.
55:08
One day I will promise and. Yeah, maybe it's.
55:11
Also to follow up on this one is like, what do you think we
55:14
should do? I mean we as people are working
55:17
in tech to help that maybe to to have more women and women plus
55:21
in general. Yeah, I do think it's.
55:24
So there is a few things if you do look at it like from schools
55:27
encouraging more girls to take up STEM subjects like science,
55:30
technology, engineering, math. Because if you look at like in
55:33
the UK, if you look at the statistics like every year about
55:37
20 to 30% of the students that go into these degrees are women.
55:42
And then if you look at the ones that graduate and come out,
55:45
24,000 women graduate from a STEM subject every year in the
55:49
UK. Like the workforce in the UK is
55:52
like what, 16,024,000 like. Tech is booming.
55:57
The pipeline is like so small and it's like encouraging them
56:02
from schools. Super valuable.
56:03
But also in the workplace, if there is, like, you will see it
56:08
that engineers, data scientists will be four times more male
56:12
data scientists than females. Like just being an ally for
56:15
women there, like if they're in a meeting and someone's like
56:18
talking over them, like to speak up and be like, hey, hang on,
56:21
someone speaking like, that's hugely important.
56:23
I know it's like the small things like that can make a big
56:26
difference because I do. I do feel like it can be a
56:29
really good place to work and learn with people, even if the
56:33
diversity isn't 5050 or if it isn't perfect.
56:38
If people are kind, courteous, and helpful, it shouldn't matter
56:42
what your gender is. Yeah, I just I.
56:46
Know from organising the conferences.
56:49
That. It is.
56:54
I try really hard to make sure that we get strong
56:57
representation. From everybody.
57:01
In different colours and genders and ethnicities, whatever you
57:06
whatever it is and. It it's a lot of work to go out
57:11
there, but I think that. There's some incredible work
57:15
being done by females in the field right now and it is.
57:19
It's super cool to see and so. Whenever we get the chance, I'd
57:23
love to highlight it, and this is actually.
57:26
As a coincidence, this is my second podcast today and the
57:30
last podcast that I recorded was also with a a female guest.
57:34
So it's super cool to see that. Yeah.
57:37
But I think it's good because if you are spending that time to
57:40
get like presenters at the conferences that are all
57:43
diversities, it means when the people go to attend and see
57:46
that, they'll be like, oh, that's cool, that's like someone
57:48
that kind of represents me on that stage.
57:50
Oh, maybe I could be more technical in that space and it
57:53
kind of inspires. And like, it's that natural,
57:56
hidden role model that people don't think about.
57:58
But it's hugely important. People feel connected and they
58:01
see. Things that they can relate to
58:04
and it's like really years that use another Shemale podcast
58:07
interview. I always so I.
58:10
Felt really weird. I'm just gonna.
58:13
Be straight with you. When I would ask different
58:19
people, hey like do you know anybody that would be great to
58:23
speak? And then so some, some people
58:27
that I was close with, they would be like, hey, do you know
58:29
anybody that's great to speak? And then like preferably female
58:32
because I already had, you know, 90% males that I knew and I just
58:37
know a bunch of males. And so we already were.
58:43
Over indexed on the male gender, right.
58:46
And then when I had when I I didn't have to, but I wanted to
58:50
find more females And so trying to find females, it was just
58:54
like I felt so awkward trying to be like, so I remember we were
58:59
in this call with NVIDIA and we were talking to them and it was
59:03
like NVIDIA is a gigantic company.
59:06
They've definitely got females that work there, right?
59:08
And so we're talking to them and they're like, can we.
59:10
Yeah, can we speak? And I'm like, you can totally
59:12
speak, but. It has to be a female that
59:15
speaks because we have to highlight like some incredible
59:19
people from your company that are females and and it just
59:24
feels so awkward sometimes when I when I ask that right.
59:27
And it's almost like, but you're not doing that to be, like rude
59:32
to the male employees. You're you're trying to be like,
59:35
hey, I'm trying to break Norm and the stereotype so that
59:40
people don't think that all these companies are just guys
59:44
and you're trying to do that in a really courteous way.
59:47
Um, I will say I am an ambassador for women in data and
59:51
it's a network of around 35,000 women that work in data, mainly
59:56
in the UK, but they have expanded around that some in
59:59
Europe and America. A situation looking for ohh yeah
1:00:03
perfect. I go right to the course that
1:00:05
many things. I'm sure you will have plenty
1:00:08
women. I would really love to talk
1:00:10
about my life and to get more diversity and leave out the
1:00:13
awkwardness of like so many other types of people.
1:00:17
It was I'm. Telling you it was like, yeah.
1:00:20
And I had, there's two other women on our team and I had to
1:00:24
like, ask them, like, is it weird that I'm asking that when
1:00:27
I was after the meeting? Like, is it weird that I'm like
1:00:29
forcing it to only be a female? No, I think that's something in
1:00:34
the group. We are very big on that.
1:00:35
If we do create panels and stuff like that, going to have it
1:00:38
diverse as possible so that it's not.
1:00:40
Sometimes it's it's hard to do that because especially if it's
1:00:44
a really senior leadership panel and you're like ohh gosh,
1:00:47
there's 10 people and we can only find one or two women
1:00:49
because unfortunately they they don't exist yet in that space.
1:00:54
Yeah, but I have. To say also.
1:00:56
Like attended the conference which was women speaking women
1:00:59
plus speakers only and and and men were allowed to to like just
1:01:03
attend the conference And even from my point of view although
1:01:06
he was way more exclusive, way more inclusive, sorry than what
1:01:09
you would have usually and it was actually very inspiring as
1:01:12
well to see like yeah it was amazing.
1:01:15
Ohh no. Well it was like.
1:01:17
I'm totally fine with that, but I just thought he was, yeah,
1:01:20
amazing. And he was like a very different
1:01:22
vibe actually to to like usual conference and I really, really
1:01:26
enjoyed it. So I hope we'll have also more
1:01:27
of those later. Yes, now.
1:01:30
We definitely will. We when?
1:01:33
It comes to the virtual conference that we did.
1:01:36
I would say we had probably 30% female speakers and somebody
1:01:42
wrote in the feedback form like, oh, it's so good to see 5050.
1:01:46
Like people felt like just with the 30%, it felt like there's
1:01:50
way more than normal. Exactly.
1:01:53
Because it it can be such a gap that sometimes it is so easily
1:01:58
to fall into not the trap, but just again, that norm, that
1:02:03
stereotype. I went to the Pie Data
1:02:05
conference in London recently and I think there was.
1:02:08
I think I remember two women speaking too.
1:02:12
I imagine for that planning it might not have been actively
1:02:16
sought to have a balanced speaker set, because if you
1:02:20
don't actively seek it, it ends up being very skewed.
1:02:23
Yep, exactly. That's what it is.
1:02:25
It's too easy to not. Have it be very important and
1:02:30
and if you don't actively try to find females because I I
1:02:35
remember I was getting so stressed that I would we would
1:02:37
be looking at the speakers and it's just like man man man man
1:02:40
man and I'm sitting there like we have to find more females we
1:02:43
have only men. Like I cancelled anybody that
1:02:47
would reach out to me and say hey is there another speaking
1:02:50
spot? And it was basically towards the
1:02:52
end it was like no, unless you are female and or you like your
1:02:58
cofounder or whoever is female. Like we cannot have anybody else
1:03:01
or the other piece and this is. Probably make me sound horrible,
1:03:05
but it's like or you're paying us and so if you're going to
1:03:08
sponsor or you're female then I can find a spot for you but if
1:03:11
you're just you know but I'd be interested it costs money to do
1:03:14
the conference and like having people that are being the
1:03:18
sponsors like it wouldn't happen otherwise.
1:03:20
So if it's kind of like a fair shout in terms of like you
1:03:25
financials need to be covered, but you understand I, I.
1:03:29
Appreciate your understanding of it.
1:03:31
I felt I, you know, it's like, I felt so bad sometimes when it's
1:03:35
like, yeah, I I do want to find a way to get you in there, but
1:03:39
umm, no, we're like saving these spots because I'm looking for
1:03:44
female speakers, so. All right.
1:03:47
We made like way bigger of a thing out of this then we needed
1:03:50
to kind of that's yeah, that's how we can end it.
1:03:56
This has been awesome. I really appreciate you talking
1:03:58
to us about everything from your your path into tech, the GCP
1:04:05
revolution that you're going through, the way that you.
1:04:09
Make sure you don't have problems at the bank and how
1:04:13
careful you are and then what you do to mitigate problems when
1:04:16
they do arise and. In the end there the.
1:04:21
The whole. Advice that you would give to
1:04:25
people coming into the field and how how it has been in your
1:04:30
experience to be a female in this field and how we can
1:04:33
hopefully have more of them exactly.
1:04:37
We're definitely moving in the right direction and that's cool
1:04:39
space to be in. So thank you so much for having
1:04:42
me. I really enjoy it.
1:04:43
It's really good podcast and it's been amazing.
1:04:46
Thank you very much. This is Skyler.
1:04:51
I lead machine learning at Health Rhythms.
1:04:54
If you want to stay on top of everything happening in ML OPS,
1:04:57
subscribe to this podcast now. Now.
1:05:03
Now. Now.
1:05:06
Now. Now.
1:05:10
Now.