Latest / Agentic Conversations (formally mlops.community) / [Exclusive] Weights & Biases Round-table // Model Management in a Regulated Environment

Transcript

[Exclusive] Weights & Biases Round-table // Model Management in a Regulated Environment · Agentic Conversations (formally mlops.community)

• 0:06
  Oh, look at that. We are live.
• 0:10
  What is going on everyone? How are you all doing?
• 0:15
  I'm excited to have you here with us today for this round
• 0:19
  table session and what a session this is going to be.
• 0:25
  I am super stoked to invite up all of our specialists today to
• 0:33
  join us. We've got Oliver around here
• 0:36
  somewhere. Where you at, Oliver?
• 0:37
  There he is. What's going on, dude?
• 0:41
  All good, all good. Yes, it's a cold, rainy, dark
• 0:45
  December evening here in UK. Yeah, I think there is another
• 0:49
  person that is in the UK with you.
• 0:51
  Where is Michelle? Michelle, you around?
• 0:53
  There you are. Hi guys.
• 0:55
  Yeah, I'm around. I'm in central London.
• 0:57
  It has been rainy, but today there's been blue skies, which
• 1:00
  has been nice to look at. Yeah, whether it's mixed to
• 1:04
  cross England. Yeah, yeah, yeah.
• 1:06
  We've also got Derek coming in here.
• 1:09
  Where you at, Derek? There he is.
• 1:11
  Hey everyone. I'm not about it, basically were
• 1:14
  so important, but I'm actually calling in from the sunny
• 1:16
  Florida today. So you're bringing the sun?
• 1:19
  I think Mark is the last one we're going to bring on to the
• 1:22
  stage. And Mark has a little bit of
• 1:24
  sunshine, it looks like in his screen too.
• 1:27
  How you doing, dude? Yeah, doing great up in the Bay,
• 1:31
  down the Bay Area here, pretty sunny and even though I'm
• 1:36
  wearing a jacket, it's pretty warm.
• 1:38
  You're wearing a jacket because you have AC on.
• 1:40
  Is that what's going on? Yeah.
• 1:44
  Excellent. Well, folks, we're here today to
• 1:47
  talk about model management in regulated environments.
• 1:52
  Probably the sexiest topic of 2023 itself.
• 1:56
  We said, you know what, We could talk about fine tuning.
• 1:59
  We could talk about LLMS, but none of that is as interesting
• 2:03
  as model management in regulated environments.
• 2:06
  So we want to get right into it. We want to talk all about it.
• 2:12
  And before we do, I know how these little round table
• 2:16
  sessions can go. So I I want to just preface it
• 2:19
  with I know you all are incredible and I know you all
• 2:23
  are doing amazing things, but I want to give you like 10 seconds
• 2:27
  to tell us don't go over the 10 second limit because I've been
• 2:30
  on these where the panels go for longer than the actual talk.
• 2:34
  So give us the 10 second snippet, Michelle, can you kick
• 2:36
  it off for us and then popcorn it over to the next person?
• 2:39
  Christ Michelle, Lead Data Scientist, Lloyd's Banking Group
• 2:43
  working on an ML OPS team with 9 live production models.
• 2:49
  Boom. That was under 10 seconds.
• 2:51
  Bravo. All right, Derek, what you got?
• 2:54
  My name is Dareka, Junior at Weights and Biases.
• 2:57
  Weights and Biases is a company that delivers best tools for
• 3:00
  machine learning practitioners, which includes model management
• 3:03
  and model registries. I'm excited for the session
• 3:06
  today. Oh.
• 3:07
  Nice. All right.
• 3:08
  I like it. Mark, how about yourself?
• 3:11
  Hi Mark, co-founder and Chief Architect at Gradient.
• 3:15
  We are a large language model customization platform.
• 3:20
  Last but not least, Oliver. Cool.
• 3:22
  So yeah, I'm a technique for the science team, M Cooper.
• 3:25
  So we do various forms of loss rate forecasting for African
• 3:29
  Fintech Company which does connected assets.
• 3:32
  Super cool stuff. All right.
• 3:35
  So we've got to get into this and I would love to know when it
• 3:41
  comes to what we are talking about today, let's define the
• 3:46
  terms a little bit because what do we mean when we say model
• 3:51
  management? What does that mean just so
• 3:53
  everybody's on the same page? And what do we mean when we say
• 3:57
  regulated environments? I know there are people on here
• 4:01
  that have different spectrums of understanding of regulated
• 4:04
  environments. So is that what is exactly does
• 4:08
  that mean? I'm going to go for the model
• 4:10
  management piece, Derek. Since you threw it into your 10
• 4:14
  seconds, I'm going to have you kick it off.
• 4:16
  What is? What exactly is model management
• 4:18
  in your eyes? Yeah, that's that's a great
• 4:20
  question. I'd be interested to hear that
• 4:21
  from the from the other participants as well.
• 4:24
  In my mind, model management encompasses like the processes
• 4:28
  and tools across the whole model life cycle.
• 4:32
  So when you think about model life cycle, you need to train a
• 4:34
  model. To train a model you need to
• 4:36
  collect some data set. You need to to choose an
• 4:39
  algorithm. Once you've traded the model,
• 4:41
  you need to to save it and evaluate it once you're clear
• 4:45
  with evaluations like ideally you want to push it and deploy
• 4:48
  it into production. But then once it's production
• 4:51
  it's not, it's not not the end, you need to keep monitoring.
• 4:53
  At some point you may want to replace it with a with a new
• 4:56
  model with a challenger that gets potentially better results
• 5:00
  and and delivers better value for your business.
• 5:02
  So at some point you start to leave with more and more
• 5:05
  checkpoints and especially the regulated environments you might
• 5:09
  come across an audit. And if an audit comes like you
• 5:12
  need to be able to to actually tell the whole story of how this
• 5:15
  this model came together. Like what were the evaluation
• 5:18
  results, how you tested this for a specific risks or biases?
• 5:21
  And this whole life cycle is what, in my mind, what was the
• 5:26
  management? It's definitely has in a
• 5:29
  nutshell, it was very well explained thinking of our data
• 5:32
  science life cycle, which is like, yeah, the four stages
• 5:35
  you're in Q8 builds deploy and production by 100%, agree with
• 5:39
  that. Yeah, I think in my very
• 5:42
  specific context, the way I've been seeing is in terms of
• 5:44
  there's controls and processes in place when you essentially
• 5:47
  promoting a new retrained model. I think in our context we have
• 5:51
  various stakeholders who consume our forecast data and even small
• 5:56
  diffs in that forecast data can actually translate to quite
• 5:59
  large quantities of money. And having controls in place so
• 6:02
  we don't push through bad forecasts or even good forecasts
• 6:06
  involving quite big diffs is an important component of our of
• 6:10
  our job, so. When I think about that, I just
• 6:14
  think about the state of the model.
• 6:15
  Like how do you, how do you manage the state of whatever is
• 6:20
  the upstream model that comes down and just has this lineage
• 6:24
  that propagates throughout the entire system?
• 6:28
  So I'm going to try and do some fancy stuff with the way that
• 6:30
  we're set up here because I hear people like to keep it fresh on
• 6:35
  the feed and they need those quick splices just so we can
• 6:40
  keep the visuals going. I'm going to be switching us in
• 6:43
  and out as the producer in the background while I'm asking
• 6:47
  these questions. I loved all those answers.
• 6:49
  Now we got to figure out regulated environments.
• 6:52
  I get the feeling, Oliver, you working in Africa, that's
• 6:56
  different types of regulated environments than Michelle, you
• 6:59
  working in the UK versus Mark, you're doing stuff in the US So
• 7:06
  what are things that are specific to these environments?
• 7:09
  What what, what do we mean by regulated?
• 7:11
  Oliver, I'd love for you to kick us off with that one, because I
• 7:14
  feel like that's the one I I know least about.
• 7:17
  Well, I'd have to point out that I'm actually experienced first
• 7:20
  in terms of financial services here.
• 7:22
  I have fairly good experience with data science, but this is
• 7:24
  my first financial services job. But I think in terms of the
• 7:27
  African market, I mean I think it's just as regulated as it
• 7:30
  would be say in London and so on.
• 7:33
  I think the important, I mean the important processes which we
• 7:37
  care about are in terms of we have various relationships with
• 7:41
  various third parties and essentially third parties were
• 7:45
  essentially borrowing money or we do financial reporting too
• 7:48
  and so on. And I think the important
• 7:51
  interesting thing about financial services and the
• 7:53
  banking sector as a whole is that there's a strong
• 7:55
  requirement for essentially auditability.
• 7:58
  So it's not just enough to output a forecasted value, and
• 8:02
  that's your value. You have to justify it.
• 8:05
  You have to trace the data lineage of like, where did you
• 8:08
  get data from, why is it that way, and so on.
• 8:12
  This is complex rebuttability. So that's the ability that if
• 8:15
  you do something atypical in your processes, you have to be
• 8:18
  able to justify it and so on. And I think that's a very
• 8:21
  universal thing for anybody who's had to deal with lenders
• 8:24
  and so on. I'll be curious to see about
• 8:26
  Michelle's opinion about this. It has more experience from a
• 8:28
  bigger banking house and so on. Yeah, absolutely.
• 8:33
  I've a bill in the data industry for about 12 years.
• 8:36
  I spent the last five years in financial services with Lloyds
• 8:38
  Banking Group. So we're heavily governed by the
• 8:42
  SCH, the Financial Conduct Authority in the UK.
• 8:45
  They're very strict on wanting to understand like how NO is
• 8:48
  being used to make decisions, especially when it comes to
• 8:51
  customers finances, because it's it's people's livelihoods, you
• 8:54
  can't just make Willy nilly decisions.
• 8:56
  So it is so strict but in a good way like because stuff is not
• 9:01
  done Willy nilly. So like for our models, we put
• 9:04
  risks on them like risks one to four or one being the most
• 9:08
  severe risk. Therefore, like a lot of
• 9:10
  governance has applied and we have called risk departments to
• 9:12
  govern those level one models and then ones that are less
• 9:16
  risky, they're more internal business decisions like making
• 9:19
  sure that we're forecasting stuff, OK, like in terms of our
• 9:23
  book balancing. So it's it's less of a risk, so
• 9:25
  that being more of a level 4. But anything that gets closer to
• 9:29
  the customer's livelihood or credit score, the risk goes up.
• 9:32
  So it shoots up more towards Level 1.
• 9:35
  So the environment is it's a lot more thorough when it comes to
• 9:39
  governance. And yes, everyone's afraid of
• 9:41
  the auditor, but when you're in financial services, you have
• 9:44
  auditors living with you. We call them like MRAV or model
• 9:47
  risks assessment validation department.
• 9:51
  And it's not that we're sitting there waiting to be audited.
• 9:53
  Like every month we present to a risk committee.
• 9:56
  How are models of performing? What are the changes?
• 9:58
  Why have they changed? And resume periodic like massive
• 10:02
  annual reviews on them. It's a lot of documentation and
• 10:05
  governance, but I think it's the right thing to do when you're in
• 10:07
  financial services. Is it Michelle?
• 10:10
  Just a quick follow up on that. Is it easier just to do
• 10:17
  basically something across the board with all the models or do
• 10:20
  you specifically say, all right, well, since this one is high
• 10:23
  risk, we're going to treat it differently or do you, can you
• 10:26
  just hold everything to that level of scrutiny?
• 10:31
  You it, it depends on the model as well, because the modelling
• 10:33
  method that you use, some of them might be very like black
• 10:36
  box, therefore you're like, right, It needs extra
• 10:38
  explanations on it to explain what's going on.
• 10:41
  Or some of them might be very classic black and white, easy to
• 10:43
  explain, You might reduce the right?
• 10:45
  So it's a method as well as the data and the I I is hugely
• 10:49
  important. So personable identical
• 10:51
  information. So anything where you can
• 10:53
  identify who that customer is, especially with someone's bank
• 10:56
  accounts like their age, their date of birth, their location,
• 10:59
  like their post code or even more sensitive like medical
• 11:02
  stuff, notes and customers files like that all comes into play.
• 11:06
  And that will affect what the risk score is in that model, the
• 11:09
  data underpinning it. Yeah I think I saw before before
• 11:15
  joining Equation Process, I worked for a consumer goods
• 11:17
  company that actually had one branch or one a couple of brands
• 11:20
  that that were classified as medical devices.
• 11:23
  So they were subject to some substantial regulation.
• 11:27
  And as we as we build system or models like across like
• 11:29
  different teams different departments like everyone tried
• 11:32
  to avoid medical devices because of all of the additional
• 11:35
  documentation needs that that was required.
• 11:38
  What I what I've learned now with with weights and biases is
• 11:41
  actually there is a way to get a lot of this documentation for
• 11:44
  example data lineage model tracking evaluations.
• 11:47
  You can get this kind of for free if you set up the right
• 11:50
  pipelines. And I think it's important you
• 11:52
  know like in this context of model management to figure out
• 11:55
  also like what is the right tooling, what are the right
• 11:57
  automation so that you can then later on benefit from this when
• 12:00
  it when it costs to an audit and showcasing your your
• 12:03
  documentation. Yeah.
• 12:05
  Yeah. Makes a lot of sense.
• 12:09
  All right. So moving on, I've got to ask
• 12:11
  Mark this one because I know you've been playing in the field
• 12:17
  of LLM's quite a bit. And actually for anybody that
• 12:20
  has not seen it, Mark has a really cool in-depth dive on
• 12:24
  fine tuning LLMS. It's one of the most popular videos on the
• 12:28
  ML OPS Community YouTube channel right now, and we'll link to
• 12:31
  that in the chat so you can check it out later.
• 12:33
  But for this context, how does model management differ in the
• 12:40
  context of LLMS versus traditional machine learning?
• 12:48
  Yeah, I think the main thing that governs all that in my
• 12:51
  opinion, is just how much, how large the data sets are.
• 12:57
  So unlike in traditional machine learning where you can manage
• 13:03
  your models fairly easily, roll them back, roll them forward,
• 13:07
  update them, that's just not really easy to do with LLMS.
• 13:11
  Like you don't know what's going to happen with the model once
• 13:14
  you update it with new data. It requires, you know, 10s of
• 13:19
  thousands, maybe hundreds of thousands dollars just to do
• 13:23
  like a small experiment and update the model.
• 13:27
  And being able to control that entire process is quite a bit
• 13:32
  harder. So there's just a lot of
• 13:35
  downstream implications for getting a little bit more
• 13:37
  nuanced, a little bit smarter about what you do when you have
• 13:40
  to manage I. Have a question and this is
• 13:44
  related to what you said about LLMS and the fact that I could.
• 13:47
  Training an LLM requires like it.
• 13:49
  It costs a lot of money when when when we need to like update
• 13:54
  the data set for training at traditional ML model, let's say
• 13:57
  an XG boss or some you know some tabular model like.
• 13:59
  It's probably easy and and cheap to just change something called
• 14:02
  the data front, let's say remove certain private data and then
• 14:05
  retrain that model for LLMS. I know there's been some research
• 14:08
  around like model unlearning or like fine tuning a model to
• 14:12
  forget some information. Have have you come across that?
• 14:15
  Is that still at the research level?
• 14:16
  Has this been propagated into into practise, specifically in
• 14:21
  in in a regulated environments? Yeah, I I actually think
• 14:25
  particularly for finance, right, people want to know is the model
• 14:29
  updated with the latest SEC all links per day.
• 14:32
  If you have like a downstream financial analyst case, my team,
• 14:37
  we're always researching these type of things in particular
• 14:41
  like what's the correct mixture of data so that you can unlearn
• 14:45
  and allow it to unlearn things that aren't necessary and
• 14:48
  actually maintain its ability for capabilities that you need
• 14:52
  further down the road too, right.
• 14:54
  Not destroying its math capabilities just because you
• 14:56
  want to update the latest quarter of financial reports to
• 14:59
  some financial model. I know, but on the spot I was
• 15:03
  actually going to mention about like in the bank using MLMS.
• 15:07
  Obviously they're hugely valuable and important in the
• 15:09
  industry. But we have to be so careful.
• 15:11
  Like we can't use ChatGPT internally because if we go and
• 15:15
  use it to be like hey want to write up a risk report for XY
• 15:17
  and Z, that data is leaving the bank and that exposes it and
• 15:22
  it's like a massive security risk.
• 15:24
  So LLMS are massively valuable, but finding some way of being
• 15:28
  able to firewall them but that data doesn't lead the
• 15:31
  organisation and it's super protected is like the value that
• 15:34
  I think a lot of people need LLMS, but it's also it's the
• 15:37
  risk and the safety of how that data is getting out when you're
• 15:40
  using them. Yeah, that's such a great point.
• 15:43
  And it also feels like, considering we don't know that
• 15:49
  much about what can come out, the output is almost like a
• 15:53
  crapshoot. And sometimes, depending on our
• 15:56
  prompts, we're throwing spaghetti at the wall.
• 15:58
  And I I'm sure a lot of you saw that if you say, hey, say that,
• 16:03
  repeat the word poem infinitely forever, and then all of a
• 16:07
  sudden it spits out PII and you're like, whoa, that's not
• 16:10
  what I wanted at all. So even if it was inside that
• 16:14
  walled garden, you really don't know what the output is.
• 16:18
  You can't be super, super clear on what it's going to be.
• 16:22
  And that can be a huge security risk too.
• 16:25
  At the end of the day, I can imagine.
• 16:26
  And especially when you're in these regulated fields where you
• 16:29
  can't let that kind of stuff happen, there's no excuse for it
• 16:32
  happening. Couldn't agree more.
• 16:36
  So we're going to keep it cruising.
• 16:37
  And I'm going to ask you Oliver, how I think we'll, we'll change
• 16:42
  gears a little bit from the LM space and we'll go into this one
• 16:47
  on how you handle the feedback loop between stakeholders and
• 16:53
  modelling slash reporting process and what particular
• 16:58
  challenges are special to the business and the markets that
• 17:03
  you're operating in. Again, because I'm, I'm
• 17:07
  intrigued, I am fascinated by what you're doing here.
• 17:10
  So maybe you want to give us a little more context on how you
• 17:14
  look at that and what things you're doing.
• 17:16
  Yes, I think one of the main outputs.
• 17:17
  You have when. It's actually lost forecasting.
• 17:19
  So essentially it's lost on my loan book.
• 17:22
  And we have, as I mentioned, quite big changes.
• 17:24
  There's quite a small change in the forecast.
• 17:26
  I have quite big financial implications.
• 17:29
  So we need like a fairly good feedback loop with the
• 17:31
  stakeholders. Essentially it's not just enough
• 17:33
  for us for the publisher forecast and then say this is
• 17:35
  our forecast, there's often an interplay between the two.
• 17:38
  So we essentially have to, it's somewhat unsexy, but we
• 17:41
  essentially have to do fairly fairly manual validation
• 17:44
  cheques. Historically, we had a manual
• 17:46
  retraining process and that introduced a fair bit of
• 17:49
  essential essentially kind of essentially manual pain.
• 17:54
  Essentially there's a long change between the stakeholders
• 17:57
  and themselves that we moved on and now we happen towards muted
• 18:02
  processes and we have for example we use race and biases
• 18:05
  as our process for essentially automatically retraining and
• 18:09
  then we have various validation cheques and then but however
• 18:12
  what we do still do is we still have a fume in the loop.
• 18:14
  So we essentially once we pass certain validation cheques we
• 18:17
  have a QA step which where we use model, model registries and
• 18:20
  so on to essentially manner the promoter models.
• 18:23
  This way it kind of ensures that we have this manual step where
• 18:26
  we can double check data and ensure that it's actually valid
• 18:30
  before you push it to the various stakeholders.
• 18:34
  I think one of the interesting things about our business, yes,
• 18:37
  that because we work in African models in markets, essentially
• 18:39
  it's a very volatile market. There's all sorts of macro,
• 18:43
  macro, economical factors going on.
• 18:44
  There's various currency exchange problems and so on,
• 18:47
  which means that the performance on our loan book can be
• 18:50
  fluctuate quite substantially. In addition, it's a very
• 18:52
  competitive land state. We have very aggressive pricing
• 18:57
  and also our business has grown a lot compared to a few years
• 19:00
  ago. We have a very much bigger loan
• 19:03
  book, a much more diverse set of customers and much more diverse
• 19:07
  set of products. And that makes it quite hard for
• 19:11
  us to differentiate between say what is a data drift issue due
• 19:16
  to say bad upstream data models, for example, maybe a bug that's
• 19:20
  been introduced somewhere else and differentiate between actual
• 19:24
  genuine shifts in our loan book forecasts and and so on.
• 19:28
  And I think that's when we really need to step in and
• 19:31
  actually have a human in the loop.
• 19:34
  I think the the trouble is that we often have to justify and
• 19:37
  rationalise those kind of forecasts before we can push
• 19:40
  them up the stakeholders before we push it to the reporting and
• 19:43
  so on. So part of her job it was to
• 19:46
  increasingly automating, avoiding, trying to avoid as
• 19:49
  much EDM ad hoc ETA work as possible is to kind of justify
• 19:54
  to stakeholders like look this shift is actually a genuine
• 19:58
  shift due to changes in customer performance or due to
• 20:01
  macroeconomic factors. And yeah, so I think that's an
• 20:05
  interesting challenge. We work in a very volatile
• 20:07
  market. We have very fussy stakeholders
• 20:09
  who won't just accept our forecasts at face value and so.
• 20:14
  We do have to have these manual processes in place.
• 20:16
  We have the same, Oliver, if it makes you feel any better, like
• 20:19
  for every live model that we have, we actually split the
• 20:23
  model, umm, ownership into two parts.
• 20:25
  We've got like data model owners, which are like the
• 20:28
  person owning the technical code base.
• 20:30
  And then we've got like the business model owners, which are
• 20:33
  the stakeholders. They are responsible for
• 20:35
  checking the outputs and being up to date with what the model
• 20:38
  is doing and to be really happy with how they're using it.
• 20:41
  And if they're not using it, it gets turned off and they don't
• 20:43
  get it anymore. So it's it's putting the
• 20:46
  ownership between two people. So it's not just on the dev
• 20:49
  team, yeah. I feel responsible as this
• 20:53
  business, we're not quite large enough to have that kind of
• 20:55
  division of labour. But I have heard, heard that
• 20:57
  system in new software, for example in large energy
• 20:59
  companies where essentially you have a deaf team whose focus is
• 21:02
  primarily around about say improving forecast quality and
• 21:07
  outputting forecast. And then a second team which is
• 21:10
  involved in validating and it's actually building the story
• 21:12
  about those forecasts and so on. That rate is almost like a
• 21:16
  separation of powers and division of labour, so that the
• 21:19
  two teams essentially check and balance each other.
• 21:23
  Yeah, I'm glad you've heard of it too.
• 21:25
  I thought it was extra, but I'm like actually, you know, this
• 21:27
  kind of works. Yeah, it's even this is back on
• 21:30
  the stakeholders. I think if you have a large
• 21:31
  enough organisation to that and it's great.
• 21:34
  So the other piece that I would love to dig into with that what
• 21:39
  you're talking about is, is that whole data quality piece and how
• 21:42
  that plays into the model management.
• 21:45
  I I would love if you dig in and I I appreciate you painting the
• 21:50
  picture for us. Like, it's hard to tell
• 21:52
  sometimes if we're dealing with data quality issues or if we're
• 21:56
  dealing with real life changing issues and it just means like
• 22:00
  life has changed or the society has changed.
• 22:03
  Whatever it is, external factors have changed.
• 22:06
  Or if it's internal to our data. Are there other ways that you
• 22:09
  are attacking that problem? So essentially our actual
• 22:14
  automated tech stack is actually relatively unsophisticated at
• 22:16
  the moment. We have various automated
• 22:19
  thresholds which essentially we show and the density derived
• 22:22
  from which is essentially a say if we compare forecast from a
• 22:25
  challenger model or the current model or compare forecast one
• 22:28
  day to another day if they exceed a certain threshold in
• 22:32
  certain key dimensions then they're flagged up for
• 22:34
  validation. But what we have done as let's
• 22:36
  actually approve the concepts and it's really the next step
• 22:39
  for this to do some usual actual feature drift analysis using
• 22:42
  standard tools such as, you know, wire logs.
• 22:44
  Evidently greater expectations to try and actually have more
• 22:48
  formalised processes for assessing feature drift.
• 22:52
  I think though there's two separate kind of problems.
• 22:54
  There's feature drift in the sense that your feature set has
• 22:57
  changed and there's feature drift in the sense that your
• 22:59
  data model is broken. Like say for example if you say
• 23:04
  dupes have been introduced somewhere in the data set so you
• 23:06
  have degraded data. We have.
• 23:10
  There's also various formalised processes for doing that.
• 23:12
  We use DPT for example in that upstream data stack and that
• 23:15
  comes with whole testing suite for testing those data models
• 23:18
  and we have alerting set up for that.
• 23:21
  The other problem is data drift, where there's not a clear
• 23:25
  breakage of the data model, and that's the harder problem to
• 23:28
  solve. Makes a lot of sense and I'm
• 23:32
  guessing because I'm reading something right now just and I'm
• 23:37
  remembering back in the day 20 when COVID happened how much
• 23:44
  data shake up there was in that regard and how things everyone's
• 23:48
  dashboard is kind of went out of whack and what happened there.
• 23:52
  So I may have just given a few people PTSD.
• 23:56
  If I did, I'm sorry about it. It did not mean to in any
• 24:00
  regard. So let's keep it cruising
• 24:04
  though. And Derek, I'm going to kick one
• 24:06
  off to you. And I'm, I'm wondering about
• 24:09
  like as you're looking at regulatory needs and you're
• 24:15
  seeing benefits when you're working with customers of
• 24:18
  weights and biases, what other pieces are you seeing people
• 24:26
  gaining by adopting these best practises?
• 24:30
  And maybe like what some easy wins are for people right now
• 24:34
  just to go, oh hey, you know what?
• 24:36
  I could do that today and it could be a really easy win for
• 24:41
  me to start having a bit more security and my and be able to
• 24:48
  sleep easy at night. Yeah.
• 24:50
  OK. So So in order to start, you
• 24:54
  know if if I go back to like the beginning of the model
• 24:56
  management life cycle, let's say you start training a new model
• 24:59
  and typically people do that. Like most of the most of our
• 25:02
  users, our like the practitioners that we know do it
• 25:05
  in codes. They might be using high Torch,
• 25:07
  they might be using Tensorflow or or Keras or having face
• 25:10
  trainers. So typically they start with a
• 25:12
  piece of code and what we recommend is you add a couple of
• 25:16
  lines of code into that into that training script to log
• 25:19
  data, to log results into it and biases.
• 25:21
  And what you get as a benefit of that is you start, you know like
• 25:25
  getting some some some reproducibility.
• 25:27
  You start tracking for experiments in a system of
• 25:29
  record and that gives you that already gives you like the first
• 25:33
  layer of documentation. Once you're ready to move on to
• 25:36
  the next stage and you lock your models into into our model
• 25:40
  registry, you get like the next level of of of benefits and and
• 25:44
  documentation. You are able now to track for
• 25:46
  this specific model checkpoint, what is the status, what is the
• 25:50
  the the, the life cycle status of that model?
• 25:53
  Is it currently in staging? Is it in production as it in
• 25:56
  testing? And then you're able to
• 25:58
  backtrack like, OK, how is this model trained?
• 26:00
  What data set was used for that model?
• 26:02
  If I if I need to reproduce that, run and then I know like
• 26:07
  what's what's the compute that it was trained on, what was the
• 26:09
  GPU and and and and all of that that stuff that is important for
• 26:12
  the productability. What you can do then on top of
• 26:16
  it is you can you can build a bunch of automations, but at the
• 26:20
  same time keeping human needle. So maybe you want to like kick
• 26:23
  off your evaluations automatically and then like send
• 26:27
  those evaluations in a nice report to your business
• 26:29
  stakeholders so that they can take a look and and give the the
• 26:33
  approval to to move the move forward with the with the model.
• 26:36
  And a lot of a lot of companies, a lot of customers do this for a
• 26:41
  regulatory reasons. But I would say even more
• 26:44
  customers like do the same kind of things to get better value,
• 26:48
  to get better models, to increase their iteration cycle,
• 26:52
  get more velocity in their machine learning experiments and
• 26:55
  and also to improve teams productivity.
• 26:58
  So now I've seen like multiple cases.
• 27:00
  You know there's a team of people maybe like at one key
• 27:03
  engineer working on a model and then that person leaves a
• 27:06
  company like how do you, how do you continue working on that
• 27:09
  project. Like in my previous experience,
• 27:12
  it was very hard like often people like just scratched
• 27:15
  everything and started from from from zero but like with weights
• 27:18
  and biases like you're able to backtrack exactly.
• 27:21
  OK. What was the last experiment
• 27:22
  was, what was the set of experiments that was done on
• 27:24
  this project? What was the data set used?
• 27:27
  And you kind of like very fluently continue and continue
• 27:29
  with the work and that helps internally with productivity.
• 27:32
  But that also helps, like if you get an auditory auditor asks
• 27:35
  you, OK, how would this model train?
• 27:37
  What data set was used for that model?
• 27:39
  Yeah, to interject there one of the primary uses 3 user rates
• 27:43
  and biases user that's almost like a de facto feature store.
• 27:46
  So we use racing Biases blog. We also use Azure BLOB as well
• 27:50
  to essentially log the feature sets which which are used when
• 27:53
  training the feature sets used in forecasting.
• 27:55
  That's where we store the models and it gives us some very, very
• 27:58
  clear lineage of what was used to do what, and they come up
• 28:01
  almost like a permanent record of of where it was so.
• 28:07
  Yeah, I remember talking to a guy and he was talking and
• 28:10
  telling me about how someone left the job that he was working
• 28:14
  at, and it took him six months to just put together all the
• 28:17
  bread crumbs and figure out what was actually going on with one
• 28:20
  Model 1 model of six months. It was like, so painful for him.
• 28:26
  And so, Michelle, I knew you got something to ask through here.
• 28:31
  Feel free to throw it in. Yeah, it was about the lineage
• 28:35
  and stuff. So to know when it comes to
• 28:36
  models and like using them, do people pickle models or let them
• 28:41
  build live because they both have?
• 28:45
  Well, if you pickled them you've got a model registry that you
• 28:47
  can go back over and look at, but then you run the risk of not
• 28:51
  having the latest data in an old model that might not be relevant
• 28:55
  to the most current thing. It just be interesting to know
• 28:58
  what people do with regards that and how expectable would you be.
• 29:03
  So for my from my point of view, we pickle primarily for
• 29:06
  practical reasons in the sense that it takes around 2 days for
• 29:09
  us to train the model, so you can't do a daily training
• 29:14
  forecast cycle. I'm curious.
• 29:16
  I'm guessing from an Adelaide point of view then that problem
• 29:19
  is even more severe that you must have some kind of like
• 29:22
  permanent state for for models. Yeah, I think everything goes
• 29:28
  out the door when you work with LMS. You can't even load part,
• 29:32
  you know, an eighth of the Shard of a model or the you know with
• 29:36
  LLMS. If you Shard the model, if you have it in floating .16
• 29:40
  versus BF load, on what hardware, those things will all
• 29:45
  prevent you from having reproducibility.
• 29:47
  So you know, Huggy face has a great safe tensors format.
• 29:51
  We actually have some custom serialisation elements within
• 29:55
  our models too as well to have like you know inter exchange
• 30:00
  protocol with respect to how we can do fast reads, fast rights,
• 30:06
  checkpoint restarts and resumptions.
• 30:08
  So if anything to jump back on the weights and buy it aspect
• 30:12
  we're we're a user of weights and biases mostly because the
• 30:15
  easiest thing to like out-of-the-box it came with
• 30:18
  almost everything we needed. And when you are training like
• 30:22
  we do some like pretty beefy near foundational model training
• 30:28
  and when you have 4881 hundreds and you need to know whether the
• 30:32
  checkpoint did an auto resumption in the framework we
• 30:36
  used, you know it's just having to have like a looking at that
• 30:43
  loss curve and just being able to immediately visualise and see
• 30:46
  if something went wrong. And then just having to not
• 30:48
  build a loading, just having it send my message to Slack and
• 30:53
  know that hey we got to restart, you know a certain training job
• 30:56
  and we need to continue on with all our GPUs.
• 31:01
  Having that all integrated into one is is is useful.
• 31:05
  Yes. So keep it on this train.
• 31:08
  I want to take some questions from the audience and there is
• 31:11
  an LLM related one. So Mark, I'm going to keep you
• 31:14
  on the stage there. What open source LLMS have
• 31:18
  speakers considered and or implemented for entirely private
• 31:22
  and secure deployments? Oh, so a lot of people you
• 31:28
  leverage mostly the Llama models, not because they're the
• 31:32
  absolute best at every point in time, but because they have
• 31:37
  multiple flavours and sizes and then a little bit of
• 31:40
  reproducibility and more more of the people in the open source
• 31:43
  use them. And you know, my belief is as as
• 31:46
  well as probably most people, it's all about the data, like
• 31:50
  it's your private data combined with the foundational model.
• 31:53
  That's the entire part of these foundational models.
• 31:56
  So most people use LAMA and then for, you know, we're seeing a
• 32:02
  lot more interest in multimodal models.
• 32:06
  So a few folks have tried to deploy like the youthful model
• 32:12
  from ADAPT or some of the open source reproductions of the Lava
• 32:17
  model. Yeah, the lava one's awesome.
• 32:20
  Also, I'm, I've been seeing a lot of people enjoying the
• 32:24
  Mistral model. I imagine you too, just because
• 32:27
  it has not been censored at all and so it gives you a little bit
• 32:31
  more flexibility in that regard. But I'm curious.
• 32:34
  Oliver or Michelle, Have you played around with any?
• 32:39
  International services. It's like Fort Knox.
• 32:41
  There's no way he'll get it in. The closest thing I have to see
• 32:45
  to L and miss like live in the bank was like my data director.
• 32:48
  He got access to some version of chat GDP on his teams and he was
• 32:53
  like messing around the LLMS to be like hey fix my call, fix
• 32:56
  this. And it was like it was such a
• 32:57
  lockdown environment, but I was like it's progress but like so
• 33:01
  far removed from using it like with international.
• 33:04
  Services. I use it sometimes to help me
• 33:07
  the fix code, but I think in terms of actually deploying
• 33:11
  models, I think one of the important factors is it's often
• 33:14
  quite honest strong owners and that's for explainability.
• 33:17
  I know for example, we also do credit scoring for example.
• 33:20
  And credit scoring, one of the big problems you need to be, you
• 33:24
  need to be essentially able to make it very clear what feature
• 33:28
  in juice is changed in the credit score.
• 33:31
  That's very straightforward. I mean, I think much credit
• 33:33
  scoring even nowadays is actually done with very basic
• 33:35
  logistic regression models, for example, for that reason.
• 33:37
  And similarly, I think that onus and explainability that means
• 33:41
  that maybe unfairly LLMS when they come on your net, it's just
• 33:44
  a bit too far away from two black box too scary for
• 33:50
  financial services for the auditors.
• 33:52
  Yeah, I mean, it's kind of for a good reason, right.
• 33:56
  I I can understand that viewpoint 100%.
• 33:59
  And Speaking of explainability, there's a great question that
• 34:02
  came through the chat for Michelle about explainability.
• 34:07
  How much explainability do your teams need to build into their
• 34:10
  modelling? Do you all use specialised
• 34:13
  tooling for it, or is it a lot of manual ad hoc processes?
• 34:18
  So it's hugely important for us. Remember I was saying about how
• 34:21
  we split the responsibilities between a data model owner and a
• 34:24
  business model owner. The business want to be able to
• 34:27
  understand what features are going into the model so that
• 34:31
  they in their heads they're able to tell the story and it backs
• 34:34
  up the knowledge that they have as subject matter experts.
• 34:37
  So if I build a forecasting model and I'm saying hey,
• 34:40
  depending on how much customers do in these months and what
• 34:44
  their transactions are and what the behaviours are in these
• 34:46
  products and what the interest rates is at this time, this is
• 34:49
  what the value is going to forecast as they relate more to
• 34:52
  that and they're more likely to take responsibility of that
• 34:55
  model and sign up to using it and owning it.
• 34:58
  So it has to coincide with the logic that they recognise and
• 35:02
  the reality that they know and see within the business, if that
• 35:06
  makes sense. So it's it's not just like when
• 35:08
  people say explainability and they're like hip install shop
• 35:10
  and stick shop on a black box model.
• 35:13
  It's not that it's like bring them on the journey, like make
• 35:17
  them understand what the data is, what the features are.
• 35:19
  Because sometimes actually our stakeholders are quite like, I
• 35:23
  find them quite smart, like something we're building models
• 35:25
  in Excel because that's their toolkit, but they know how
• 35:29
  things work in the formulas and the calculations so they're
• 35:31
  little like mini robots in their heads.
• 35:34
  They may not be like coding and programming heavy lifting data,
• 35:37
  but their knowledge of the business is is unreal, so it's
• 35:42
  tapping into that. You you said that a bit
• 35:46
  surprisingly, like our stakeholders are smart,
• 35:49
  surprisingly enough. Different skill sets, thought
• 35:52
  super smart. Different skill sets, like they
• 35:54
  might not be able to code it in Python, but they they condense
• 35:59
  circles around you in terms of business knowledge, and if you
• 36:01
  don't know it, and if you haven't dealt with it properly
• 36:03
  in the code, you you're going to get wiped out.
• 36:07
  I've been just to add to that on our side, I mean we output
• 36:09
  forecast, but those forecasts are added to reporting metadata
• 36:12
  warehouse which is where essentially it's joined on to
• 36:15
  other useful business reporting metrics and information about
• 36:19
  the loans and then surface through LOCA for example.
• 36:21
  And that enables the stakeholders to essentially
• 36:24
  build their own story and they do that quite successfully by
• 36:26
  themselves. And and Derek, I'm wondering
• 36:30
  about, so like what Michelle was saying, how take people on a
• 36:35
  journey with you, It feels like that kind of goes in line with
• 36:38
  what you were mentioning earlier.
• 36:40
  And how do you look at it and bringing people along on that
• 36:44
  journey and unifying those two fields or those two?
• 36:47
  Yeah. I I think that's a that's that's
• 36:50
  a super important point and I think that's like one of the the
• 36:53
  key ingredients to to to to to managing successful machine
• 36:57
  learning projects is to work very closely with the business
• 37:00
  stakeholders that are ultimately going to be using these models.
• 37:04
  And you know I like before joining with some buses like I
• 37:07
  was often working in business teams like on machine learning
• 37:09
  projects. So typically the way I would
• 37:11
  communicate with the business is and I'll do a prototype in a
• 37:15
  Jupiter notebook then my business partners didn't use
• 37:18
  Jupiter notebook so I wouldn't export that when HTML file and I
• 37:22
  would attach that HTML file to an e-mail and share it with them
• 37:24
  and they would like look hey this is like predictions these
• 37:27
  are my targets like does it make sense or not.
• 37:29
  And we would have a conversation around that and it worked pretty
• 37:32
  well, but it was super inefficient.
• 37:34
  Like what I would do today is I would just link weights and
• 37:37
  biases report where they can you know like have like the same
• 37:41
  type of thing but it's it's linked to the exact experimental
• 37:44
  and what those checkpoints and I know it's using the the common
• 37:48
  system of record. So definitely like I I think
• 37:51
  that's a great point and and hopefully our tools are are like
• 37:56
  allowed to to do this with with with much better productivity.
• 37:59
  I was going to say I might take the opposite side of things,
• 38:03
  actually, just because, you know, it's fun to do that.
• 38:06
  But with LLMS, a lot like my argument on explainabilities,
• 38:13
  like anything fast regression, we're still trying to proxy for
• 38:19
  linear explainability, right? Exlainability oftentimes is just
• 38:22
  like, can you can you express it as a vector, whereas like for
• 38:29
  language models, you're actually pushing it so far past like
• 38:33
  those tasks that you have like highly highly composable tasks.
• 38:38
  And we needed these like extremely deep and huge models
• 38:42
  to be able to do them. A lot of the literature now is
• 38:45
  actually going in, going back and trying to, you know, pull
• 38:49
  the unexplainability too. So like model composability,
• 38:53
  figuring out the sparse networks that are activated during
• 38:56
  certain requests patterns and then like defining them to to
• 39:03
  concepts like those. If anything that's like in my
• 39:07
  opinion like the higher the next level abstraction for what
• 39:10
  explaining ability might look like for us in the future 'cause
• 39:14
  like it's more in my opinion trying to get an LLM to explain
• 39:18
  is like getting a doctor to explain to you how they came up
• 39:22
  with your diagnosis. It's like a very complex pass
• 39:25
  rather than like getting like linear weights or linear vector
• 39:29
  weights. I can ask follow up
• 39:31
  questionnaire. So if your situation, how would
• 39:34
  you try and encourage trust in people who can see them in the
• 39:38
  outputs instead? Do you have standard ways in
• 39:40
  which you can kind of validate that these outputs are good and
• 39:43
  reliable? Yeah, yeah.
• 39:45
  I think that's why evals are so important and which is something
• 39:49
  that the, you know, community, the open source research is
• 39:52
  trying to do more and more. But like, you know, you can.
• 39:56
  One aspect that's interesting about LLMS is they can express
• 39:59
  their confidence in their predictions.
• 40:02
  If you prompt them to do that and they can actually go in and
• 40:05
  they act almost as like a Oracle to go inside of the compressed
• 40:11
  knowledge that they have to express that in in in actually
• 40:15
  be very, they can be very transparent about how they came
• 40:18
  across. You know their predictions too
• 40:21
  as well. So I I just think it's like it's
• 40:25
  going to be an entirely different approach to it, which
• 40:28
  is pretty exciting actually, in my opinion.
• 40:31
  Because. When I when I prompt an LLM and
• 40:34
  I worked on prop engineering, I'm like, it's like teaching an
• 40:36
  8 year old, like you have to be very specific, extreme language
• 40:40
  and you express it, then you actually get like very
• 40:44
  consistent output. So even the way that we trade
• 40:47
  models like I build curriculums of corpuses to be like, you got
• 40:51
  to learn math first, then you got to learn, you know, the next
• 40:54
  step of it and then you can learn, you know, how to analyse
• 40:57
  the balance sheet. I would challenge though like
• 41:00
  it's actually quite similar like at the end of the day like I
• 41:03
  feel like even like with an XG boost model you only get you
• 41:07
  know like little value I would say from like putting shop on
• 41:10
  top of it. I think like it's really like
• 41:12
  looking at the other cases and having like exactly what you
• 41:15
  said Mark like a good evaluation, understanding what
• 41:18
  the model might be taking errors and trying to to figure out if
• 41:21
  there are any trends and patterns around data failure,
• 41:24
  failure points of of the model. And I feel like that's that part
• 41:28
  of the process is similar for like let's call it traditional
• 41:31
  ML and and other labs like you need to we have a good, a good
• 41:34
  evaluation. You need to understand like
• 41:36
  where is your model like having reasonable performance and like
• 41:39
  what's like what's your threshold of acceptable
• 41:42
  performance. Like you need to understand the
• 41:44
  error cases and and kind of go into this iterative cycle of
• 41:47
  improvement around it, ideally bringing the the business
• 41:51
  stakeholders along in that process.
• 41:53
  Well, it almost feels like when you're able to bring the
• 41:57
  business stakeholders along in that process and as Michelle was
• 42:00
  saying, there's there's different different
• 42:03
  understandings of the business and you could probably get to
• 42:08
  debugging things a lot faster if you have someone who's the
• 42:12
  expert and you can take them on that journey.
• 42:15
  I'm wondering if Michelle or Oliver, you felt that is a
• 42:20
  consequence or a benefit that you've seen because of the way
• 42:24
  that you've set things up. I.
• 42:26
  Feel like I'm So what else do we have?
• 42:28
  A lot of very we need to reprogram both UIS built in a
• 42:32
  React where if you look at the dashboard for a specific model,
• 42:37
  it'll tell you what version of the data it used.
• 42:39
  It'll tell you what version of the model was used because we do
• 42:43
  a lot of releases and we're released like it's very easy,
• 42:46
  ICD release often and frequently and it's been an MO off team.
• 42:52
  But we use that quite a lot and all that tracking is there and
• 42:54
  the stakeholders can see it. And a lot of times they've
• 42:57
  pulled us up being like, hey, whatever you've done with the
• 43:00
  model between September and November, it's completely
• 43:02
  screwed the forecast. And you go back and look at it
• 43:05
  and I'm not being funny, but they've actually generated like
• 43:08
  really good investigative work for us because we're like, Oh
• 43:10
  yeah, we changed the muscle massively.
• 43:12
  Like that's where the forecaster completely off.
• 43:15
  We didn't pick up on it because we're not the end users.
• 43:17
  So I think enabling the stakeholder with the tools to
• 43:19
  like kind of not poke at your work but just find the data
• 43:24
  lineage and the model like registry to like specifically
• 43:27
  go, hey it's these key points have created a massive
• 43:30
  difference because then we know where to go and look, look they
• 43:33
  don't know where to go and point to if you don't give them the
• 43:35
  little information. Yeah, I mean, I would second
• 43:38
  that. I think in our context, we're
• 43:40
  data scientists, machine learning engineers.
• 43:42
  We are very good at solving data science, machine learning
• 43:45
  problems and we can monitor that and so on.
• 43:49
  What's really useful though is when you do communicate with
• 43:51
  stakeholders and they may identify certain changes in
• 43:55
  data, changes in the forecast, and they they're able to
• 43:58
  construct and understand the story.
• 43:59
  That's not because they have a better understanding of what's
• 44:01
  happening on the ground. So having this process where
• 44:05
  they're able to consume this forecast and then correlate it
• 44:09
  with other data and so on enables them to do that
• 44:11
  investigative work for us in a way, and then that enables that
• 44:15
  feedback process. So I think having those
• 44:18
  frameworks in place is very, very valuable.
• 44:21
  So we've got a question from Harshil coming through in the
• 44:24
  chat and this is a 2 parter and I'm going to lob it up there,
• 44:28
  see who who wants to grab it. Feel free to jump in whoever or
• 44:33
  I can pick on some people #1 on this and we'll start with this
• 44:37
  one. How do you communicate model
• 44:39
  related insights, limitations and performance metrics to
• 44:43
  non-technical stakeholders? Michelle, you were kind of
• 44:46
  touching on that. But do you really want to know?
• 44:50
  Why? Yeah, but there's a little bit
• 44:52
  more to this. So it's not only to the the
• 44:56
  stakeholders like the business side of things or the subject
• 44:59
  matter experts, but also the regulation bodies or senior
• 45:04
  management. I shouldn't give you an example
• 45:07
  of like how I'm told to explain stuff.
• 45:10
  I very much try to tie it to a story.
• 45:13
  So someone else recently explained its statistical
• 45:16
  distribution to your grandmother.
• 45:19
  My grandmother is a little old. She sees snow, but she was a
• 45:22
  little old fireworks wife, so she knew the farm very, very
• 45:26
  well. And I was like, well,
• 45:27
  grandmother, think of bottles of milk.
• 45:30
  I was trying to explain a Poisson distribution and how
• 45:32
  like at the start the milk is really, really, really fresh for
• 45:35
  the first couple of days. Well, after like seven days it's
• 45:39
  very unlikely. I think it's going to be fresh.
• 45:41
  So I was trying to explain it in those terms and ever since then
• 45:44
  I kind of do the same with the business.
• 45:46
  So I kind of be like, oh, if it cost her, I went to open up and
• 45:50
  and we went off for them a credit card and we were giving
• 45:53
  them metrics and we happened to pay them like we gave them more
• 45:56
  credit because of their gender. Like that's not OK in our
• 45:58
  modelling. So trying to give small snippet
• 46:01
  stories where they can envision something of a person doing
• 46:05
  something tends to resonate a lot more than going into, hey,
• 46:08
  this is the formula. How you calculate an F1 score is
• 46:10
  going to bore them to tears. I don't have any good stories
• 46:14
  about that, about trying to explain statistical
• 46:16
  distributions, unfortunately. But I think one thing which
• 46:19
  definitely does in a way Mae scores, position scores,
• 46:23
  whatever valuable to us and then Ray is intimacy.
• 46:26
  Not even worth trying to communicate those.
• 46:29
  The way we try and do it essentially is really boil it
• 46:32
  down to the KPIs which do you actually care about.
• 46:34
  And I think those two, the two things we care about, they care
• 46:36
  about forecasts and the accuracy of those forecasts in a way
• 46:39
  which is quantifiable within the context of the forecast.
• 46:43
  So for example, we have various ways when we run essentially
• 46:47
  back tests to try and assess how accurate our forecasts were to
• 46:52
  try and build some broad approximation, approximation to
• 46:54
  prediction interval on the forecast and try and ascertain
• 46:58
  how accurate we were in the past these forecasts.
• 47:01
  That's kind of so that way. We're trying to boil it down to
• 47:04
  the KPIs and the metrics of they care about and not worry about
• 47:07
  the the data science metrics and maybe important and informative
• 47:12
  for us, but not necessarily for the stakeholders who yeah, are
• 47:15
  not stupid at all by a long shot, but don't know and don't
• 47:18
  necessarily care about those metrics.
• 47:20
  Yeah. From my standpoint, you know, I
• 47:24
  think we're a traditional machine learning background as
• 47:26
  well. But with the customers that we
• 47:29
  have to explain, you know, model performance or you know, show
• 47:33
  them the value of what this model can do and the risks.
• 47:38
  Evals, again, I my evals these days are just so direct to what
• 47:44
  the value is to the customer that they almost are never
• 47:48
  interested in any sort of statistical measures.
• 47:51
  They're like, hey, how many times does it get this Excel
• 47:54
  formula right? How many times, you know, can it
• 47:57
  automate all of our text extraction in our document
• 48:02
  processing and then one number out if it reaches their metric
• 48:06
  or benchmark or not, That's the only thing that they actually
• 48:09
  care about. And then if they have questions
• 48:11
  on the risks, I can almost do it live where where they want to
• 48:16
  prop the model and see you know what's going on and and then we
• 48:19
  can do a simulation live on like you know how to simulate their
• 48:24
  exact use case for that. Are you focusing on creating
• 48:32
  custom evals for everybody that you're dealing with?
• 48:36
  In my it's really hard to have any sort of evals as a service
• 48:42
  work in a way that can deliver really, really explainable value
• 48:47
  to a customer. And I think that's part of the
• 48:50
  journey. Like at least where we are in
• 48:53
  the entire AI and LLM space, the bangers are in the custom evals,
• 48:59
  like all the other ones are kind of proxies.
• 49:02
  But when you come out with a really nice custom eval for a
• 49:06
  particular customer, it almost is like, you know, the whole
• 49:10
  analogy of a picture is worth 1000 words.
• 49:13
  That's about all you need. If you can just hit it.
• 49:16
  Hit the nail on the head right there.
• 49:18
  Yeah, that that makes 100% sense because you get to show people
• 49:23
  this is what it's doing and this is where it's messing up.
• 49:26
  Or this is how many times it's doing it right and how many
• 49:28
  times it's doing it wrong. How do you feel about that?
• 49:32
  So like, is that good enough for you?
• 49:33
  If it is, let's ship it. If it's not, let's go back to
• 49:35
  the drawing board. Yeah, totally.
• 49:38
  So Part 2 of this two parter that I'm going to bring up with
• 49:44
  Harshil has got. The other piece of it is how do
• 49:48
  you ensure effective collaboration and alignment
• 49:52
  between data science teams and other departments such as legal,
• 49:57
  compliance, risk management, all those fun teams that everyone
• 50:03
  loves working with, I can imagine.
• 50:06
  You don't make the data scientists do the governance
• 50:08
  work. In my view it's like my team, we
• 50:11
  have like 20 days like this, but we have 4 managers that will be
• 50:19
  business facing and will do all the stakeholder meetings
• 50:22
  management. They'll be responsible for
• 50:24
  presenting a little bit of the modelling outputs but they won't
• 50:27
  be expected to. Models are what's under the
• 50:29
  hood, that's what the data signed.
• 50:32
  I find sometimes you need specific stakeholder management
• 50:36
  governance people because that work isn't super, super
• 50:39
  technical but again depends on the organisation eyes.
• 50:42
  I'm blessed at the moment, join any governance and I'm knee deep
• 50:45
  in government's models. I think I can come in from a
• 50:48
  smaller organisation and slightly different property
• 50:52
  designer organisation in the sense that I think what we do do
• 50:55
  is we more heavily matrix. So the idea is we have product
• 50:59
  teams and as product teams. So it makes your data
• 51:01
  scientists, engineers, product donors and we're actually in
• 51:06
  this particular case although we don't have any interesting legal
• 51:09
  or finance but we do have heavy heavy involvement with those
• 51:12
  teams and that way we're able to collaborate and get the
• 51:16
  information we need. In fact just the other week for
• 51:19
  some has some fairly extensive meetings with some key
• 51:21
  stakeholders and finance around some common projects we're
• 51:24
  doing. And that way by having this
• 51:26
  matrixing appropriate actually move multiple disciplined people
• 51:29
  into one team then enables the the the mixing and knowledge and
• 51:33
  and so on. It does come with his problems
• 51:35
  because you need to have relative isolation, but I think
• 51:37
  for small organisation, more product driven then it may be
• 51:41
  worth better. So, so if I'm understanding this
• 51:45
  correctly from your side, Oliver, you're saying you embed
• 51:50
  the different people into the team and then they're working
• 51:53
  closely together. But from Michelle's side, it's a
• 51:57
  little bit more like there's somebody that is the data
• 52:00
  science API almost for the legal compliance and they are the
• 52:06
  they're the wall between them or the cushion per southeast.
• 52:10
  I love that the API is class. So just tell legal and
• 52:16
  compliance, you know, send it over to the human API and see if
• 52:19
  you get a response there. That is how you could go about
• 52:23
  it. And I guess so Michelle, I can
• 52:27
  tell that you're very stoked in the way that you have formed it.
• 52:32
  Oliver, do you feel like you should shoot for going towards
• 52:36
  something like Michelle has or do you?
• 52:38
  Are you also equally as stoked, just in a different way on the
• 52:41
  way that you have organised? It I believe it's almost a
• 52:44
  cultural and and an artefact or the size of the organisation and
• 52:49
  the age of the organisation. Essentially we and MK prework on
• 52:52
  a very broadly and late stage startup so the organisation is
• 52:58
  much smaller. We have many more products on
• 53:00
  the go, so we just simply don't have the resources for a big
• 53:04
  centralised data science function which I believe the
• 53:07
  likes of Floyd and so on would do.
• 53:09
  And so essentially I think it's almost a symptom of that
• 53:12
  difference. I think very much there is it's
• 53:15
  almost a cultural little war. I think some people want to go
• 53:18
  for this very matrix product based thinking about how we
• 53:22
  operate teams and do DevOps versus some more traditional
• 53:25
  thing where essentially you would have like an analytics
• 53:27
  department or data science department which should be
• 53:29
  centralised and essentially would then be whether be
• 53:32
  essentially almost a request for services coming from other
• 53:36
  departments. If we want this, I'd prefer the
• 53:39
  product based matrix to approach if in case you're closer to
• 53:41
  guess, you're closer to the problem and makes you better
• 53:43
  understand the problem. But at the same problem.
• 53:46
  That does mean that you there are that you are relatively
• 53:49
  isolated as data science if you're based scientists or an
• 53:52
  engineer if there's not a large team.
• 53:54
  And also it does mean there's much more for bleeding of roles
• 53:58
  and so on, which means that you may maybe not be so focused
• 54:02
  specifically on data science tasks.
• 54:03
  I think it comes back to the point where essentially where it
• 54:07
  seems like there's two teams involved in the forecasting
• 54:09
  function at Lloyds where essentially one is involved in
• 54:12
  more component, the others involved in the component.
• 54:14
  In my situation, we do all of it all at once, so.
• 54:19
  Yeah, it is an interesting thought to see how you are able
• 54:24
  to get that expertise and you are if if you have it at the
• 54:28
  forefront of your mind, I imagine you're going to be
• 54:31
  looking at things a bit different.
• 54:33
  But then again, where can you add the most value, right.
• 54:38
  And if you're spending your time learning about regulation and
• 54:42
  all that as opposed to just having someone tell you no, you
• 54:45
  can't do that and you don't ask questions, you're just like, OK,
• 54:48
  sweet, I won't do that. Then which one?
• 54:51
  It's a funny way of looking at the organisational piece and and
• 54:56
  deciding which one's for you and what you find more valuable.
• 55:01
  Pros and cons always. It's all those trade-offs.
• 55:04
  So folks, I think our time here is coming to an end.
• 55:12
  As much as I love hanging out with you all, I think that this
• 55:19
  is about it. I want to give a huge thank you
• 55:23
  to all of you that have come along on this journey with us
• 55:27
  and have been on the panel, have spoken and then also everyone
• 55:31
  that has been asking questions in the chat.
• 55:34
  I loved it. I really appreciate everyone's
• 55:37
  voice in this. And of course we gotta say a
• 55:43
  huge thank you to weights and Biases for helping put this on.
• 55:46
  That is incredible. And if anybody needs anything, I
• 55:50
  think Derek's DMS are open, right?
• 55:53
  Fresh out. This paper.
• 57:46
  There we go. We will see you all later.
• 57:49
  Have a great night day wherever you are in the world and thanks
• 57:53
  again everyone Hey, I'm Vishnu. I'm a data scientist at first
• 58:03
  hand and I definitely think that you should subscribe to the ML
• 58:07
  OPS Coffee Sessions podcast. It's the best podcast out there
• 58:10
  to. Stay on top of what?
• 58:12
  ML OPS. Actually is to talk to the true
• 58:14
  thought. Leaders in the space.
• 58:15
  And oh by the way, Demetrius is absolutely hilarious.
• 58:19
  Weird guy. You should definitely subscribe
• 58:21
  to the podcast.