#95 How to Build a Data Science Team from Scratch

Hi, I'm Maggie, VP of Curriculum at datacamp, and I know firsthand how overwhelming it can be to keep up with the rapid changes in data and AI. But staying ahead doesn't have to be hard. That's why our curriculum team has curated the best learning content to help you and your team stay ahead, covering everything from data and AI literacy for busy executives to cutting edge generative AI for developers, and everything in between. Whether you're upskilling yourself or your team, we make learning practical, engaging and flexible. It just works. Power your data and AI transformation journey on datacamp.com.

You're listening to Data Framed, a podcast by Data Camp. In this show, you'll hear all the latest trends and insights in data science. Whether you're just getting started in your data career or you're a data leader looking to scale data driven decisions in your organization. Join us for in depth discussions with data and analytics leaders at the forefront of the data revolution. Let's dive right in.

Hello everyone, this is Adele, Data Science Evangelist and educator at DataCamp. As data science becomes more and more integral to the success of organizations, now more than ever, organizations of all sorts and sizes are building data science functions to make the most of the data that they generate.

However, I think given all the data.

Framed episodes we've covered thus far this year, it is definitely no easy feat to launch a data science function from scratch. So I'm excited to have Elettra DiMaggio on today's podcast. Elettra is the Director of Data Science at Stonex. She has been deeply embedded in the data and digital transformation space and financial services and played a crucial role in launching the data science function at Stonex.

Throughout the episode we talked about the.

Main challenges associated with launching a data science function. How data leaders can PR the roadmap between low hanging fruit and long term vision how to earn trust with stakeholders within the organization as a data leader, use cases she's worked on, advice she has for aspiring practitioners and much more. If you enjoyed this episode, make sure to rate, subscribe and comment, but only if you liked it. Now let's dive right in.

Eletra, great to have you on the show.

Thanks guys for having me.

I'm excited to talk to you about your work leading Data science at Stonex. Best practices for launching a Data Science function from scratch how to manage short term objectives and long term priorities and more. But before, can you give us a bit of a background about yourself?

Yeah, sure. So I started like studying computer science a long time ago so I graduated. I Had my Bachelor of Science and Master of Science in Computer Science and then during my master I majored in AI and databases and I graduated in 2009. So yeah, a long time ago. And at that time I had to say data science wasn't yet a thing. Although I went through all the, the neural network vision, NLP type of projects that you might imagine. So I started to work in consultancy and then after a while I got bored of that. So I, I won a fellowship in Paris and got my mba. It was really interesting to get a business. I was a business educational background as well. It actually was very useful to me to learn a lot about how company works and what's behind the product or the service that a company actually offers. And after that I went back to Italy to work in Gartner Consulting. So again in consultancy was just, it was a little bit of a curse on me at the time. But then I moved finally in BMP Paribas and so as in the client side, as in consultancy they used to say in financial institution, retail mostly. So retail banking. First bmp, Piparipa in digital transformation, then HSBC and then finally I moved into Gain, aka StoneX. Now we've been acquired in 2020 and then rebranded into Stonex. And I actually transitioned from a more like retail banking type of service to trading instead. Trading services. As you might know or not know, stonex owns the two brands in UK and worldwide, Forest.com and CT Index. Yeah, provides trading services to people.

That's really great. And I want to set the stage for today's conversation. You led the data science team at Stonex. You led the data science function as well at Soanex. It's always interesting to talk to someone who played a key role in launching a data team or a practice within an organization because I think there's many growing pain stories that are often missed by practitioner data scientists who join a relatively mature data team within an organization. So what are the key ingredients of launching a data science team or function within an organization?

So when I started in Stonex, at the time it was still gain. It was 2019. I started as principal analyst and the hope for my boss was for me to start the conversation about using data science and machine learning within the company that wasn't really using anything of this type of application. And I started with two analysts under me and now I have nine. So it was quite a journey in terms of, I have to say I was very happy to see that the organization was ready to use the data in a certain way and this is one of the key points. So data needs to be ready to consume. If you don't have that, you can definitely not start data science very quickly because first of all, you need good data. And I was lucky enough that all the other people and the other teams and the enterprise data system teams and all of these people spend a lot of time and effort to set up a good data set, a good backend. From a data perspective, that helped very much. This is definitely key. And I understand that sometimes in huge organization where you have the so called data swamp issue, where a lot of people just dump their data in the cloud and then they say, okay, now do something with that. That is really one of the biggest pain points of a data science practice. So that said, when you start a data science practice, the first thing that you need to understand is what can be your scope of action. And your scope of action is directly linked to the quality of data that you have within that scope. So in my case, I think the good, the secret ingredient of the recipe was to understand, okay, where can I bring value based on what is ready to be used? So not start from, you cannot do top down. Because if you do top down and you say, oh, you know what, we should have a machine learning algorithm that do xyz and then you just, you fold this requirement into the tech spec and you understand when you go to the tech spec that, wow, to do this we would need to basically to work on all these data sets. And if all these data sets are in a huge mess, you will just spend months and months, if not years to fix things. So in this case you need to be smart and understand, okay, where I can drive the most value with what I have. It's like you open your fridge and you have, I don't know, eggs and you have maybe avocado and you have something, okay, what I can do with this, instead of taking the recipe book and say, you know what, it would be great to do a carrot cake. And then you don't have anything to do to do a krk. That's basically the same thing. You just start with what you have and try to. And we say, I know it doesn't sound maybe very fancy from a data perspective, but a lot of things, usually from a business perspective that brings value right away is good data or linked data or integrated data view.

That's really awesome. And what I really would like to ascertain from some of your answers here is the challenges related to launching a new data science function. So you definitely mentioned the technical Challenges of data quality. What are other categories or challenges associated with building a new data science function from scratch?

Definitely the talent recruitment and also I would say understand the tech stack that you want to work on. The way we did that was to incrementally find our use cases that we now would, or we were fairly sure that would provide value to the business, try to deliver a pilot of those and then just get more money from the company, more investment. It wasn't a that's all the money, go ahead. We had to earn all our tiny steps and we were fine with that. Because a big bang approach might not be the best. Because the point on machine learning, I believe is that it's very much experiment driven. You need to understand everything that you can work with. You need to run all your experiments, you need to understand and the more you learn and the more you have an understanding of how many people you need, what type of tech you need, maybe someone knows. But personally, if you just join a company and you don't know anything about or you don't know anything yet about the status of the data, the status of the organization, the status of the business per se. I mean in my case, even though I was in financial institution, I was coming from retail banking. So trading was a new thing for me. So I had to learn a new type of service. So if it's a new type of service, a new industry, maybe not an industry, but a new area of the industry, you need to get an understanding of that as well. So data, organization and business, before you have this understanding, really deep inside your head, it's very hard to say I need these people, I need this tech, I need this capacity. So the way you need to do that in my opinion is that you need to learn and readjust. It's lean startup type of thinking. You just start with a pilot with your MVP and then you work on it and you just evolve and add on top and understand if you're still in the right track, if you're doing something that is useful for the business or not and you constantly readjust and you add on top or you just like fine tune. So this is definitely the way I'm doing it and the way I would suggest someone else to do it. And the challenges in doing this is definitely find the right people and not just in your team. But also it's really key for a data science team to have a very good dev team, an architecture like team that can support you with suggesting the right tools for your need, suggesting the right architecture, suggesting everything that you need for example, to process stream of data. There are so many aspects in delivering a data science product that is really hard for one person to know everything of everything. So you need to make sure you have good people advising you in all the steps that you are not an expert on.

That's really great. And let's harp on that organizational challenge, whether it's building out your own team. Can you walk me through in more detail? How do you earn trust as a new data leader within an organization when working with different stakeholders, such as a dev team, such as a business stakeholders? And that's the first set of questions. But the second set of questions here as well would be how do you build out a team knowing that it's still early out in its juncture and that you want to be relatively disciplined in the type of resources and the amount of resources you add to a new team while maintaining the fact that you're adding value. But you also want to make sure that you have the best hire. So what is the type of profile you look for in an early data team?

As I said, those success stories that you can drive in your first six to 12 months, those are keys for you to build your trust. If you can deliver a success stories, say within your first year in the business and it could be something that people can say, oh you know what, who's that person delivered that? And they can associate you to a certain type of deliverable. So start to build this type of trust by actually have direct content and being, how to say, lead by competence. So make sure that everyone thinks associates your name to something that works. And that is definitely step one. And then you start from that. And I would say if you can secure that, it will be all good, everything will fall a lot smoother instead compared to something that just barge in and say, oh, we should do this, we should do that, and so forth and so on and on. The second point instead, about what type of people hiring and your early data team. So because the tech stack was very simple at the beginning, like very, very simple. Because we were building the practice, let's say. And that is also related to the iterative approach. If you start with a very complex tech stack, you know, very complex, a full tech stack from your cloud, your machine learning and ops platform, your data engineering, ETL and all the works, okay, all the works are you have GCP or AWS or Azure cloud and you have on top your email. Of course you need people that are skilled on all this tech to deliver something. So automatically you will need more People because you will need. You cannot have someone that knows everything about all this tech if you start with an easier tech stack. We started with Python having a server that was running our Python script to test and and then we, let's say partnered with other dev team to deliver some models in production. So we didn't do the delivery in production but we handed that over to other dev team that had other tech stack. So with that in mind, the type of people I hired in the first place where I would say data scientists that had a little bit of coding, if not coding experience, just coding appetite, so they didn't mind setting up Python scripts that were just getting data from API scraping, website or whatever to get the data that they need to have to develop their machine learning models or to just test and experiment the machine learning models that we had in mind. And once we developed this couple to, I would say a couple of success stories, we finally started to have our own development platform. We have been completely included in the DevOps process because when I started the analytics team wasn't considered part of the DevOps. It was an old school Excel BI type of team and that was all it was reporting. Most times it was just reporting. But of course there was an appetite to evolve that. So we started with that in 2020. They said, you know what guys, you are developing software, it's good that you are included in our DevOps. So we started to be included in the DevOps. So we had some training. I already knew a little bit of Git and BitBucket or GitHub or whatever. We switched repositories in between. But the other guys were so type of people eager to learn. Definitely they need to have solid foundation from a statistical and mathematical perspective, but they need to have that, I would say that, that appetite to develop things, to not just analyze things, but to really develop something that is a product. So it's more, it's more an attitude type of thing that you need to associate to a strong quantitative background. That was the type of people that I was hiring at the beginning.

That's really great. And how has your hiring practices or what you look for evolved as the team grew and it became more established and provided roi.

So now that the team is a little bit more established, the way I set up my team is that I have guys that are more focused on the data engineering and machine learning, actually engineering things as we are setting up finally our MLOps tech stack. So I don't know if this is like a very mean differentiation, but the way I see this is that you have. There are like people that are driven to write what someone might call production code. They're like other people. They're more driven to analyze experiments and see things like what. How I see the data scientist at the moment in my team is very much an R and D function. So it's a person that needs to have a business acumen, so needs to know about the business or needs to be able to understand the business. So has a strong commercial, organizational and business understanding and of course has that statistical and machine learning knowledge so that can, you know, just join the dots and say, oh, you know what, I can use this data to solve this problem. But once I would say the data scientists molded the infinite space of solution and caged it in a little bit more manageable space, that thing is passed on the machine learning engineering and the engineering function that will industrialize and set up the pipelines and everything that needs to be done in order to operationalize and make of that mold a product that is reliable, sustainable, sustainable and reusable within the business. On top of these two group of people that I have in my team, I also have a BA that supports me and the way a BA which I think is really useful because BA is that type of person that first of all has a constant relationship with different stakeholders that are customers of your products and can gather requirements and have a conversation with the data scientist or the machine learning engineer to say, you know what, maybe we should do something to either change the product in this way, an existing product, or maybe design something new that would include, that would solve this type of issue and also is the person that really helps you embedding the product within the business. You know, training business stakeholders, talk with them, maybe guide them at the beginning on how to use and how to interpret data and how to interpret the model workings. Because one of the things that when you develop a machine learning model is that it's very hard to explain these to known data people. So you need to have that person that has that constant relationship with them so he can or she can wrap that up in a way that is understandable and so that you can have sponsors outside your team. That's key. You always need to have sponsors outside your team.

I love that answer and I love how you create the delineation within the data science team from a more research and development type, small mini data team that transitions its outputs to a more applied engineering team that industrializes the work. A lot of data scientists do. But harping on that last note here, when it comes to the business analyst role and when creating a relationship with other stakeholders and gathering requirements and feedback. Oftentimes when talking to data leaders, a big obstacle they face when it comes to providing value with data science. Analytics is data culture or analytics mindset or lack thereof within the organization. I'd love to understand from you how did you approach conversations with the remainder of the stakeholders within the organization that may or may not have analytics mindset or a data culture, or understand the value of data science and how you were able to maneuver these obstacles, whether through the use of a BA or within your own team and how you approach these conversations.

So first of all, this has nothing to do with your data skills. I'm just putting this as a disclaimer on top. This is all about your, I would say, political skills or relationship skills. So as I said, it's very key for you to see start understanding where you can find your sponsors. So first of all, you need to have conversation. For example, in our case, our company is organized in commercial leaders and we have global teams as well. And commercial leaders. Of course, you have commercial leaders of the biggest regions, commercial leaders of maybe smaller regions. And you need to gather an understanding on who has the most driving role within the community of executives. And I'm sure if there is a data science team in the company, you will be able to find your sponsors from day one, the ones that are really keen to get involved into that. It might be easier or harder in some cases. So first thing, try to understand what are your easier sponsors, the ones that maybe they're the keenest in sponsoring you, but there might be still on the lookout because you haven't delivered anything yet. I'm interested in data, I would like data science. So try to understand what are their key requirements. And as I said, I remember when I started, I was like, this is a little bit of your Jedi trick. So you don't want that, you want this. So when you have a conversation with them and you know what you can deliver, you need to in a clever way sell something that is useful for them, but you can deliver in in a reasonable amount of time. So you try to drive them to that type of solution. And this is your personal negotiating skills, let us say like that once you have secured your good sponsor, your big sponsor with that, you just word them one by one, if that makes sense. And this, I know this might seem okay, but what happens when I deliver the model, I have to explain that to them. This is not related, right? It's actually very related. Because if you Know that they, in their heart, they're already sponsoring you. The day you are going to them explaining, they will have a different attitude listening to you. So you will have your chance to explain it to them. And I would say don't never be condescending, never been the lecturer there. Always try to. You know what, I delivered this because the main goal for this is to provide this additional benefit for you. I'm using this. Do you want me to go through the details of the model? I can. Most of the times I have to say that we're interested in knowing the performances. So whatever type of performance metrics you want to use, try to save for the business stakeholders, the ones that are mostly understandable, all the performance KPIs that you use to understand if the model is sustainable, if the model is robust, maybe just save it for the annex. But at the end of the day, commercial stakeholders want to know how often this works and if it doesn't work, what is the risk? So for example, we had a churn prediction model that we started to share. It was our first XGBOOST random forest, actual real machine learning type of model. And we tried just to. We went through that just explaining the features and we explained the confusion metrics to the commercial leader. And that was already too much because it's a new thing for them. And the way we were talking to them about that was the model on average, 90% of the time predicts correct. But what the mistakes we worked in a way that we are over predicting churners because at the end of the day we slightly over predict churners. This is why we don't have higher performances. Because at the end of the day it doesn't cost cost to us a lot sending another email or calling another person that is at risk of churn. It might cost more losing someone that we're not calling and wrapping in this way. It was very understandable for them and they were really happy with that. It required multiple explanation, like multiple times to go through. But after that you just build trust and it's easier and easier because they just start trusting you and they say, okay, you know, I don't have a full understanding, but if you say it's working, it's fine, we'll see. We'll review it after a couple of months that we have this running. So this is the type of, I would say massage that you have to do at the beginning. And you need to be patient and not too rush or aggressive. Definitely not aggressive.

That's really awesome. And I think at the crux of a lot of the different answers that you've done so far. And I think a key central tenant.

When it comes to succeeding and launching.

A data team is managing both the short term priorities and the short term wins that you can get, but as well as making sure that you're working towards a long term vision. So there's always a North Star where we want to be in the long term and quarterly OKRs and objectives that guide the short term objectives for a data team. Can you walk me through the process of prioritization between these two objectives?

I have to say it's not something that you do alone, especially if you're joining a new business. The first thing that you want to do is also have a talk to the people that have been long time in the business so they can share with you what I've been 20 years in the business or 15 years in the business and I think that one of the things that really would disrupt us will be a way to predict that, to understand that and then it's okay. Wow. And never underestimate the fact that if the guy has been there for 15 years and they didn't manage to do that, it doesn't mean that because you're a data scientist in one year you're going to do that just because you have machine learning or whatever, it's probably harder than that. So you just put that and you gather all of this thoughts and you understand okay, you know what? So let's define a roadmap to go there. So for example, one of the things that we gathered from our I would say key internal system speakers or applications that we can apply to online stream of trades and transaction. And of course being able to apply machine learning model on an online stream of data, it's something that requires a tech stack that we're building towards that. But if we started to do that from day one, we wouldn't have delivered anything valuable or will just be cost for the business and will probably still be working on it after three years because it requires time to do that. So you have that and so that's your top down checklist if you wish. And this allows you to understand what is the roadmap. So what we have now, okay, now I have my desktop, a SQL data warehouse and Excel because that's how we started. And I need to go where machine learning, online streaming, set up this, what do I need to do that? And you can do it yourself. I would always advise to talk with other people as well on the architecture side. And gather like their view because I'm sure other people would have thought about it as well. And you start defining your roadmap and milestones. We would need to have at least an orchestrator like Airflow to run our scripts in Python and all of these things. We would need to have a DevOps process and that is step one. And then you go okay, you know what? We will need to have probably a cloud based approach to run our machine learning not on our desktop, on a cloud as compute that is scalable. And we don't need to leave off our laptop to run overnight to train models. We will have something on the cloud to do that and have some platform to connect to different data sources like for example, I don't know databricks or this type of Azure cloud and all of this platform. And then to actually get stream of data you will need something like Kafka. And then you start using Pyspark and all of this things. So you have this plan and this is your vision planning that is always easy from a certain perspective. You just, you plan and you say okay, what do I need? I need all of these things as your grocery list. On the other side you have short term and short term, as I said before, you need to start with what you have. So what do I have? This, what can I do with this? And what is priority for the business? Priority you get from your sponsors or commercial stakeholders. So you get the priorities from that from the business. When I joined I got two priorities. We need to understand how much we're spending on acquisition marketing and how much we're getting from that spend because at the moment we have no idea. So that was one priority. And on the other side we don't know how we're targeting our customers. We need a way to segment our customer and define the journeys based on our segments. So very much acquisition focused. And I have to say having an MBA or whatever business course or marketing course that you can have really helped me there because I know how a marketer would think about these things. Defining Personas, defining the user journeys, defining all of these things. This is a knowledge that I got from both my MBA and also my previous job in retail banking. Because I used to work in the UX team as a quantitative ba, I was analyzing data and defining journey with the user experience designers. So I knew how much that acquisition artifacts were important for designers and for marketing in general. Thanks to that I was able to capture that. But I have to say they were very vocal in saying that they had these issues. I Said, okay, well can we do that? And I had a look at our data warehouse and as I said at the beginning, it was really key. I was very lucky to have a neat data warehouse. Even if it was just our backend on premises data warehouse which just our onboarding system and customer activity. I was very lucky to have a very neat data set to start work with. Of course there were some glitches in the process, but nothing too messy. So that was key of the first successes that we had. So that's how I started to prioritize like more short term goals.

That's really great. And if you want to abstract this out and propose a framework that can enable other data leaders to extract small wins as well as low hanging fruit that demonstrate early value for a data team, how would you go about that?

So the way I would go about that, I would start with the data. Don't do that alone. Start with your business stakeholders and ask them what are the data sets that you use in your day by day job and how do you use them? Because if how they use them you can understand, oh you know what, you could automate that. I could do something that will help you in using that data in a more efficient way. And by going this way you're able to first of all understand right away what are the data sources involved in the process and have a look if the data sources are usable. And second, you have your use case and even if it's not the fanciest use case, you can start delivering something very quickly because you have a workable data source to start with. And by doing that you start building your sponsors. And once you start building your sponsors, even if it's like with tiny deliverables, you can start building up things on the other side. I would say based on how messy is the data situation in the company, you can start involve other teams and raise awareness. If it's not there already, maybe it's already there awareness, but raised awareness and investment of the time and resources on fixing the data so that the data will enable you producing something that is more of higher value, this is the way I would do that, as I said, is a very entrepreneurial, lean startup type of approach. MVP first and then you just build up your way to the top.

I couldn't agree more. That bias to action and having that lean approach is super useful for a lot of data to teams. Now as we end out our episode, Eletra, I'd be remiss not to talk a bit about your work at Stonex, especially on the data science use cases that provide value in financial services. With the recent war in Ukraine, Covid induced supply chain issues, economic uncertainty. I think it's never been more important from a data scientist perspective to understand the role data science plays in commodities trading, foreign exchange trading and more. So I'd love to understand some of the ways data science has been providing value in the industry.

I have to say we haven't been requested and in our case the overall international situation didn't affect too much directly. Of course we know there are like some people that have been sanctioned so accounts have been blocked. Stonex didn't have a huge impact on this point, so we've been lucky. But as a trading company we of course experience a lot of volatility in the market and that made our business very active from a certain perspective in terms of how data team in our case we haven't been involved too much apart from making sure that what we were seeing in our system wasn't affecting other processes in the business. But in terms of doing anything, we haven't done anything also because when you have this I would say delicate situation, it's left to human handling this because you never know if you automate things, you are prone to, I would say embarrassing mistakes. And this is something that of course no company wants because likely, I mean because we have a manageable volume of customers, a manageable volume of accounts, the data team wasn't really involved in doing anything specific.

What are some of the main use cases you've been working on as a data leader at Stonex?

So we have definitely a lot of things related to marketing. So segmentation, attribution modeling, churn prediction, lifetime value prediction. Last year we did our first NLP application to classify customer communications. At the moment we're also working on client sentiment in trading and definitely one of the things that we would like to work on, as I said before, is online streaming of data, but I don't have yet workable use cases to share. We need to build the ground to do that.

That's awesome. Eleara. As we close up our episode, I'd love to look at any future trends and innovations that you're particularly excited about in the future.

At the moment I feel that we are achieving a sort of Data science has been a very wild and chaotic type of aria. There was a lot of buzz, not many companies achieved to tangled the data science practice. So at the moment the focus that I have is to try and industrialize the approach and make the data science practice like solid. So the type of for example, tech that we're looking around is definitely MLOps and pipeline tech in terms of like pure innovation and machine learning. Honestly, there's nothing purely innovative that we're looking for. We have so much ground to recover and to work on it before we do something like more innovative. But especially for marketing, there's a lot of innovation in terms of combining multiple models. So ensembling, for example, but also combining multiple models to dynamically select advertisements. This is something that is in our mind and we will definitely do that. So using internal and external data to understand what are the trends, what are the things that are actually grasping people's mind at the moment, and dynamically select the content of your advertisement, serving them at the right time to the right person, that is definitely something that is becoming machine learning heavy, especially with all the cookie policies that are becoming more and more strict. So this is definitely something that is in my mind. I don't know when I will be able to implement that, but this is definitely one of the things in my mind.

That's awesome. Finally, Letra, as we close up, do you have any call to action before we wrap up today?

I would say just it takes patience and hard work. So if you're not ready to have patience and do you know your your hours to to get your success stories, do something else. But it gives you a lot of satisfaction. But it definitely gives you a lot of satisfaction. In the end is worth is worth your while. But it's a hard way to the top if you want to rock and roll as they say.

Yeah, 100%. Thank you so much Eletra for coming on the podcast.

Thank you for having me. Thank you.

You've been listening to Data Framed, a podcast by Data Camp. Keep connected with us by subscribing to the show in your favorite podcast player. Please give us a rating, leave a comment and share episodes you love that helps us keep delivering insights into all things data. Thanks for listening. Until next time.