Data Science Hangout | Jacqueline Nolis, Saturn Cloud | Structuring Teams to Empower the Business

Transcript#

This transcript was generated automatically and may contain errors.

Welcome, everyone to the Data Science Hangout. I'm your host, Rachel Dempsey. But for anyone who's just joining for the first time, this is an open space for current and aspiring data science leaders to connect and chat about some of the more human centric questions around data science leadership. But really, it's just an opportunity to focus on questions that are most important to you all as well. So you can jump in live or questions in the chat. And we also have a Slido link so you can ask questions anonymously, too. But just a quick note, this will be a recorded session. But I'm so excited to be joined today by my co host Jacqueline Nolis, head of data science at Saturn Cloud. And Jacqueline's led data science, machine learning and AI projects also at Microsoft, Adobe, Airbnb, and T-Mobile. And she's also co wrote the book on building a career in data science. So happy to have you here today. I would love to have you introduce yourself and maybe just share a little bit about some of the work that you're doing today.

Sure. So hi, everyone. I'm Jacqueline Nolis. I got an undergrad and master's in mathematics. Then I went, I want to go get a job using data to help businesses. And this was before data science was a term. So I did that for a while. Then when I said, Nah, I want to go to academia, I got a PhD said, No, I want to go back in the industry. So then I consulted for many years, including at a bunch of the companies Rachel mentioned. And then I recently started, maybe not that recently anymore. I'm now the head of data science at Saturn Cloud. And Saturn Cloud is a cloud platform for data scientists to run Jupyter notebooks and things like that and use Dask clusters for distributed computing. And I will say I've used R for about 10 years now. And it's been a really fun journey. And R is my favorite language to use these days, even if the company I currently am employed by is more of a Python place.

And Jacqueline, I think it's also helpful for people maybe to understand some of the projects that you work on at Saturn Cloud as well and like what your team looks like too.

Sure. Well, okay. So throughout my career, I have started to lead data science teams. Like when I was a consultant, I created a data science department on my company that had like eight data scientists and managers managing. Then like when I went to T-Mobile and helped consult there, I like helped them start their AI team. So I've done a lot of helping teams start, and then growing up teams. At Saturn Cloud, I'm actually because we are a product for data scientists, I'm less of a like do analytics and build models and more of like product development of like what do data scientists like want? And, you know, how can we build documentation and examples and like help our customers solve their problems, that kind of work. And so that's what the team I lead does now. But it's actually the thinking of it is fairly similar. It's not as different as you'd expect from purely I'm gonna arima a neural network kind of data science.

Awesome. So I get to ask the questions in the beginning when people are thinking about what they want. So I'm curious, what's something that you're really excited about in data science right now?

Um, that is a great question. I think that I have found it really fun to watch. And like the last three years, I would say, maybe a little more than but it's gotten really easy to use the most cutting edge models. Like it used to be that like, well, you could use a linear regression or a logistic regression, or maybe if you really put some work into it a random forest, but if you want to do something advanced, you need a blah, blah, blah. And now it's like, man, with PyTorch, TensorFlow , XGBoost, like GBL, like they're really the difference between the most powerful models and the powerful models that are powerful enough that you can use them on a day to day basis gotten really small. I think that's really cool, because we have to, there's less like, oh, you only use a blah, blah, blah, you haven't even tried to use a blah, blah, like there's less of that, because we can all just kind of use everything now. So maybe the thing of data science is it feels like there's less gatekeeping than there used to be. And I think that's great.

Centralized vs. decentralized data science teams

That's awesome. And I know I got to talk to you a little or I feel like I got to talk to you at the RStudio webinar where we had the building effective data science teams. Something that I thought was really interesting from that conversation was this whole idea of like distributed versus centralized data science teams. And I know you had a kind of set idea on that. And I think it'd be interesting for everyone to hear a bit more about that, too.

Yeah, so I will actually tell a story around this. So this is back when I was the head of a data science team at a consulting firm. And some like one of these Gartner reports, like Forrester kind of a company was asking us like, what is your stance on distributed versus centralized data science? Meaning? Do you think at a company, the data scientists should all live on individual teams, like you should have the data science marketing and the data science blah, blah, blah? Or do you should you have one central data science team that all the data science report to you? And there's more like a consulting firm that kind of goes out and helps. And at the time, when I was working on this, and we're a consulting company, we didn't want to harm and offend anyone, we did a very consulting company thing, we just said, you know, we think there's, there's reasons why you'd want to use all of them, you know, and we just kind of like absolutely punted on like, who's to say what is the right thing. And now in 2021, I feel actually pretty definitively, like in most situations, the decentralized method is better, which is to say, have your data science teams really close to the people who are there helping is better than having data scientists close to each other. Because the data scientists get close to each other, what do they do, they really like talking to each other about really complex models and building really complex things. When you have the data scientists close to the teams they're working on, then you actually can like solve the more practical problems.

And now in 2021, I feel actually pretty definitively, like in most situations, the decentralized method is better, which is to say, have your data science teams really close to the people who are there helping is better than having data scientists close to each other.

I think I have almost, well, like I've either been at like a small company where like, we just have one data science team because there's only 50 people. Or usually, well, actually, I've kind of been on both. I've been at companies where there's just one central data science team. And I've been on companies where each team has its own data science team. I do think there are times where all of them backfired. I worked at a certain, I consulted for a certain large tech company. And they had within one product, they had like 10 data science teams all basically doing the same thing all fighting with each other. And like, that's not a good use of decentralizing. But yeah, I think having your data scientists close to the people they're trying to help is still like the generally right call.

So specialization right to an extent is a positive thing. How do you feel about having the person that is awesome at forecasting sitting in just one of those business unit teams, versus the person that right has spent a lot of time doing cluster analysis and then figuring that out? When you have these individuals who are really, really good at things, and then they get stuck kind of in a corner, which is great for that team that has them, but not so great for everybody else in the company. How do you see that balance playing out?

I think that what you just said is the justification for having some sort of internal consultancy. Like if you really have a rockstar forecaster, put them on the team of people that goes around and just talks to the other people to like get their forecasting up. Like I think if you have 100 data scientists at a company, you could have 10 or something who are that style of like, I'm just going to go around and make all of our forecasting better. And then the other 90 are actually like, more in the field of doing it. I'd said, I think there's actually a lot of risk to in having people specialize. I've been in a lot of situations where like, people think have a desire to specialize too quickly. And like, oh, we're gonna have this person, like, there's so much risk and having this one person be the forecast person, and they're going to do forecasts and no one else knows how to do forecasts, but that one person, like, then you, that person gets bored of forecasting, if they quit, your company's really bad position. You're much better off, I feel like, avoiding those sorts of scenarios and making so everyone knows a little forecasting. I think I've seen occasionally problems where people don't specialize enough, but much more often problems where people specialize too much.

How do you also have the team like, I'm imagining when it's decentralized, and you said it's not as many data scientists, it's all talking together, how do they also help train each other on different topics, or, like, keep up to date on certain things if you're working on a separate department?

I don't know, like, I think you can get clever with like Slack channels and like monthly meetings where we all get together and just socialize, right? Like, that's more about like, creating human connections than it is about like having an internalized process. But like, if everyone's friends, that doesn't super matter. Because like, if somebody goes roggy, you just go ask your friend. So that's more about like creating human connections than like formal processes, I would say. But I don't know, I think that like, also, like at some level, like, I don't know, I usually learn more about data science from blog posts than I do from co workers. So like, it's less about having all your co workers talk necessarily some internal information maybe but like a lot of it's also just like making sure your data science teams have the time to go out and learn new things and stay informed.

And then the other thing I think is that, yeah, you really can't document your way out of talking to stakeholders.

What kind of system do you use for requests?

So I do think that you have a person on your team who's the project manager person, and maybe that's actually a project manager. Maybe that's the principal data scientist, whatever, but someone whose job it is to triage requests. You then have, when anyone has a request they want to make of you, they either talk to that person or drop it in the JIRA board or whatever board. And then that person looks at it, tries to understand it, goes back to them, has a meeting about it to really clean it up. And then from there, the data science team tackles it. I think some sort of intake method like that works best. I think creating a standard form that they fill out, that kind of stuff doesn't really work because every question will never fit into a form because always have its weird edge cases. And I also think, yeah, I think just having a person who everyone knows the point of contact on this just makes it a lot easier. And like, I don't know, I find that job really fun. I love talking to people about what their problems are. So it's not necessarily like a bad job, but it is something where you should have one person who's really the point of contact on it.

Being the first data scientist

You mentioned going in, starting all of these teams from scratch. Was there a common denominator, this common thing that you would experience in companies where there's maybe not a data centric culture or there's not a big understanding of what data science is? Was there a common thing that, you know, hurdle or something you had to overcome to kind of get buy-in or, you know, just to even start the team off?

Yeah. So a couple of things. I mean, usually if they are hiring you, there's got to be like some big buy-in at the company in the sense of like, you wouldn't have been hired if at least someone somewhere didn't think data science would be useful. So I used to worry that like, am I going to get buy-in for this? That's kind of gone away. Usually it's like, there is buy-in, but there's a total lack of idea of how to actually make it work. Which is to say, and I've had this happen a few times where you come to a company, you start a team, but the question is like, is there actually data we can use? Is this data actually useful? Right? Like, so if they hire you with the expectation of like, you're going to build machine learning models on the revenue data to blah, blah, blah. And then it turns out like two months in that, like, there's not enough data here that you'd ever be able to build models on it or stuff like that. That stuff keeps me up at night, but I don't think there's any real way to solve it besides before getting hired, really try and like probe on if this seems like a good chance for data science and really just being flexible with once you do start.

I will also say to the previous or earlier point, when you are the first data scientist you do, you are the first data scientist, everything you do sets a precedent. If you pick R or you pick Python, now everyone's going to use R or Python. If you use the Git repo or blah, blah, blah. Like, so it really, more than any other time, you being careful about what you're doing is going to pay off in the future. And also you have to, at the same time, and here's the twist, you can't just be like, I'm going to build a beautiful cathedral of Git repos, this CI, CD, blah, blah, blah, because you'll spend the next six months of your job setting up all this tooling, no one else to use it yet. And after six months of not actually doing work and just building tooling, you will be let go. So it's like, it is like a strategic game of like, where do you actually invest, given you can only do so much investing while concurrently trying to produce revenue. I think being the first data scientist is extremely fun, but that is quite a stressful thing.

It sounds like maybe that has happened before where you get there and you don't have the data that you actually need to do what they want you to do. Like how did you actually handle that?

Well, you have to have some real frank conversations with your leadership about, Hey, you hired me to do things X, Y, and Z, but think X, Y, and Z is impossible. Let's talk about what is possible and how I can do it. And if you have a strong set of leaders, they'll be like, you are a data science expert. I trust you on that. Let's have that conversation and pivot. And if they are weak leaders, they will say, well, you just don't know enough about data science. We're going to keep doing the same thing because we know more than you. And then that is like a, I don't know, not great situation.

Denying requests and organizational hierarchy

Sure, absolutely. So I work at a hospital where we have a fairly new data science team. And the problem with hospitals is that they tend to be very hierarchical in that everybody's first idea is always would go as high up in the hierarchy as possible to ask what we should solve. Whereas I maybe have the understanding that the good ideas actually come from the people who work with the patients and do all of the stuff where whose work we're supposed to be helping. And I haven't figured out a good way to triage where people I work with are not necessarily comfortable saying to the high ups that this might not be the best use of our resources. And I also am not senior enough to maybe be able to do that myself. So if you have any ideas on how to, as you say, like, try to bring the reality into the conversation, that would be really cool.

So I will first say you said, well, I'm not necessarily senior enough. I think that is a student observation, which is I think a lot of times people get on these calls and do the way like you can be the lowest person in a company. And if you pitch your idea, just right, people are gonna love it. And like, that's not really like, like, positions are real. And sometimes being more senior gets stuff. So like, I think it is good that you are aware that that stuff exists, and you're navigating around that. That said, the most luck I have had in these things, in these sorts of situations is just trying to like pivot the pitch from not like, hey, more senior person, you're wrong. And here's why. Although that's maybe how I feel but much more like, hey, how can I show you, senior person, the things I am seeing that are making me think that the more junior way of doing things are, right? So like the senior person, like we need to build a whole new data platform or whatever. And you're like, no, we really just need to like, go survey some hospital beds, right? Like the more you can be like, well, let me explain why. Let me show you some data. Like the more you can kind of, it's not like, hey, I need to change your mind. It's like, I'm going to bring you into this new, even better idea.

Like that's kind of thinking helps, which is to say, most of what you're kind of asking about, and I think has repeatedly come up on this call is like, organizational dynamics and like the things that have helped me the most of my career get good at these are books on like negotiation and like getting to yes and difficult conversations, just like thinking, answering the question of how do I convince another person to do the thing that I think we should do? Like that's that is totally unrelated to data science and very much like business philosophy, like argument, meta conversation, you know, like, and I think just really thinking about it that way and trying to look at those resources help.

Hiring and upskilling

As a data science leader, what are the pros and cons of hiring new talent or trying to upskill existing talent?

So I will say, I don't super feel like it's, um, it's like a binary. I always feel like I, as a leader, I always want to be upskilling my employees, which is to say, at every turn, I should be trying every day, regardless of hiring other people or anything like that, I should always be trying to make sure my employees are working on things that are challenging them and helping them grow. And they learn new skills, both because that gets me more things that they can work on at the company. And it's good for them because they get new skills and have presumably more fun doing it. So like, I just, I think that the upskilling is just like a perpetual push forward of people. When it comes to like the pros and cons of hiring new talent of like, when is it valuable to your team, I really struggle with this, because I do think that when you hire more people than that, that actually is a really big burden on your team, both like, you have to give them work to do and make sure they are content and communicate with them. So I, I think I tend to be a little bit more conservative in my hiring, just because I find that terrifying. I'm more terrified that and then I am terrified of not having enough people, which is, I think, just my personal, like personality showing up. But like, there's just, it's a very real calculation of like, how much do you want to take on the risk of like, have not having enough work and your team getting too big and blah, blah, blah, versus the risk of missing opportunities, because they didn't have people around.

Data analyst vs. data scientist

Oh, okay. Great question. I would say a data analyst job is to take raw data and get it into a consumable format for an executive, right? So if you have a SQL table, it's an analyst job to turn it into an Excel file that has just the summary monthly numbers and put that in a dashboard and these sorts of things. This is a challenging job because like, well, what are you doing with sales data? What about returned objects? And what about this? And like, well, if we do that, then that actually changes last month's numbers. And that would be weird. Can we do that? Like, there's like a lot of work and thought that has to be into getting executives that data.

A data scientist, in my opinion, is a person who is more focused on, they get a data set in an abstract question of like, why are my customers churning? What is the best offer to give my customers? Things like that. And a data science job is to like build models, do regressions, forecast, whatever, to try and answer those things. So a data scientist is more about using models to answer questions and like do inference, whereas a data analyst doesn't really build models. They just take data and group, aggregate, summarize, and visualize in thoughtful ways that people within the company can use it. And I think companies have like a thousand times more data analysts work than data scientists work. And I think there are a lot of data scientists there who like, I would never do data analysts work. And I know because I used to be one and I was very snobby about this early in my career. And now I do a lot of it and it's fine. But like, there's like a mental hurdle you have to get over of like, things aren't less interesting or numerically challenging just because they don't have a regression in them.

Roles on a data science team

So as far as data science roles, I guess teams, you mentioned kind of a project manager earlier, data engineer, data scientist. What do you recommend as far as the types of roles in a data science team?

It's super depends on what your team is doing. If your team is like primarily like making reports, doing analysis, like doing an analytical, like trying to answer these questions with data. That is different than you're building actual models that are going into production, that sort of thing. I think, I mean, obviously you need data scientists. I think after you have like two or three data scientists, you probably want to have like a principal person or like this, a person who is like job it is to like coach and help out on the other data scientists. I also do think having someone whose job it is to do project management, which is could very well be that principal. That's valuable too. I think having a manager is good if you get big enough. Okay, so let me, I'm going to start over. Here's what you need. You need data scientists to do the work. You need someone whose job it is to help those people technically when they're not sure about stuff, which could be convincing them to do unit tests or whatever, and checking for that. You need someone who's a manager whose job it is to figure out that the work coming in makes sense, right? Like a manager's job is to like, hey, are we asking, are we working on the right stuff? Are we asking right tests? Are our employees growing and are they getting the right comp or whatever? And you need someone who's a project manager, which is like literally like, hey, person X from this department asks us to do this. Does that request make sense? Do we need to flesh it out? How does that fit into our like list of work? That's like five roles on that team. A lot of those can be bunched into the same person, but depending on what your company is like, it may make sense to do that bunching differently.

Transitioning from analyst to data scientist

If you're a data analyst today, what tips do you have on making that transition over to data science?

So I think it's a lot about learning, learning the regressions, learning the models, getting more comfortable because analysts are more likely to use SQL and Excel and those sorts of tools and R and Python. And the more you can get comfortable with R and Python, the better like there are boot camps, right? Like a data science bootcamp is a thing designed to get you those skills you need real quickly. But a lot of times when people ask this question, it's because they are an analyst and they aren't just in this stuff and they learn some R in their free time and now they're trying to figure out, but okay, but how do I actually get the better job? And then at that point, it's a lot of like crafting your resume and stuff to like actually like highlight that, hey, analyst work is very similar to data science work. So look, it's not really much of a stretch at all for you to switch positions. So at some level, it's about like marketing yourself as it is as much about getting the right skills and things like that, or using internal opportunities. If you can switch teams and that sort of thing to see if someone could switch you within your org to get more work. I will say, I have a co-wrote a whole book that has discussions about these sorts of things, and you can get it at bestbook.cool with the offer code buildbook40% for 40% off.

Explaining departures and data science consulting

But when you do determine that it's time to leave a company, how do you explain that departure when looking for the next job?

Okay, so we do have a whole chapter on leaving your company. What I do is, I mean, I guess it like super depends if it like, I think it's very one. Oftentimes, people don't even ask because they don't see it like it's often not too real. But if they do ask, you can be like, Hey, you know, I've been at the company for a few years, and I was looking for a new opportunity. And that's fine. You can I've been in situations where like, I actually left because of the stuff we talked about earlier, like I have incredibly toxic boss or whatever. And like what I do there is when I go to my interviews, I say, like, well, why did you leave? I'm like, well, you know, after being there for six months, they're just very clear that there wasn't necessarily a fit in terms of what they wanted data scientists to do. And that what I thought was the right stuff to do. It just like wasn't a quite right fit. And that's like a wink, wink, nudge, nudge for like, it was a bad talks environment, but like no one questions that. So like, there's a lot of ways you can say I don't like working there anymore. That's still like middle class sound like look, we're being professional and fun, aren't we? And I think that you just kind of have to like, practice saying the words a few times, then go into the meeting, and it's fine.

Pitching data science to executives

One is if you're talking to an executive leader, how would you explain the core value that a data science team brings to the organization?

So how would I purse if like, I get asked into the office by a CEO of big name company, like fortune 40 company, whatever. And they're like, Jacqueline, what is data science? Why is your company's data science? I would say, look, companies, you generally collect lots of data, what products people are buying, how they are using them, blah, blah, blah. You also have to make a ton of decisions, both like in the abstract sense, should we use that data to do stuff like, or in the abstract sense of like, should we launch a new product or not? You could totally make that decision in a vacuum. And that's fine. And companies have done that for a long time. But wouldn't it be a little bit better if you use all that data to try and do things like, Hey, let's not send the customers who will, who will purchase from us, regardless of if we give them a coupon, let's not just hand them a coupon and lose money. So I don't think data science ever gets you like your company went from a failing company to a huge success of data science. But I do think like you can like on the margin improve your product a reasonable amount, and like make things a bit better by having data science. So I don't think like data science is the right thing for every company to suddenly grow huge. But I really do think especially for bigger companies with lots of data and lots of opportunities to improve stuff, that data can be just a very nice way to kind of flesh things out.

Data science consulting

You mentioned that you worked in data science consulting, you know, at one point in your career. I was wondering if you have any sort of tips, tricks, best practices that are unique to finding success as a data science consultant, and in the consulting world, dealing with a lot of non-technical stakeholders as well, the sales process of selling your value.

Yeah. So I will say, I do talk about the book. I have like loosely toyed with the idea of like writing like a, like a five chapter ebook just on consulting data science, because so many people have your question of like, I find this stuff interesting. How would I, a data scientist get into consulting? So consulting has a pattern like, okay, so super high level, how do, how do companies hire people? They have employees, full time employees, they have, it's generally called like contractors or vendors who are like, hey, we just need you for a certain number of months to just work as like our team. And then they also have like consultants, like more consulting consultants, like we're bringing you in for a very specific task. So it is possible to, as a individual, be in that last group of like people know you as like, you're a good person for this kind of stuff. So they're just going to bring you in. By far, by far, the hardest part of all this is finding, getting companies to know and trust you. And like, like finding the work is a thousand times harder than doing the work. For me, I had worked at consulting companies for many years. So like I had built connections with like companies I'd consulted for as part of one point. And then I was able to talk to them as an individual. So like a lot of it is just how that is how I built the network. And more than anything, what I recommend if you're interested in this kind of stuff is like try and find someone who would hire you, and then do it in your free time, like do it in your evenings and weekends and that sort of stuff. Because if you do it that way, it's like a nice you're testing it out your and if it works, that's a nice side income, which can be really valuable. And I recommend charging like real, like charge more than just whatever your hourly hourly is like double or triple that. And then if you do that, and do that enough times, and you like it and stuff, then you can start thinking about making it your full time life.

But that's a hard thing. I also think working at a consulting firm is a very valuable way to learn these sorts of things in terms of like, well, how do you make that first contract? And what is the sales cycle like, there's just a lot to that, that is unrelated to data science and just a standard consulting thing. But like, it just takes time to like figure out the beats of it. And so that kind of stuff, which is to say that being a data science consultant is very little about being a good data scientist and extremely much about being a good salesperson thinking through contracts, thinking about like, what's going to happen if this company won't pay you for four months, like, it's like that kind of management and all that kind of stuff, which can be fun and eventually a very financially lucrative if you can pull it off, but it is also tricky, and I no longer do it.

Being a data science consultant is very little about being a good data scientist and extremely much about being a good salesperson thinking through contracts, thinking about like, what's going to happen if this company won't pay you for four months.

That's super helpful. Thank you, Jacqueline. And as we get to the top of the hour, too, I just want to make sure to ask, what's the best way for people to get in touch with you? If they have follow-up questions, is it Twitter or LinkedIn?

A great question. I don't really check LinkedIn. Please Twitter DM me at Sky Tetra, S-K-Y-E-T-E-T-R-A. That is the best way to get it. And like a lot of my talks and blog posts and stuff, I have all my personal websites. If you just go to janeowles.com, you can see a lot of my past material.

Awesome. Thank you so much, Jacqueline. It was really fun chatting with you. I really appreciate your time, and looking forward to hopefully having you on here at a future session, too.

Perfect. Yes, this is a lot of fun, so thank you for inviting me. It's mostly just getting to rant about stuff I like ranting about, so it's great. Awesome. Thank you, Jacqueline. Have a great rest of the day, everyone.