DataTalks.Club

Home Articles Slack Events Podcast Books

Building a Data Science Team

Season 1, episode 3 of the DataTalks.Club podcast with Dat Tran

Did you like this episode? Check other episodes of the podcast, and register for new events.

Transcript

Intro

Alexey: Today we have pleasure to have Dat as a guest. Dat needs no introduction. If you have a LinkedIn account, you probably already know him. If you don't have a LinkedIn account, Dat has a lot of experience in building data teams, and this will be the topic today. So he led a team at Idealo. This is a popular price comparison tool in Germany. Then he was a head of AI at Axel Springer. This is a big publication house. And now he is the CTO and co-founder of Priceloop. So thank you Dat for coming to the show today. (2:06)

Dat: Yeah, thanks for having me, Alexey (2:53)

Dat’s early career

Alexey: Yes, we'll start with your background. So can you please tell us how you started your career? How did you get into machine learning? And how this all led to becoming a CTO of your own startup? (2:57)

Dat: Yes, sure. I would say, my career is not very straightforward. I didn't study computer science which would probably naturally grow into the area of machine learning. I actually studied business — I studied economics at Humboldt University and I was more into investment banking. (3:12)

Dat: But since I was a child, I was a lot into gaming, and due to gaming, I also got into coding. I started to do coding very early. I think I was 12 or 13, I'd create my first HTML websites. I did my own forum and other stuff, and then over the time, I got into more of these areas and then I learned to program as well. (3:42)

Dat: While I was in investment banking, I solved a lot of problems with coding and programming. It's like monkey business: you have to copy paste, you sit into 2 or 3am. And you copy and paste things. Where I just wrote a simple VBA script on Excel, which solved my problem within two minutes where my peers were working for three to four hours. And I was thinking, “Okay, maybe I should do something with software engineering,” (4:23)

Dat: It was well paid, but it's really a monkey business. Then I went back to my graduate studies I majored in, in operation research and econometrics. Then a friend of mine, he was studying statistics in Munich. And he told me, “Hey, well, there’s this cool course – machine learning. You should do it”. I think a lot of people did that course at that time. (4:57)

Dat: I think it was six years ago, something – six or seven years ago, I did that course that was like, “Oh, yeah, that sounds interesting what he's doing.” And a lot of the stuff that he said in the course, is something that I already did at university, but it was presented differently. Because now he called supervised learning for linear regression problems. So I was like, “Okay, so I know some of this stuff.” And now I have a little bit of machine learning context, but also – how you would actually apply it in the real world — for the business world. Because while you're studying just the concepts, you learn theorems. And you don't know how to use it in real life. (5:28)

Dat: At the end, I had to decide, “Okay, where do I want to go?” Should I do a PhD or go to the industry? Luckily, at the time, Accenture was having a new team called Advanced Analytics. Then I decided, “Okay, I should apply there, because it's a new team”. There I was not doing a lot of data science or something like that. It was simple statistics, but six or seven years ago everyone was into big data. (6:13)

Dat: So you had to learn Spark. In Spark, they had simple stuff like linear regression, and logistic regression. But that was not my biggest interest, because I knew at the time that big data was just very difficult to do. A lot of people just thought that it’s so simple to do, because, yeah… who brought up MapReduce at the time, and everybody was thinking, “Okay, in Germany, that would work as well.” (6:44)

Dat: But that was very difficult. That's why my focus was more rather into starting the data science practice in Accenture. I was one of the first ones who did a data science project here in Germany. But then I realized that a company like Accenture is not something that I wanted to work for a long time. So I left it pretty early. Accenture is a big company, it's a consultancy. There's still a lot of overhead that you need to do. For example, you need to do a lot of presentations, you need to talk to clients – it's more like product management. (7:10)

Dat: And then you have people in India, in Spain — in offshore centers, where people do code delivery. But I thought “I'm still pretty young, and I like to code. I don't want to spend my whole time just working on concepts and not doing the real thing.” (7:46)

Dat: And then I moved on after a year. I joined Pivotal. Pivotal is a US software company. The main focus was actually to do CloudFoundry. CloudFoundry is similar to Kubernetes. I joined Pivotal Data where you have databases like GreenPlum. It’s similar to Snowflake or Redshift. It's an MPP database. (8:06)

Dat: They also have a data science team, which teaches customers how to do data science. It's also consultancy, but more hands-on. At Pivotal I got to know one of the best engineers in the world. They taught me a lot of things – what’s programming, what’s test room development, DevOps skills, bringing data science into production. This was quite new five years ago. At the time everyone was talking about that, but no one really understood how you get a data science and machine learning model into production. (8:37)

Dat: I learned a lot about this. I devised my own ideas on how to make it happen. Because at the time, no one was really thinking about that. What I was thinking was: how do you create this fancy machine learning model? How do you do all the hyper parameter tuning? But no one really thought about how “What happens afterwards?” What happened after day two? Day one — okay, it's in production. But what's day 2, day, 3, day 5, day 6? Because it's not a simple app. (9:20)

Dat: Even an app – when you create an app, you have more feature requests, you have feature development, you have bugs coming up, and you need to think about that. And this is something that was pretty cool. Other than that, I also had to work on nice projects. I did my pet projects as well. I got into window vision a lot. And also I work on interesting projects, like, for example, Hydro planning prediction. I worked for many interesting customers. And I could also travel to many nice locations, because Pivotal was based in Silicon Valley. (9:55)

Dat: They had their main European office in London. It was a very cool experience that I had with Pivotal. I left Pivotal after two years. I realized, “Consultancy is nice, you see different customers, you get to know different problems, but actually…” As I said, at Pivotal, I started to think about machine learning in production. And I was like, “Okay, how do I actually find a company where I can test my ideas?” (10:29)

Work at Idealo and Axel Springer

Dat: I looked at some companies, which were interesting for me at the time. I applied for several companies. I interviewed for companies like Deutsche Bahn, Telekom and whatever. But then, none of these big companies were really interesting for me. At that time, I was actually looking for a Head of Data role, like a more managerial role. But when you just have three years of work experience, you never get this managerial role. They will ask you, “Okay, you're too young, you don't have the experience to showcase this”. There was this position at Idealo and the role required to have eight years of work experience. (11:07)

Dat: But naive me… Of course, I applied. And luckily, it went through. The CTO liked my CV. We had our first conversation. After this conversation I talked to the CEO, the CPO, and so on. The process only took two weeks. And the coolest thing is — while I was in the meeting with the CEO, and CPO and CTO, when we met on-site — I could do a little presentation about what you can expect from me in the next two years. (12:10)

Dat: That was a very interesting kind of experience — going through this job application. It was not like the usual one. I was literally creating my own role in a sense. Then I was hired as a Head of Data. I was responsible with two other co-heads, for areas like business intelligence, data warehouse, web analytics. (12:50)

Dat: While I was at the interview, I was pitched, “We need a data science machine learning team.” Idealo is a data company, but we haven't made use of all the data that we have at the moment. My two years at Idealo, from my perspective, were really successful. (13:24)

Dat: I had a nice team. We did a lot of open source projects. We kind of birthed a brand for Idealo — that we have a strong machine learning team. It's not like when I left, everything was gone. I tried to build a very sustainable culture in the company. I don't want to leave a company with everything, but I want to leave a place there that I built up. (12:51)

Dat: That was something that I really liked. But after two years of Idealo, I also realized, “Okay, I built up this team very successfully. What's next?” I cannot stay there forever. I was young. I was like, “Okay, I managed to do that. I did my learning but what is the next step?” Then I decided, “Okay, I did it for Idealo. But how do you do this on a corporate level? For a holding like Axel Springer?” (14:27)

Dat: Axel Springer at the time, one and a half years ago, was like… They didn't know what research was. It's not a company that is driven by research. Because they just didn't know what research was. But if you want to be a tech company, you really need research and an open source component within the holding. Then I approached Stephanie Caspar, she's one of the board members at Axel Springer. I was telling her, “Hey, these are my ideas. What if we try to create these main central functions? With evangelizing around Axel Springer, what AI is about, what research is about, and how we actually can work together.” (15:08)

Dat: One of the driving factors why I did that was I wanted to turn Axel Springer into a tech company. So it's not just taking the AI angle – to turn it into a more tech-oriented company. Then I did it. I did it for one and a half years. I built an AI team again. I took some people with me from Idealo — they wanted to come with me. I had a few people. And my role there was… sometimes it’s called a “machine leader”, because I didn't have a lot of people function. But more like, “How do you talk to the Managing Directors?”. Shape them so that they go into the right direction. (16:06)

Dat: It's much harder when you do this. Management is very difficult. It's not like you have a team and this team is one you've hired. But now you have many, many different managers on the same level. How do you make sure that they are going in the direction that you want them to go? Also, when looking back at my time at Axel Springer, there were a lot of ups and downs. The downside, of course, is that it's a big corporate. It was a big challenge for me. (16:52)

Dat: So I started thinking “OK, two years is enough”. To do this whole transformation, it takes much longer to do this transformation on an organizational standpoint. And also, one thing that I didn't think through was “Hey, to do this kind of transformation, you just need much more people under you to drive this transformation, and also kind of budget.” You need profit and loss ownership. (17:25)

Dat: But despite that, I managed to drive a lot of things. I managed a couple of open source projects. I was one of the initiators of the Axel Springer techcon, which is the first big tech conference that we had within Axel Springer. (17:54)

Dat: In pre-Corona when we did it, we had 700 participants from all around the world. That was one of the first big things where we could say, “Hey, we're driving Axel Springer in the right direction of being a tech company.” And also, the other thing we did was the Axel Springer tech blog — where people within the holding, within the companies can create articles and blog posts on this tech blog. (18:12)

Dat: This is something that is still living after I'm gone. I really like it, it’s still thriving. Because I started it. If I think about this, these are really small things, not very big things that I did. Everyone could do them. You just need someone to do that, to start this thing. You give people the freedom to write articles, go to this conference, and so on. (18:44)

Dat: Some of you know that I resigned from Axel Springer. When I joined Axel Springer, I thought, “Ok, I'm not going to stay there forever and I'm either going to do my own things or find a niche, managing director / top management positions, so that I can drive more things.” Because my credo is “Always know that you need more ownership and sometimes older ships come with power.” And power you either acquire through a major director position, or you create your own company. (19:18)

Dat: During Corona time the idea came much stronger. I was like, “Everything is so slow. Everything at Axel Springer is so slow. It's a COVID world. You cannot be there forever. Because you're still 32, not 45 – you don't have a family yet. So you really need to go out and think about — what's the next step?” (19:58)

Story of Priceloop

Dat: Then I was talking to a few friends. One idea was “Okay, maybe you go back to Vietnam.” I'm not from Vietnam, I'm from Germany, but maybe go to Vietnam and go to a consultancy, because the tech is really strong there, and maybe an idea grows out of this (20:26)

Dat: But luckily my current co-founder, Dr. Richard Schwenke approached me. We left at the same time from our companies. Richard co-founded Contorion, which is an ecommerce for tooling. He co-founded that company, he was the managing director, but he sold the company in 2017. And he wanted to create another company, again. He's also not that old, and he still really wanted to create a company from scratch. (20:48)

Dat: Because if you're too old, it can be very, very tiring, in a startup to create something from scratch again. He approached me with his idea of pricing because he created a data science team at Contorion already. They implemented a pricing algorithm. They've been dealing with pricing for three years already, and they had an uplift of 25% with the things they've been doing. (21:28)

Dat: They did a lot of A/B testing to find the right calibration, hyper parameters for the machine learning models, and so on. It's a nice idea — I had pricing at university. When you're doing operation research, you also will focus on revenue management or dynamic programming. It has a lot to do with pricing. Pricing is a decision that you need to do in a control theory. (22:00)

Dat: I was like, “Yeah, that sounds like a good idea.” He was looking for a technical co-founder. And I was looking for a business co-founder. So, it was a really good combination of us two. We had a discussion around August. And then it was clear, “Okay, I'm going to resign from the company.” And he was also “Yeah, okay, I'm gonna leave in October as well”. We left at end of September, both at the same time. And now we started Priceloop. So, Priceloop is my new venture together with Richard. And our goal is to disrupt the pricing industry. (22:26)

Dat: As far as you know, there's many AI software systems out there, also for pricing. Most of these pricing servers are actually more closed solutions. You get the data from your client, and then you put it into your system – maybe you have a login – and it's probably hosted on some cloud provider, and then you give out the price. (23:19)

Dat: That's what most of the services do: you give the data, and then you get the labels and so on. But what we want to do is, we want to create a pricing framework or library at the end. We want to give data science teams, pricing teams a pricing framework. So that people like you, maybe OLX, will use us in the future. (23:53)

Dat: It's so easy to create your own pricing strategies. Then, of course, we have commercial solutions on top of that, which we also will offer this to other customers. Because if they like to use our framework — we are a company — we also need to finance ourselves in some way. But the overall goal is actually to create a whitebox solution. (24:20)

Dat: We don't want to take away the pricing manager. We don't want to tell them “Hey, if you're going to use this, you don't need to hire a pricing manager or you can fire the pricing manager.” No, we want to give them a frame of a tool. So that they can make better decisions with their pricing teams. And pricing is a core component of many, many companies. And that shouldn't be a blackbox solution. That's where I am so far. It's a nice, interesting ride. (24:52)

Team at Priceloop

Alexey: A long story. But very interesting. What stood out to me was, first of all, you mentioned Andrew Ng and his course on Coursera. I think so many people ended up where they are now, because of that course. Including myself. Yeah, it changed the lives of so many people. I remember, I started following you on LinkedIn when you were already at Idealo. And your team contributed to so many open-source projects. I think there was an image quality library. I though “this team is doing great on the open source front, pushing out amazing stuff.” The projects also have a lot of stars on GitHub, which shows that a lot of people are interested. That's a great job. And I'm curious about your startup now. So you said you just started it. So Richard and you – do you have somebody else working with you now? (25:25)

Dat: Yeah. The way we started our startup is not very usual. So since both of us are experienced, we finalized our funding and already signed four people. One of the machinery engineers just started this week. Three more are coming. We're going to make two more offers. We want to be 10 people by Q2 2021. We got a plan to do our seed round in 2021. Then we're going to hire more people. (26:38)

Dat: Our goal is to create a strong tactical product team. Which focuses on disrupting one of the industries. We believe that the future is in open research, and contribution from outside and contributing into ideas for many, many different organizations. We see that direction from other startups like Hugging Face — a similar example. It’s getting so strong and people are using it in production. Because of the open research. And at the moment I haven't seen that in pricing so far. (27:25)

How to start building a data team

Alexey: That's an amazing topic. Many, many different companies, ecommerce companies will benefit from that. I know that it will all work out. So now you're already in the process of building a team. Some people already signed their offers, and soon start working. How do you start building a team? How do you approach this process? What should you do first? Do you first select a project or you start immediately with hiring. How to approach this process? (28:16)

Dat: It's hard to rationalize my mind. I would say it's a combination of both. Some companies just start with hiring people, and then build. And some companies, they need a big, big plan, and then they're going to hire people. Our approach — this is together. I'm driving it from the technical perspective and we have a project. We know what we are going to build, but it's still unclear. Unclear in this way that we don't know. We know what the end goal should be – like the vision. (28:57)

Dat: But we just don't know, which features will lead to this kind of thing. We are hiring for different roles that would take us to that point to get a better understanding of our vision. We’re building like an open framework. Like a library. Which means it's a strong software engineering project, which means we need very good software engineers, who understand how to create abstract libraries. (29:40)

Dat: Since we're dealing with machine learning, we need machine learning engineers. Since we're dealing with data, we need data engineers. We need a product manager who will prioritize these kinds of things. We need designers who will guide the API. We will also need a front end for the commercial solutions. Which means we need a UX/UI person who will drive that kind of thing. (30:12)

Dat: There's a lot of roles that need you to think about before. In the beginning you also need to think about – do you need very experienced people or inexperienced people? Also generalists with specialists? This is the question that you really need to ask. At our stage we really need more experienced people, because we are an early stage startup. We need to get traction as fast as possible, so that we can raise on next funding and also get this product market fit with our customers. The second thing is, do we need a generalist versus specialist? Now we need more generalists, because as a start up, when you start, you have no lines of code. There's nothing, which means, you know, you need to do back end, front end, DevOps, and whatever. Whereas when you're a specialist, you focus more on things like “I just want to tune this specific hyperparameter.” (30:39)

Specialists vs generalists

Alexey: It’s an interesting discussion – this specialist versus generalist – and I'm wondering. Let's say, if you were still at Idealo. Who would you prefer to hire back then? If you wanted to hire somebody in your team? Would it be a generalist or specialist? Or would it actually matter? (31:45)

Dat: When I started at Idealo or? (32:06)

Alexey: Let's say you have a team – you're working already in a big company. In a startup, it's clear — you want to have generalists – people who can do pretty much everything. But let's say it's a midsize company, like Idealo – it's not a large corporation, but it's already not a startup. For these kinds of companies, who would you prefer to hire? (32:09)

Dat: I think it depends as well where you are in the organization transformation. I think there's this graph where you see how data-driven the company is. If the data is very immature, like at the beginning, they don't have a data analyst or a data team. Then it makes sense to hire data analysts and data engineers who build up the kind of backbone of this. (32:33)

Dat: And then over time you can start to hire more different roles, like data scientists or machine learning engineers. Who will bring up the business intelligence. There was this famous pyramid, where on the bottom, you have data. You have very messy data — you clean the data, and then on top, you have this very thin slice with intelligence and then the machine learning part. (33:03)

Dat: If I would map it to Idealo. Idealo was not very mature, but also not completely immature. It was in the middle of this transformation. They had a data analyst before – they had business intelligence people – they also had data engineering who work on a very old database. And then you see that you have to compliment this. That means, you don't need generalists in data science, you need more specialists in data science. Because the topics are there, but you need people who work on that. (33:35)

Dat: Of course, you don't need super-specialists. They need to be a little bit towards the level of generalist-level. Idealo was very new in machine learning, which means, these people I hired, also needed to put things into production. They needed to cooperate with data engineers. Because the data engineer didn't understand what machine learning was about. You really need to have empathy to work these people together – to bring things into production. (34:15)

Dat: For example, if Idealo would’ve already been at that stage – they already have machine learning in production, they know how to use it, then it makes sense to go towards this super-specialist, which would mean more research-oriented jobs. Because they’re only researching and not really like taking care of other stuff. (34:44)

Who to hire first

Alexey: Going back to your current company, Priceloop. You mentioned you want to hire a lot of different people. You want to hire a product manager, you want to hire a frontend engineer, backend engineer, UI/UX designer, data engineer. You said that machine learning engineer also is starting soon. How do you decide who to hire first? Or you know already who to hire – five different roles and you just start hiring? Or you'd rather focus on one specific role first? (35:06)

Dat: Yeah, we open a couple of roles. In my head, we have to start with certain roles. Our first goal was to hire machine learning engineers and software engineers. And actually machine learning engineers, who are very close to being software engineers. Actually software engineers who know a lot about machine learning. Then we can hire product managers and data engineers. At the beginning we need people who can work on the prototype, who will work on MVP, who do a lot of coding, who will work on the product. It doesn't make sense for you to hire like a UX/UI designer, when you have no work for them. (35:41)

Dat: So you really need to understand, at which stage you are. And what kind of roles do you need now to solve this problem. Then also, who would you need in the future. You cannot hire someone just for six months – then you just need a freelancer – but we want to create a company that is more sustainable, longer. We know we want to keep these people longer. (36:25)

Alexey: So what you're saying is start with hiring engineers – backend engineers who know machine learning, and then they will build the backbone. They can also probably take care of data engineering and all these data pipelines. Then you add on top of that, maybe analysts, UI/UX, product managers, but first you need to have this backbone, and then you need to hire an engineer for that. (36:52)

Dat: Right. (37:21)

What is a strong product team

Alexey: You mentioned a couple of things previously. And one thing that stood out to me was – you want to build a strong product team. What does that mean to you – a strong product team? (37:23)

Dat: A strong product team for me is a team that is building a product that the customer wants. Strong means very customer-centric. Which also means we deliver features very fast. And test these things out with our customers very fast. I want to build a product that a customer or user would say, “I love to use your product!” It’s the same thing, you want to create these libraries and put them open source. We're creating libraries, so people will say ”Wow, the thing that you built is very useful for us.” This is something that I would like to hear in the future. (37:44)

Alexey: Being customer-centric, being able to iterate fast, get this feedback, and make sure like you have this feed that customers really want to use what you're creating. But how do you make sure the team can do that? Is there any secret sauce? (38:30)

Dat: I wouldn't say there is a secret sauce – it is just how you create the culture. If you look at high-performance teams and the culture with their managers, it actually just boils down to the culture that you create in this environment. For me, we don't want to do a bullshit bingo – like a scrum bullshit bingo. We want to have people who work towards a mission. Who like the job. So we need to keep these people motivated. Do everything, as much as we can, so that they can work on the problem. (38:49)

Dat: And help them when they have problems and when they get lost with the vision. Telling them again, “this is the vision that we want to go, this is direction.” Have very short feedback cycles. Also, allowing them to do open-source stuff. Not a lot of companies here in Germany and overall – are contributing to the open source community in some way. Or are doing stuff that is open. This is something that I rarely see in startups as well in companies. There are big corporations who do that. But overall in Germany not a lot of companies are doing open source. (39:31)

Motivating team to write articles and contribute to open source

Alexey: Yes, that's definitely true. With this open source, many developers want to do this. But when it comes to actually doing this… sometimes it's difficult. Do you try to give some extra motivation? How do you motivate people to actually go ahead and release something to open source? Or with writing articles, it's also something people want to do. But it's often difficult. You want to write an article, but then you end up doing something else instead of writing. How do you motivate people to actually do that? (40:16)

Dat: If you look at my past teams, most of them before joining, never did anything open source or wrote an article before. I'm a very pragmatic manager, so I do one-on-ones. In this talk, I give them a suggestion, “it would be nice, if you wrote something like this.” Or “this would be nice for the community, if we do some open source like that.” And I just talk to them and give them inspiration. The rest is up to them. First, some of them started to write and then they were stuck and I was telling them, “maybe you could do something like this, then do that.” Then they just take that as an idea. (40:56)

Dat: If you really want to create this kind of culture, you need to work with people. You need to give them inspiration. Some of them don't have the courage to start with that. If they found the courage to start, they may be a little bit clueless. There's so many articles out there – where do I start? As a manager I wrote a lot of articles already. I could start with telling them, “if you want to do that, you could look at some of these articles, I think they are good. Try to do the introduction like this, or in the main section a diagram like this would be nice.” Give feedback and then it works. (41:41)

Alexey: Basically by setting the example. You said you already did that in the past, and then you’re just sharing this experience, sharing this motivation that you had with the team, then it gets contagious and people just start following that and doing that, right? (42:25)

Dat: Another example is conferences. Before I created this team at Idealo, Ideal was never at a machine learning or data conference. If you're thinking about the last two, three years, I could see that many Idialos went to conferences. I was very proud of Chris Ovanenin. He was one of my first hires. He spoke at Strata San Francisco. There were two German companies – Idealo and Flink. I was really proud of that. “Wow. We made it to Strata, San Francisco.” (42:43)

Alexey: It's a pretty high bar. (43:23)

Dat: Yeah. (43:25)

Hiring a data scientist

Alexey: That's awesome. Coming back to the hiring process. So you need to hire engineers to make sure that the infrastructure is there. The process for collecting data is there. But at some point, you want to hire a data scientist. How do you do this? What is the process like for you? How does it look like? What are the qualities you check? What are the things you're looking for in data scientists? (43:27)

Dat: My checking is driven by how I feel. I’m looking at the CVs. I don't have a checklist where I say, “Yes, yes, yes, yes.” “He studied at Stanford or Harvard or whatever”; “He did computer science”, “He has a 4.0 GPA”, and blah, blah, blah. I look into the team. If I already hire someone with this quality, I have to look for the CV with someone who has a different quality than the other one before. Of course, they are similar in some ways and they need a basis. So the basis is – you need to know how to program. This is 101 for me. (43:57)

Dat: If you don't know how to code and especially if you don't know about software engineering, you are already out of my process. Unless you are a junior. When you’re a junior it’s a little bit different, but still I require people to have very high coding skills. Other than that, I look at stuff that may be interesting for the team. “That person studied mathematics. Wow, cool. That means that person knows how to do math.” or “That person already did some open source projects.T they know, the open source process.” or “Somebody already did a Kaggle challenge,” — like a real challenge. It means “the person can work under pressure, it's competitive”. (44:45)

Dat: Then I combine these things together, and I say, “this could be a good fit to the team,” and “It could be a good fit to the skill set that we're looking at in the future.” Then I do a first interview, where I just talk about all the experiences, whether that person is interesting. We are also people. We have to work together. If someone is just plain boring, this is very difficult for the team. That person also needs a hobby. I don't know, go to the cinema, do some sports, or whatever, hiking is also healthy. But that person needs to do something. You work more time with that person than you spend with your girlfriend or your wife. You spend more time with them than with the person that you love. So that's why you need to understand that person really well. I also ask 10 basic machine learning questions. Some of the questions are in your interview guide that you have on GitHub, so I'm not going to discuss them. (45:37)

Alexey: There are 160 questions, so 10 of them are there. So if somebody goes through all of them, then they will pass your interview. (46:49)

Dat: Yeah, but they should not copy your answer one on one. Because… (46:59)

Alexey: It's a good idea to actually look at the questions and try to answer themselves. (47:05)

Dat: Right. But what I do is, I ask one question, and then from choices I ask a random question. (47:13)

Alexey: And when it comes to coding – to programming – is there any specific process that you follow for checking? Or how do you do this? (47:21)

Dat: The second interview is a homework assignment. I send out a homework, which is not very difficult. Then they send me the code, whether it is Jupyter Notebook or whatever. Then I check it. From this simple task, you could already see how much people are working. For example – quotes. Some people don't make sure that quotes are the same everywhere. Like double quotes or single quotes. When I see that people are using single quotes, near a double quote, then they have a single quote, again, I would already see — that person is not really taking care of the code. (47:31)

Alexey: Small things, yeah. (48:19)

Dat: Small things, these kinds of small things, you can always see. Also, naming. How does this person do the naming? Whether the person does some extraction of classes well? Or is the person using a pipeline? From this simple task, you could already see how someone would work in the future. There are small things that I check, because these small things make a difference at the end. (48:20)

Alexey: So basically, whoever is listening, if you want to go to Dat’s company, make sure you use the same quotes throughout all the code. (48:44)

Dat: I don’t think it’s for my company – I think it’s for every company (48:55)

How to pick the next projects to work on

Alexey: That's interesting… It didn’t occur to me to look at these things. But that's an interesting perspective. We just wanted to remind you that you can ask Dat a question. You can go there and ask Dat a question. And we already have one question. The question is – “For a company that already has an established data team, how do you decide which project to take?” You probably have a big list of different projects initiatives – how to pick the one to work on? (48:59)

Dat: This is always a very difficult question. It's risky. Let's say you have 100 projects. You have only limited resources, which means you need to pick the one that has the highest return on investment. What I do is – I have this matrix. A two by two matrix. On the Y axis, you have the business impact. And on the X axis, you have the technical feasibility. Then from these two dimensions, you can map up some of the projects. You go into your different dimensions. You're thinking – this is impact, is it impacting revenue or cost? So these are the two driving factors and cost revenue. And you can also distinguish it as well. (49:51)

Dat: And the technical feasibility — is there a lot of legacy involved? Do we need data engineers? Do we have a data dictionary? Is the problem solvable? If you think about self-driving cars, it's not an easily solvable problem with just data science. You need much more than data science. You need hardware, you need infrastructure, you need a whole ecosystem behind that. Then from there, I would just prioritize this list and then look at the top three. And then work on the top three. Very important: don't work on just one project. Because if you work on one project for one year, and it's going to fail, you're going to fail with this one project. At Idealo, you only saw the successful project that we open sourced. We had a lot of projects that no one saw, because they never went live. It's fine. The only thing is you need to decide on a project. Because you think that it has a high business impact and it's technically feasible. Then you also need to iterate fast. To work towards the goal very fast. And then if you see that this thing will not work and I’m failing, then you should really cut it down and try the other idea. Fail fast. (50:42)

Alexey: Okay, iterate fast, fail fast. It brings us back to the topic we discussed previously of strong product teams. So this is one of these aspects. We have a couple of questions. I'll share my screen now. (52:10)

Questions to Dat

Alexey: Question from Pratap. “If I'm about to set up a complete data science, AI team in a product space – from where I need to start with?” (52:32)

Dat: I answered this a bit already with my Priceloop question. You need to think about what's your product. Is your product a software engineering project? Then you need software engineers at the beginning. If your project is just a consultancy, then you can hire any role. (52:54)

Alexey: A question from Kai, “How do you see the role of corporate IT regarding data science?” I'm not sure I completely understand it. Do you have an idea what corporate IT is? So probably maybe something like in companies like Axel Springer? (53:18)

Dat: In the corporate world, in a company like Axel Springer that has corporate IT – I don't think an corporate IT system makes sense for a company like Axel Springer in the future. Axel Springer is turning into a tech company. Everything that is within the company is driven by technology. So there will not be this central corporate IT department. And the corporate IT will be like a DevOps role — within a whole technology company. And then data science will play a part within this technology organization. I hope this answers the question. (53:35)

Alexey: Another question is “How do you keep a good team? Good people tend to get great offers and might leave soon. So how do you keep them?” (54:23)

Dat: This is the question that I always had so far. I didn't have the problem of people leaving me because I was always ensuring that people are getting paid fairly. And also have interesting projects. If you are going to balance these two things out, then you can keep the people. When you start to give people shitty projects, or when you micromanage them, then they will leave you someday. Also when you're not gonna pay them fairly. Then this is also a big problem. That is something that I learned over the time. This is what kept the people working for me. Very simple ingredient actually. (54:43)

Alexey: So, two things, pay well and give interesting projects. Dat, will you prefer a mathematician, or a computer specialist for a machine learning position? And computer specialists, probably somebody who graduated from the computer science department. (55:32)

Dat: It doesn't matter. If you are a mathematician, you also need to code. Which means you need to be on par with a computer scientist. If you’re a computer scientist, you need to understand the math behind machine learning systems. Which is not so complicated. And then you know – it doesn't matter, actually. (55:50)

Alexey: So you need to have a certain set of skills, and it doesn't matter where you pick the skills, right? Was it from your university… (56:13)

Dat: But you could also not study at all? Right? There are many who didn’t study at all and they are very good, so.. (56:21)

Alexey: Yeah, thank you. “How to deal with hype on management when building a data science team?” So it's probably like a question of expectation management. (56:28)

Dat: This is the problem when you are a company and you create this new data science team, everyone will expect a lot of you. They know that “Wow AI!” They read these things – “They can do so much! We’ll increase our revenue and cost” and so on. If someone is opening this new data science team, you really need to communicate as much as possible. I also do a lot of education for higher management: “this is what I can do, and this is what I can't do.” Otherwise, you will have so many expectations that you're expected to fail in some way. (56:40)

Alexey: So basically work with management and explain them. Cool. “What do you think about this establishment of the data product management role?” So I think this is a question about this, maybe a new trend? Data product manager? What do you think about this role? (57:21)

Dat: I would like to have it here in Germany. Most of the product managers are not really good. There are some really tough product managers, who understand what data-driven is, and they understand what machine learning is all about. But most of the product managers are business-driven. If you're a product manager, you should also be tech-oriented. In the US, there are much more tech-oriented product managers. They were software engineers, they were machine learning engineers before, and then they became product managers. Here, there are many people who are doing marketing and then they also did product management, and now they’re doing project management, but project management is not a product management role. Managing a project is different from managing a product. (57:48)

Alexey: The last question is “How to start doing data science in a new company, when data quality and organization in the company is not good? What are the required steps before starting hiring?” (58:49)

Dat: It’s difficult, if the data is not really good, it can be very challenging to create a data science team. If you're doing that, if you are at this stage, then you should hire data engineers to clean up the data. Also, have a proper data quality process. If you don't have clean data, you can create new data. For example, for a lot of our projects at Idealo, we didn't have the data. We collected the data and this takes. Sometimes it takes half a year or a year to collect enough data to solve a problem. But you have to start someday, because many companies just think you come in as a data scientist, you think that we have the data already. And then we're just gonna do that. But it’s not going to work this way. (59:03)

Alexey: Yes, I can only agree with that. And It also brings back to your point that when you have this backbone with data pipelines and all that before, you know, thinking about machine learning and hiring data scientists. (60:08)

Alexey: Yeah. Thanks a lot for taking time to come here and share your knowledge with us and your expertise. Thanks a lot and thank you everyone for attending and you questions. And we will put the video out soon. And yeah – that’s all, I think. Any last words from you? (60:19)

Dat: No. Thanks for having me. I think it’s always nice talking to you Alexey. See you then someday live in person – after Corona. (60:40)

Alexey: Hopefully it will be soon. Good bye. (60:50)

Subscribe to our weekly newsletter and join our Slack.
We'll keep you informed about our events, articles, courses, and everything else happening in the Club.

DataTalks.Club Hosted on GitHub Pages