Failing Fast With Data Science and ML

Read Interview

Alex Mysak:
Welcome to the Champions of Data and AI series, I’m Alexandra Mysak, your host for today’s episode. In this episode, I’m joined by Dan Jeavons, the VP of Computational Science and Digital Innovation at Shell. Dan will share his views on how persistence and a fail fast approach regarding experimentation is helping a global brand like Shell transform their business across multiple industries from energy and commodities through to retail and commercial real estate, all with the applications of data and AI. Let’s get started. Dan, welcome, it’s great to have you with us today.

Dan Jeavons:
Thank you so much for having me. It’s great to be here.

Alex Mysak:
So it is a complete coincidence that for one of the first podcasts that I’m hosting, I’ve been matched with a coat red. So I thought maybe as an icebreaker and also to introduce a little bit of Brit to our audience here, we could start with a quick British pop quiz, if that’s cool with you?

Dan Jeavons:
Love it, lets go for it, sounds great.

Alex Mysak:
Okay. Favorite chocolate bar?

Dan Jeavons:
Well, it’s got to be Cadburys, right? I mean, there’s no other chocolate.

Alex Mysak:
Which specific one? And that’s a wide variety.

Dan Jeavons:
So I’d go with Dairy Milk to be honest, although they messed up the recipe recently so it’s a very controversial topic and it got taken over by an American brand, of course, who do it wrong.

Alex Mysak:
Ooh, careful. And then which Brit, past or present, would you invite to a dinner party and why?

Dan Jeavons:
So that’s a great question. I’m going to be a little bit odd here so my hero growing up was Lord Nelson. That might sound like a really weird thing to say. I just was completely fascinated by the sort of persona and this guy that was on top of a column in Trafalgar Square and everything that went with it. So I think I’d go with that. I’ve kind of re-discovered this whole thing because now my kids are really into this story and they like to have it read as a bedtime story. So Lord Nelson.

Alex Mysak:
Was there anything specific about Lord Nelson that inspired you?

Dan Jeavons:
I don’t know, It’s kind of the heroic personality. When you read about him, it’s just remarkable. I mean, he just kept going despite the ridiculous levels of adversity and the fact that at almost every stage he was likely to lose. And so I think it’s that kind of, and maybe there’s an element of that my personality has picked up, we’ll touch on that later on, but that kind of insistent desire to overcome the odds, I guess, I found quite inspiring.

Alex Mysak:
Yeah. Yeah. I know that we definitely chatted about resilience in the face of adversity when we’ve spoken previously. So I look forward to getting to that. And one thing that I recognized last year in hearing you speak for the first time in a data context is the massive reach of the Shell business which goes far beyond anything to do with oil and gas. And I think to set the context for the viewers today, it would be really nice for you to just showcase a little bit of an introduction about Shell, because I actually see them as being equally akin to a retail business, for example. And so then that gives much greater replicability to maybe some of your learnings and takeaways that you want to share today.

Dan Jeavons:
Yeah. And I think that’s exactly right. And I mean, I think it’s getting even more interesting if I can say so. And it’s simply because the global reach and the level of the network that we have, but it’s also growing in all sorts of other dimensions. So to give you an idea, now we operate more than 80,000 charge points so these are charges for electric vehicles and we have access to about 225,000 in terms of tracking where they are and some of those with our customers as well, of course, now. And so it gives you an idea that almost on any dimension, and that’s just our retail business and our EV business. And then you’ve got our producing assets, you’ve got wind farms, solar parks, you’ve got large real estate portfolios. Of course, a big trading business. I can keep going, but almost on any dimension, the scale is just phenomenal. And that’s one of the things as a data guy that I find so fascinating is there’s so many different data problems right across the entirety of Shell’s value chain.

Alex Mysak:
Yeah. Well, it’s a fundamental commodity at least to the world today and you actually mentioned some of the future as sources of energy as well. To give you a little brag potty for a second, just again to maybe marry those two things together so the background in Shell, and then sort of the application of data, we’re going to talk about building a team in a second, but just to showcase for a second, everything from the energy suite to the retail suite, some of the projects, just tangible projects that you’ve worked on and ways that Shell is using data today.

Dan Jeavons:
Yeah, for sure. I mean, I think maybe to start at the beginning, I think I’ve been on this data journey, I guess the current incarnation of it anyway, since about 2008. So it’s been sort of a growing passion and over this time we’ve seen a massive acceleration in Shell’s development of AI and data science. And now really, and we’ll talk more about this no doubt, but this is infusing every part of Shell’s business. So we’re using AI in the subsurface space. So when we look at the oil and gas that’s in the ground, we use AI there to better understand seismic surveys, we use AI to improve production processes, to optimize them, also to predict failure events in pieces of equipment.

Dan Jeavons:
We’re using AI in our retail business to provide differentiated offers to customers that give us better understanding of what customers want from Shell. And we’re using AI and in new energy solutions in order to start to optimize the layout of wind farms or to develop new solutions to charge vehicles more smartly. And I can keep going for a very long time, but it gives you an idea of the sheer breadth of the sorts of things that my team covers.

Alex Mysak:
I mean, it’s just awe-inspiring, which is why I asked you to have that little brag party for a second to see the range of application, particularly when so many friends are very nascent in this journey, which is really the thing that I wanted to showcase with you here today. I guess another angle on that same question is, and you mentioned 2018 as the first point of the journey for you with Shell and data, how did you scale this team of one to the successful team that you have today and perhaps some learning lessons along the way and the initial pitfalls that you’ve had to overcome?

Dan Jeavons:
Yeah. Well funnily, I would actually put a few milestones on this and I’m really reminiscing on this today because I’m going right back through my career, which is kind of fun. But I always talk about three milestones, so 2008 was when I actually left Accenture and joined Shell and the reason I put that as the milestone is that even back then the focus, the reason I joined Shell was it was all around, how do you use data to transform business processes? And so that was kind of the initial thing that I came in to do.

Dan Jeavons:
But then through a whole series of different projects which were largely analytics led, in 2013, we actually created the first advanced analytics center of excellence. So that was the first time we said, “Look, we’re going to really double down on this data science thing and we’re going to put a team on it,” and I was asked to lead that team and I was employee number one. So back then it was just me. But you’re right also to call out 2018, because I think in that 2018 time period, we really kicked off in earnest the creation of a digital center of excellence, which has been almost the foundation of catalyzing rapid growth in our digital journey and then starting to really accelerate and making that a, if you like, executive committee level sponsored program, which has led to a rapid growth over the past few years. So a number of milestones along the journey, but I think the key thing is actually, it’s been a building momentum within Shell over many years, which I think is why we’ve reached the scale that we’ve reached now.

Alex Mysak:
Yeah. Actually, there’s something I really love about what you described there, because it is a typical pattern that we see across customers, which it takes a certain amount of critical mass before then you start having the executives focus on data and AI. If you would allow me just to back up a little bit, and you mentioned that the initial projects were targeted to change your business process, so I’m guessing that you maybe had a business outcome in mind and then you backed the project into it, could you give us a couple of examples of that? Because I know many of the viewers today are literally at that point where they’re trying to get some of the initial projects off the ground and the stakeholder buy into those projects.

Dan Jeavons:
Yeah, for sure. And I think I was very lucky. So one of the themes throughout this is mentorship and leadership around me, and I think I was very lucky when I started that the people I was working for were ruthless on saying be clear on the business outcome and be clear on the problem statements. And I’m happy to dwell a little bit on also some of the things that I got wrong in that space, because there’s a lot of things, we’ll come to that later on. But I think like where we’ve been successful is where we can understand who is the person that has the problem, that does something different as a result of the insight that you generate, and importantly, are they going to really use this as part of a work process and how does it fit in that? And I think the projects where I’ve been successful, I’ve been able to really put my finger on those things and be very, very clear.

Dan Jeavons:
And maybe give an example, I’m in example mode so just to make it tangible, one of the first big successful projects that we had, and we’ve talked about this with Databricks before, because of course Databricks was the underlying technology was around spare part inventory optimization. And it’s such a great example of what we’re talking about, because what we were able to do is say if we can identify a way of better predicting two numbers, in this case the minimum stock level and the reorder point, and we can use all of the data history that we have and all the things we know about this space in order to generate those numbers, the inventory analysts will be able to give a better recommendation to the reliability engineer, which will result in a reduced or optimized stock level for Shell.

Dan Jeavons:
So very clear who uses it, how they use it, where it fits in the business process and how the impact is going to be generated and what the financial impact of that was going to be. And I think it’s that clear line of sights in these projects that leads to success and ultimately also gives executives the confidence because they can understand that, they get how this is going to generate the impact.

Alex Mysak:
One of the best lines that I think I’ve heard you say is if you don’t fail, you’re not part of the learning. And we all know that 85% of data science models don’t make it to production so this is the pain point. I would love to know whether you think Shell is higher or lower than that, I’m guessing lower given the success that you’re having, but I also know that when you’ve democratized data, it didn’t always go that smoothly. So could you perhaps think about that as a learning lesson, just in the context of what you just shared?

Dan Jeavons:
Yeah, of course. No, so what’s interesting is the art of problem definition is non-trivial. And what I’ve just described with inventory optimization seems very logical and simple, but actually in the majority of data use cases, it’s not easy. And I’ll give another example of where I really got it wrong, we were engaged in a project which was all about supporting traders with additional information about things they were buying. And the concept was, if we can enable the trader, so who’s the person, the trader, if we can enable a trader to have more information about their purchases based on additional data that we scrape from different sources, they will be able to generate more value because they’ll have a better understanding. It sounds very logical, sounds like the inventory optimization scenario.

Dan Jeavons:
The issue with this particular scenario is that the underlying assumption is if I give this person more data, the outcome will be better. And it comes down to a lack of understanding of how that business process really ran. So going back to my process point, if you think about how that process ran in practice what was happening was that trader have access to all this information, they just got it in a variety of different ways. And we didn’t understand precisely how we could drive a better decision by giving more specific information to that person in particular contexts. So we tried to solve the problem broad, give them more access to data, and what that resulted in was no additional value. And I will say this was a very unsuccessful project, we built a brilliantly engineered solution, it had all of the cool things in it, it had know natural language processing, it had really advanced data lake technology, we were using the knowledge throughout, it was very, very cool. And as an engineer and a technologist, I was so excited.But when we put the thing live, no one used it.

Dan Jeavons:
And the reason why is we’d missed the boat in what that user really wanted and where the value was going to come. And I think that the point is that hopefully you don’t have too many lessons like that, but I think these things teach you and help you understand how you get good at some of these core disciplines, like problem framing. And to answer your question directly, where are we now? I mean, I think I hope we fail less than 80%, I’m pretty sure we fell less than 80%. I think a good failure rate is healthy because we don’t always know if it’s going to work when we start, but at the same time, I think it’s about failing early, that’s the key thing. So we talk a lot about minimum viable products, can we test something? Can we test the worst thing possible that we can to see if there’s value there before we do the engineering?

Dan Jeavons:
And that is deeply unpopular, it’s not something that people are very comfortable with, but actually it’s really important because the engineering is essential to take it to scale, but you first want to prove whether the value exists or not. And if you solve it, whether somebody is going to do something different. And the lower cost of failure that you can have in that space, the better. And what we’ve also tried to do is to also make sure that we take learnings from things that haven’t gone well. And so we’ve had awards for people that have taken risks and haven’t succeeded, and that’s also started to build a behavior where people are comfortable taking those risks, but where also we try and learn from it and embed those learnings into our processes. And I think that’s also part of the cultural aspect of what we’ve been trying to do.

Alex Mysak:
I love that idea of having successful failures rewarded in some way, that’s very British. The other question I’d love to ask you about learning, I was speaking to one of your peers last week, and we were talking about the global challenges for the business of getting data to the cloud, particularly in the regulatory environment that you’re operating in, similar to them. They’re very nascent in that journey for global management of data, she explained to us, and I would love to hear from Shell. It sounds like you might be a little bit further along in that journey. And particularly when we’re dealing with regulated businesses, we’re seeing that being one of the key sticking points for companies being able to scale their enterprise data strategy faster. So I’d love to hear sort of any learnings from Shell that you’re able to share on this point.

Dan Jeavons:
Well, I’m going to give some credit here to one of my great mentors, a guy called Johan Krebbers who was really instrumental in this. And he was my first boss when I started looking at this whole space and he was super clear from very early on we need to go cloud, we need to go public cloud and we need to go as fast as we possibly can. And I think that drive towards clouds has been very much part of our journey. So we have been from, I think, 2014 onwards, everything I’ve done has been in the cloud pretty much.

Dan Jeavons:
And I think what’s interesting about the regulatory point that you raise is that there are many, many areas within any business like ours where you can’t put things in the cloud for good reason, and I can give you a thousand of these things. However, the vast majority of our data, there is no reason why we can’t. And so what we tried to do was say, “Look, we’re not going to touch this stuff where there’s a regulatory problem, but anything where there’s not, we’re going to go really hard. And when the answer comes back as no, we’re also going to ask why and where there’s not a good reason we’re going to push to put it in the cloud anyway.” And I think that clarity of purpose early, which helped us overcome some of the barriers as well as the clear drive, the clear direction, and the focus on the non-regulatory environments has enabled us to get ahead, I think, in our cloud journey.

Dan Jeavons:
And just to give you a sense of this, if you look at our global asset operations now, we are running our core backend system, which is our real-time data lake, all in a cloud native environment. I should say, all enabled by Delta, but actually underpinning that-

Alex Mysak:
Thank you.

Dan Jeavons:
– is a very powerful… No worries. Actually, it’s been a very important technology for us, and I’m happy to unpack that a bit as well.

Alex Mysak:
So Dan, I’m actually going to take you up on that and just backtrack. So what was the importance of Delta and then Lakehouse within that global architecture that you’re building? It’d be really interesting to hear a little bit more about that.

Dan Jeavons:
Look, it’s a great question. And I think we got very excited about the potential of Delta early on, because I think for us, most of what we do is time series. So a lot of our core business is based on time series data, and being able to aggregate that data at scale, and to move away from what we’ve had to do historically. So I’ve always said I don’t understand why I have to have multiple different databases running in parallel with the same data in them in order to run simple queries. And if you look at some of the early architectures that we were putting together in these spaces, you often had eight or nine tiers of ETLs going on. And I’ve been around long enough to have been involved in designing their systems and it’s deeply frustrating and of course we understand why, there were technological constraints, we didn’t have the cloud. There’s all sorts of good reasons why we ended up with architectures like that.

Dan Jeavons:
But as you will know, the more points of transition that you have in a data architecture, the more points of failure. And so my question has always been, “Can we not just go back to the raw data and run queries on the raw data at scale?” And I think this is where, again, another shoutout to some of my colleagues like Bryce Bartman who’s been working really closely with you guys at Databricks to start to help you understand our requirements for this and to embed Python and SQL and these things into a lakehouse architecture that allow us to both empower analysts to ask simple questions. So the engineer that I talked about earlier who wants to be able to run a query across the performance of equivalent equipment types globally, why can’t we do that? It’s a simple question. It should be achievable.

Dan Jeavons:
And I think actually the reason is scale, that’s an extremely large query, and having that lakehouse architecture allows that sort of capability, which is extraordinarily powerful. But at the same time, what’s also really great about it is the ability to bring together both the analyst and the data scientist onto a common foundation. And I think for me, that’s been the power in it that we’re able to say, well, actually, no matter what role you’re playing, you can use the same data sets and hopefully therefore come up with similar or the same answers without running off on different query environments that then you have to reconcile at the top level.

Alex Mysak:
It’s really interesting to hear you talking this way, because even just this week, I was speaking to a chief data officer of a major asset manager and she was describing what the challenge is to be bifurcated, data assets, because they’ve grown by acquisition. And then it’s the production code on top of those assets. So any advice for, I guess it’s like, what is the urgency to get away from a production environment that is working, but you know that you need to futureproof further and what causes that urgency to accelerate, do you think?

Dan Jeavons:
It’s this constant challenge that you have in a large organization between agility and autonomy and scale and replicability. And I think that the fundamental problem is one that you want to enable the agility, but you also don’t want a thousand different solutions because then you can’t get the benefits of your scale. And so there’s this constant trade-off, and I don’t think any large company will claim to have the answers to this. What I can say is what we’ve tried to do, and I think one of the things that we’ve been focused on is saying, “Look, let’s create a common foundational data architecture that runs on common platforms across our business and no matter where you sit in the organization, you subscribe to those same core principles.”

Dan Jeavons:
So for example, if you’re an asset in Shell, your data will be aggregated into a common environment, which we call SSIP, the Shell Sensor Intelligence Platform, which will allow you to bring all that data into an integrated way and therefore anyone can query anybody else’s data within that same environment in an easy way. Now on top of that, there’s going to be 45 different things that are happening on top of that and those are going to be run by different product teams and they’re going to be different solutions, it might be a digital twin or predictive maintenance algorithm, a query for an engineer, a visualization tool like Power BI and all of these different apps will run on top of that, but creating that common structured foundation that everyone relates to actually really helps the acceleration.

Dan Jeavons:
And I think what’s also been interesting is as we’ve started to do this, it’s built its own momentum because people suddenly realize because we’ve had a lot of customers for it, investing in the common asset means the asset gets better and better. And we’ve now reached I think something ridiculous, like 1.7 trillion rows of data aggregated into this centralized environment and it just keeps going and it’s growing all the time and new customers are coming every week because they see the benefit of it. So rather than a push up the hill, which it was initially, we now see the flywheel effect starting to happen.

Alex Mysak:
And I’ve got another slightly question for you just given your passion in this area, where do you see Shell and data and AI going within the energy and your equivalent to retail space within the next five to 10 years? What do you think is going to be possible?

Dan Jeavons:
It’s a great question. And maybe I’ll sort of bring it back a level to the macropicture. So I think if you look at society, there are two big things happening simultaneously. So one is digital transformation, which we’ve been talking about so far, which is the acceleration of data and AI and associated technologies in general. I would argue also accelerated through, through some of the recent challenges and so on. But I think on the flip side, we also see a big need to focus on energy transition. And so that energy transition narrative is also really important and I think will be a growing narrative within society.

Dan Jeavons:
And for me, if you ask, “Where is Shell going?” I sit at the nexus of those two things and right in between them and try to hold those two things in tension and harness them to Shell’s benefit. And I think actually I find that extremely exciting because I think if you look at what data and AI has the potential to do for our company, it has the potential to make our existing business much more effective and efficient by allowing us to explore things we couldn’t explore before. But it also allows us to create new business models which are more appropriate to the energy transition, which allows us to move towards lower carbon sources of energy. And I think, again, this is what gets me so passionate is the ability to do both of those two things, because both are really important in transforming the energy system.

Alex Mysak:
And then I think as we close out, what would be the one piece of advice that has set you well in your career that people could pick who in potentially action here?

Dan Jeavons:
Well, it’s interesting. Maybe I’ll go back to the Lord Nelson example at the start. I think a lot of it is about perseverance in the face of adversity. I think change is hard work, I think what you have to recognize is for many people who are trying to embed digital and AI into large companies, you’re running against the grain and you’re trying to, if you like, I use the analogy, push the rock up hill and that’s hard work. And I think you’ve got to be able to take a few punches and to recognize that you’re not always going to be thanked for what you’re doing and you and the organization has a natural resistance to it because you’re trying to bring transformation to the ways of working and to business processes to drive value.

Dan Jeavons:
And so I think one of the things that I’ve learned is that perseverance aspect is really important. And we had some really tough times. I joke about it now, but some of these failed projects, when you get a big one wrong, it really hurts. And so I think it’s that sense of I understand that this is really impactful, it’s going to make an impact, I know it’s the right thing for the company and I’m just going to keep going.