EP 61 - Is Now the First Time We Can Map Flood Risk at Scale? - Dr. Andrew Smith - Chief Operations Officer at Fathom

Speaker A: Hi, this is Michael Waitz. And welcome back to Insurtech Amplified. We are joined today by Dr. Andrew Smith, a co founder and the COO at Fathom. Andrew, thank you so much for coming to the show. I really appreciate it. Before we jump in, let's give the audience a little bit of your background just for some context, please.

Speaker B: Sure. Um, pleasure to be here, Michael. Thank you, uh, so much for having me. Um, yeah, I mean, if you go back, uh, long enough, uh, I guess I could have called myself a scientist. So did a PhD in, uh, this quite niche thing at the time called computational flood modeling. So trying to build computer models that effectively simulate floods where they happen and what the impacts are. Had, um, a pretty brief career actually as an academic because whilst I was undertaking my PhD, um, I was fortunate enough to sit next to another guy doing his PhD and his PhD was sponsored by insurance companies. And um, we kind of realized that, um, the things that we were building were actually quite useful for insurance companies. So actually, uh, formed the company whilst I was still undertaking my PhD, um, and kind of did both things for a while and then transitioned really from an academic into, I guess a co founder of a company, uh, immediately. And um, since then I've been, uh, building and helping to run a company called Fathom, which I, um, think we can say is the best flood modeling company in the world.

Speaker A: Cannot wait to talk about it. Was it always kind of a goal of yours to be an entrepreneur? Do you know what I mean? Or were you considering just being an academic for the rest of your life?

Speaker B: Um, I would love to say that I kind of had the vision early on to be a tech entrepreneur. The reality is I always just did stuff I was interested in. That was it. Um, I kind of had a real lack of attention for anything that I found, uh, uninteresting. And I just did stuff that I found interesting. And, um, fortuitously it led me down this path. I just kept doing things that I thought were kind of cool and interesting. And um, it eventually led here, but, uh, it's not really by design, unfortunately.

Speaker A: Well, the word interesting is part of my next question, and it's a beautiful segue. Do you think that there's ever been a more interesting or even important time to understand the risk of floods and frankly, all natural disasters than today?

Speaker B: Well, the answer, unsurprisingly, um, is no. Um, but the reasons for that are a bit more nuanced than people listening might realize. Um, so the thing that everybody will jump to is climate change, a warmer climate amplifies lots of hazards. And that is true. For that reason. There's this phrase non, uh, stationarity, which effectively means things are changing, uh, and the climate is changing and it's making a bunch of natural hazards worse. That's a reason why it's never been more important. But also actually, and the science tells us, perhaps even more importantly is that, um, exposure is going up, and that simply means there's more stuff in harm's way. Um, and that's actually, we think, going to be a bigger driver of risk in the future even than climate change is doing things like building in areas that we shouldn't be building. So if you take that into consideration, then I would argue, yes, having tools at our disposal to be able to identify where is risky and where is not. There's never been a more important time, um, for that reason.

Speaker A: So this is, this gets actually even more interesting now. How do you know and measure, like, where, like, how does the predictive model of where floods are potentially going to happen dovetail with locations that actually have the ability or are, uh, more conducive to having that thing happen? And then how do you control whether things get built there, what gets built there, and frankly, how they get built? And let me give you a little bit more detail. I was not here when this happened, but there was an earthquake in Bangkok a week and a half ago or two weeks ago, I can't remember anymore. I wasn't here for it. But there's not supposed to be an earthquake here. We're not on a fault line, so no one's really that concerned about it. Which means that the building codes, which are still really good, aren't like they are in Tokyo, where we're sitting on a fault line. And the tests they do for those buildings are insane. So even if there's a massive earthquake, which there was back in 2011, nothing really fell down. Right? So how do you control those things in the flood space, which is not exactly like an earthquake, but similar enough, I guess, in some level?

Speaker B: Uh, the answer is with great difficulty, um, if you're building models of these things, um, what you're aiming for in a model is to have something that's, um, as close to a physical representation of the thing that could happen that you can get. Um, and the reason for that is that you mentioned things happening in the past. Well, unfortunately, with extreme events, the past is a really bad, um, predictor of what will happen in the future, because extreme events, by their very nature, are rare events. We actually haven't observed that many of them. So if you're only using, um, if you're only using history as something to go by, then you're not going to get a very good understanding of risk. With flood modeling, what we try to do is we try to effectively build physical representations of the peril, I guess, in a computer simulation. So you're simulating all different variations of things that could happen. But to do that with flood modeling in particular is extremely difficult. Um, I have this phrase that, um, people at FATHOM now make fun of me for saying because I say it so much, but it's true. And I blame my PhD supervisor. All flooding is local. Uh, so you have to represent things at really fine scales. Right? That's something that with flood modeling is really, really difficult to do. So, um, to represent all of that complexity in a model is really hard. And that's really the heart of what FATHOM does. But back to your question. What we're trying to do is to build physical representations of, I guess, reality and then run simulations to, to give us an idea of things that haven't happened yet but could happen in the future. And then you use those things to make decisions with that could be insurance companies using it to ensure they stay solvent. Right. During extreme events. Or it could be engineers and even governments using these models to understand. Well, to do I need to have certain, uh, building code in this area or not? Um, all of those things you can do with these models.

Speaker A: Can I ask you this though? If it's so hard to predict, right, because again, just the amount of times that you've had these massive occurrences are small. How then do you know where not to build? Do you know what I mean? So again, getting back to the earthquake thing. You're on a fault line. That fault line is not going away. It's probably just going to get worse over time. So you know that if you, even if you are going to build, you have to build in a very specific way. Um, but how do you. If you don't have enough sort of predictive abilities, then how do you know where not to build and what not to build?

Speaker B: I can't really speak for earthquake modeling, although, um, earthquake modelers out there will curse me when I say that I think earthquake modeling is much easier than flood modeling.

Speaker A: Way, um, easier.

Speaker B: Way easier.

Speaker A: Sorry, I'm on your side on this because it's there. Like you know it's going to happen. You just don't know when, but you know it's going to happen. Sorry, go ahead.

Speaker B: Um, yeah, I agree. And Also, the event footprints are generally just quite large, uh, footprints. Anyway, um, back to floods. I think we're now reaching the point actually where, um, we do have predictive skill with the models. The thing that's happened in flood modeling in the last 10 years in particular, and I don't use this phrase lightly, there's been a real revolution in our ability to build these models. Going from 10 years ago where there were effectively no models for most of the planet. We had no idea, really, um, to now where we have models simulating risk for the whole planet everywhere, um, from the smallest rivers to the biggest rivers. And where we test these models, or where we can test them and we can test them against things like local scale models built by engineers, also real world events. If an event happens, well, did we capture it? Like, how well did the models do? We do all those things. And what we're now seeing is that there's real predictive skill in the models anywhere. So we can use these models, I would argue, to make decisions around things like planning for anywhere in the world.

Speaker A: So what changed in the last 10 years? In other words, if there was none of this 10 years ago, like literally zero, and now you can model any small creek all the way up to the biggest ocean, what changed? It can't just be one thing, but what are the series of things that had to change or that did change that makes this possible?

Speaker B: Now, I would probably break it down into three things. Um, two of them, you probably won't be surprised when I say this. One of them is data quality. We just have much, much better data available now. Things like better satellite observations, better terrain data sets collected by principally by organizations like the European Space Agency and NASA. Um, compute power has been a big thing. So we're able to run these models now at finer resolutions with complex physics because, um, compute power has gone up. But the third thing, and perhaps most importantly, is just how we conceptualize a whole bunch of the things you have to simulate in models. I'll give you one example, right? Um, if you're trying to build a model of, say, riverine flood risk around the world, something that you need to do is to try and estimate how big the river channels are. So you need to try and simulate, well, how big are these channels? How do they convey water, and what happens when they overtop? Well, to do that, you need some quite innovative thinking, actually, around how you actually conceptualize that problem and come up with a solution. And there's been a bunch of those things. So how we conceptualize river channels how we process terrain data sets has been arguably the biggest thing. We have much better terrain data sets now. How do you conceptualize and understand flood defences around the world? All of those things which are really, and I keep using the word, but all of those things are really conceptual modeling problems. We now have, I think, some pretty neat solutions in place for. So when you take the combination of better conceptualization of problems, uh, levies, channels, terrain data sets, um, better raw input data sets and better compute power, all those things have come together to enable us to simulate these models now for uh, the whole planet.

Speaker A: This is so interesting to me. You can see me when you're talking, like my brain is like running through like, I wonder what the math around this looks like, right? I wonder what these simulations look like. And you're right, there are all these new concepts. What makes the modeling that you guys are doing at FATHOM so innovative? Maybe you can share some of these concepts or conceptualizations that you have so I can get a better understanding of what exactly you're doing and why these models are so different.

Speaker B: Um, so I'll start with the base of the models. Because you use the word, uh, the math, um, the math. Actually I, I often say that Fathom builds physically based models. That's really the USP of the whole company. Like we're building models that literally simulate the physics of this thing happening. So we simulate like flood wave propagation coming down river channels. The actual math, uh, is really old though. So at the heart what we're solving here are some really, really old, um, mathematics. So they go all the way back to the 1850s actually. Um, so how we actually simulate waves of water propagating across the earth's surface. The actual mathematics is pretty old. But how you actually, and again, how you actually use it, how you actually solve those equations in an efficient way, is the kind of unique thing about fathom. But the big USP again is we basically build physically based models. And um, the fact that we're simulating, um, floods happening in a computer means that the outputs are also useful for a whole bunch of different people. This, this is not a, uh, an industry specific solution. We actually sell these data to a whole bunch of different organizations which we might come on to. But, um, the reason we can do that is because the models themselves are just a physical representation of flooding.

Speaker A: So did you have to build all of this software from scratch? Like you couldn't just go down to, you know, software house and buy like flood modeling software. It just wasn't there. Right? I mean, this is the USP of the stuff that you're building. But like, how long did that take you guys to build that? Because it sounds like it's very sophisticated stuff. And again, because it's all local. I want to get back to this. Right. Because all flooding is local. It's just an interesting concept. It's been running around in my head since you said it a few minutes ago. Are you modeling, like real world places like you're going to Mississippi and then modeling the Mississippi river and what happens when that thing floods? Because you know the topography, you know the depth of the water, you know the speed at which the water is running, you know what the weather's. I'm just making stuff up. Right. But you know what the weather's going to be like there. And then you put it into a simulator and think what's going to happen if these other variables change. And then what's the impact of the flooding around that area? And you're building that all in software on some probably super powerful computers. Is that what you mean when you talk about local, that you're doing it in all those individual places? Or are you just simulating what something like the Mississippi river could look like?

Speaker B: You know? Yeah, Michael, you probably just, could. You probably just articulated much better than I will, um, how the models work. It's all those things. Yeah, we're literally. If you, if you try and think about. And I actually, I hate this phrase for some reason, but there's this phrase bandit about a digital twin. Um, what we're effectively doing is.

Speaker A: We are trying to use it, but go ahead.

Speaker B: Yeah, for good reason. Um, I need to come up with a better phrase. Um, yeah, it's. The model effectively takes places like the Mississippi. We try to estimate how big the actual channel is. We try to estimate how much water flows down it during extreme event. We try to get the best representation of what the land looks like around the Mississippi. And then we simulate wave propagation down the channel and simulate what happens when that wave disperses across the land surface and, um, simulate what risk looks like. So that's really how the models work. To your question, around the code base. Yeah, this is all built from scratch. Um, and, um, I would love to say that we built all of the code. We did build a bunch of it from scratch, actually. Um, but we kind of, um, again, an overused phrase standing on the shoulders of giants.

Speaker A: Um, yeah, for sure.

Speaker B: We joined a research group that had the very first version of this model already built for, um, academic purposes. And that's what we Used, um, really to build the first parts of our model. The unique thing about Fathom, though, the actual spreading, if you think about spreading water, um, across a grid, right? That's how the models work. At their heart, they're simulating how water moves across a surface. That part was, was, uh, kind of done, right? And we use that model and built our own version of it. But effectively how you do it was something that was already done. The really unique thing that Fathom does is processing all the input data sets for that model. So how do you, how do you represent river channels? How do you process terrain data sets? How do you define extreme rainfall, extreme river flows? All of that input data processing is something that we did build literally from scratch. And the important thing here is how do you build a model that's kind of automated? Because if you want to build a model of the whole planet, you can't do that manually, right? You can't go into every single river and kind of tweak things. So it has to be done in an automated way. That's what we did early on. And so the early code base was literally two ah. Of us really building that code base from scratch. And it's kind of, uh, if you want the hard numbers, kind of hundreds of thousands of lines of code. Um, but that's what we use to build our first model. And that's the first thing that, um, thankfully some insurance companies bought from us. But that's where it all started, really.

Speaker A: I want to get to that towards the end of this conversation as well. But this other thought that I'm having right now, and I don't know if I'm going to be able to be able to articulate it properly, right? But the Earth is this living mechanism, right? It's not static, it's highly dynamic. And the stuff that's happening, like I said, like in Indonesia at some level could impact what's happening in California. But because you have access to all this, you talked about getting data from satellites, getting data from NASA, like all this kind of stuff, all that's happening in real time as well. So do the models actually work in real time and can they help predict things that, uh, may happen? And the first thing, I actually wrote down a little note to myself for surfing, right? And I know surfing is not related to flooding, but it's related to water movement, right? And since I'm a big fan of watching surfing and big wave surfing, it just made me think like, okay, these guys always say, okay, there's going to be this massive swell on the north side of Hawaii. How do they know? Right. And then I thought about all this modeling stuff that you're, that you're building. Um, but do you do this in real time? I guess that was the real genesis of this question is the Earth is moving. It's always moving, it's always giving you data. Can you just be like, oh, something's going to happen in Minnesota kind of thing?

Speaker B: Well, Michael, you're going to take this conversation in a different route in a minute because I also love surfing. I could talk to you for a long time about big wave surfing. So this makes it into a surfing podcast. Um, the answer, in short, is no, these models are not run in real time. Uh, and the reason for that is that if we're, if we're trying to simulate flooding for the whole planet, just one simulation of the whole planet takes about three months on our servers. So they're updated kind of, uh, every 18 months or so we run the whole planet again. Um, but when we rerun the planet, all the things that have been observed in the past 18 months can be plugged into that model to kind of help us understand the extremes a bit more. Because as I said at the start, extreme events are very rare events. So the more observations you have, the more certainty with which you can try and define, uh, what the extremes look like. But they're not really run in kind of a forecast mode per se. They're not kind of constantly running in the background like a numerical weather predictor would be. That's just not how they work right now.

Speaker A: So they're not now. And again, I didn't think about this before we started recording, but does quantum computing, which is not fully there yet, impact the ability to go from this three month cycle to one day cycle? Or am I just misunderstanding the way quantum works?

Speaker B: I think it would, um, I'm not a quantum compute expert. Um, neither am I. If you had unlimited compute, um, then I, uh, guess you could run the models in kind of forecast mode. But the important thing to say here is that the actual use case for these models right now, um, isn't really for forecasting. They're used for understanding risk. Right. So it's not, we're not kind of helping in another direction. Yeah, yeah, like when something will happen. Um, actually forecasting right now, I would say that some of the flood modeling, the real difficulties with, um, with flood forecasting is not really a compute problem on the flood side. It's getting better predictions of the weather. Yeah, that, that's kind of, uh, A real problem right now when it comes to trying to forecast flood risk, like forecasting actual flooding is really, really difficult. And we're doing it a bit right now and we try and do some event response work. But, um, if we had unlimited compute, some people on our team might even disagree with me. But I don't think like unlimited compute right now would solve all of our problems when it comes to flood forecasting.

Speaker A: Yeah, it's such a gigantic, um, there's just so much data out there. Can I ask you this? What was it about what you were building? Because I want to talk about insurance, but then I want to talk in a moment a little bit about where else this can be used. Right. Because I don't think it's just an insurance issue, although that's where a lot of the risk is going to get mitigated, priced and disintermediated. Yeah. What was it? So what was so interesting about this from the insurer standpoint? And why was this something that they hadn't built already?

Speaker B: That's a really great question. So why was it something that's not being built? It was kind of impossible to do 15 years ago. Like trying to build these models at large scales was just not possible owing to, um, the models from, even from an experimental basis not being built yet. Um, a complete lack of data. Um, we had no real great data sets to build these models, principally things like terrain. Data sets didn't exist in a suitable way. So, um, I use this phrase often. All water flows downhill. And if you don't know where downhill is, the models are not going to work very well. Well, for most of the planet we didn't actually know with enough detail where downhill was. So for insurers building these physically based models to understand flood risk, um, hasn't been done until very recently. And indeed I was kind of shocked actually when I began to work with some insurance companies. Like some of the models used to understand flood flooding, believe it or not, not really flood models, they were just all they really were at the heart were statistical models built to replicate past losses. So you're just building models that take in things you've experienced and you kind of extrapolate with them slightly. Right. Um, they're not really physically based models. Um, that was a solution that insurers had in place and indeed some still do today. Um, I would argue, however, that you really need physically based models because, um, there's this concept in modeling that you want the models to be, um, as conceptual as you can. So to represent reality really. As much as you can, and you want to calibrate them as little as possible. And calibration here is effectively kind of, uh, tweaking the knobs to make them align with things you've experienced in the past. Right. That's a kind of a calibration phase. Um, in insurance, in reality, the models were almost entirely a calibration process, and that still kind of exists today. And that's a real problem, because the more calibration that's involved with a model, the less conceptual it is, the worse it will perform when you extrapolate. So once you start to move beyond things you've experienced in the past, the models will quickly fall apart. Um, and we are indeed, as we mentioned at the start, like climate change coupled with, uh, rapid urbanization, we're now moving into a world that doesn't look like the world that existed over the past few decades. So you really need the models to be as conceptual as you can. In insurance, again, that kind of wasn't really the case until recently.

Speaker A: Can we talk a little bit about the terrain modeling, if you don't mind? I find this really interesting as well. Um, and again, the only thing in my head is if there wasn't enough data out there, or if the data wasn't organized well enough or clean enough until recently. Once we have the terrain of the globe mapped, did the world look, and I know this is going to sound like a stupid question, but did the world look like a completely different place than we thought it was? Right. This idea of water flows downhill, and if you don't know where downhill is, you're never going to be able to price this stuff effectively. You're never going to be able to understand where that water is going to go. But now that we know, does the world look like a completely different place than we thought it was? Because we know now all of the terrain, or most of it, at least. Yeah, yeah.

Speaker B: The answer is yes, in many places. It does look very different to some of the early models. Um, I'll give you a quick potted history of terrain modeling because, uh, when we first started to build these models, actually we moved quickly from being flood modelers into terrain data processes, because that was a big, big, big problem. So the early data set that we used is something called srtm, and that's an acronym for the Shuttle Radar Topography Mission. And this is effectively, um, as I understand it, it was kind of done by NASA on a whim for something to do. So they stuck a rad out of the top of the shuttle and then flew it around the planet and collected this kind of radar image of the Earth's surface. And um, it was the first time actually we'd ever had that. The collection of that data set was the reason why we started our company. But. And there's a big but. It was pretty hopeless. It was the only thing we had. But it was pretty hopeless. Um, and it was hopeless for a bunch of reasons. So first of all, it's what you'd call a surface model. And all that means is that the radar beam bounce it off the top of everything that it saw. And that includes trees and buildings. So towns and cities look like hills and forests look like hills. What you need is the surface. So you have to do a whole bunch of processing to get rid of all that stuff with srtm. The other problem was that um, the radar, the spaceship actually kind of fluctuated, um, back and forth slightly in space and the radar boom bounced a bit. So you had all these instrumentation errors as well.

Speaker A: I love it.

Speaker B: But we learned so much from that. We processed that data set for 15 years and in the end we really got the maximum signal out of it that we could. It put us in a brilliant position then to process the next thing. And that was collected by the European Space Agency, uh, a mission called Tandem X. Uh, and they released a much, uh, it's native resolution. So the raw signal was much more accurate than srtm. But we applied a whole bunch of the processing from that, uh, from SRTM onto that to produce the next version of our terrain data sets. Now that leads me on to the where we are today. And where we are today is, uh, really, really exciting because I hate using this phrase because it's become like a panacea for all problems. But this is one of the real instances where uh, machine learning, AI is revolutionizing what we do. Um, because when it comes to terrain data processing, what you're really talking about is image processing. You're trying to remove errors from images. That's what you're doing. And what we're discovering is that actually this new tool that we have is absolutely wonderful for processing terrain data sets. So we, we have a new terrain data set, uh, coming out, um, soon. The academic publication came out literally weeks ago. It's called Fathom Dem and it'll be the world. It is the world's best terrain data set. Um, there's, there's a gold standard terrain data set. And this is a long answer. Apologize, Michael.

Speaker A: Uh, so interesting.

Speaker B: Well, I'm glad you find it interesting because I certainly Do. Um, there's a data set called lidar, which is laser altimetry data. So if you wanted to get like the best, best gold standard image that you could, you'd fly a plane with a laser and it would collect a very precise image, an accurate image of the Earth's surface. Fathom DM M We're now getting pretty close to lidar actually in many places. So the gap from where we were 10 or 15 years ago to where we are now, I honestly can't really believe how far we've come. Um, and it is, in some areas, it is really changing our perception of what risk looks like in some areas because the terrain data set in a flood model is the thing that governs quality above everything else. So in some areas it looks, it does look very different.

Speaker A: I mean, am I just completely misguided here? But doesn't the terrain also determine at some level how the wave flow, how the water flows over that area? Right. Because if it's, if it's built in a specific way, and there's probably terminology for it that I don't know, the way the wave, the way the water is going to move is going to be different than if it's just like a flat surface, if there's nothing there. Right. So if you don't understand the terrain itself, your ability to, not to predict what's going to happen there, but to understand what would happen there if it flooded is like night and day, no?

Speaker B: Yeah, yeah, absolutely. So the models, at their heart, they critically, the models conserve mass. So there's only a certain volume of water that's available in the models. And if you don't have the correct, uh, let's call it, image of the terrain, then that mass could be spread in areas where you don't want it to be spread. Right. And that could, that could alter absolutely how a flood wave propagates downstream. So, um, the terrain data set for sure can, can alter things like flood wave propagation and result in a pretty, uh, erroneous idea of risk if you don't have it correct.

Speaker A: Yeah. Which means all the pricing is going to be wrong, all the risk mitigation is going to be wrong. All the potential claims data is going to like, everything's just going to be wrong because your ability to measure what that risk is like, it's going to be upside down at some level. I don't know. This is so interesting. Now I know why you do this stuff. Um, this is a really esoteric style question, but I'm really curious. I want to get onto where this, where this other stuff, where this can be used in other places besides insurance in a second. You know, I used to say when I was trading the stocks in Japan, right, that every, every company that was listed had a stock code. So like 6758 was Sony, 7203 was Toyota. I can't remember a lot of them. 801-8316 was one of the big banks. I can't remember them all anymore because it's been over a decade. But I used to go running and you'll see where this is going in a second. I used to go running in Tokyo, you know, just for exercise and I'd look up and say like, oh, there's a, uh, Sony thing that's in my head. I would just see 6758 because I was just in the market all the time. And I'm just wondering if you, who you're so deeply embedded in all this terrain data and all this flood data and all this risk data that just when you're driving around or walking around or on vacation, you're just like, oh, that's a problem, you know what I mean? Where you just see things that other people wouldn't see because they don't know the stuff that, you know, it's a really esoteric but you know what I mean, right?

Speaker B: Yeah, absolutely. And the answer would be yes. My um, wife would tell you that I can be really boring at times if we're on holiday. And I, I uh, see what I think are quite interesting things. Um, that, that is true. But I think the reality with a bunch of the stuff that we try and simulate is that it's kind of obvious you don't need to be a real expert sometimes to understand what is risky and what is not. So um, yeah, I do, I do see some, some things and I think, well that's a potential problem. I think the one thing that when you're trying to build these models that simulate really extreme events, the one thing that you do have a perception of is kind of just how bad really extreme events can get. And if you literally have no knowledge of kind of what a kind of waterway could do during an extreme event, then there are sometimes things like pick up like, well, high watermarks or things that have been transported downstream that um, that other people probably wouldn't think about. The other thing that's amazing I think to me is that um, when it comes to extreme events, people have really short term memories. Like you can, you can have an extreme event happen and kind of within a matter of years people kind of forget it's happened and then it happens again and people are really surprised all over again. Um, like humans expected that. Yeah, yeah. Uh, I definitely have, I think, a good appreciation of, um, I guess, uh, extreme, uh, extreme value analysis and kind of understanding kind of bad things happen. And they can often happen more frequently than you might imagine. Um, yeah, yeah.

Speaker A: So talk to me a little bit about where this other, where this stuff can be used outside of the insurance industry. It's probably tangential at some level, but I'm just curious, like where else is this stuff being used or can it be used?

Speaker B: So when we started, so my co founder is a guy called Chris. And um, Chris's PhD was sponsored by a big reinsurance broker. So we had this kind of inside knowledge and we'd go down and talk to insurers in London. So when we started to build these models, it really was focused on, uh, insurance companies. Um, but actually the technology itself, going back to my earlier points, they're really just kind of physical representations of flood risk. So the reality is that anybody concerned with flood risk, anybody concerned with climate risk can use these models and these data. Uh, and pretty early on we had, um, clients start to use these data beyond insurance. Uh, my classic example is Microsoft. Uh, got in touch very early on. Uh, I thought it was spam, actually, I thought it was a spam email because the company, and uh, I use inverted, uh, quotation marks on this. The company was myself and Chris in a room and we had an email from Microsoft saying, uh, we'd like to talk about your flood models. Uh, so I didn't think it was real. It was real. And they're still a client today and they use the models to map risk across all their data centers. So that was the real first inclination we had as well. The market here is much bigger than this fathom now serves. Again, anybody concerned with climate risk, that's insurance companies, it's financial markets, banks, some of the big banks in the world. The biggest banks in the world use our stuff. Asset managers, people like BlackRock use all of our stuff. Um, engineers. We have a whole bunch of the world's biggest engineering companies now using our data and our models and then governments as well. We build models now for the, for governments around the world. Um, the Australian government uses our data sets. We've built the flood maps for all of Texas. We're now embarking on a partnership with AON to build flood maps for all of Canada. So lots of governments and corporates use our data sets. I think maybe in some ways most rewardingly, people like the World bank use all of our data sets in really data poor parts of the world. And we're actually able to give that data, uh, in some areas away for free. So they have countries that they determine to be fragility, conflict and violence afflicted. These are some of the really kind of some of the most difficult places to live in the world right now. All of Fathom's data is available for free in those locations. And um, you mentioned the earthquake. Um, Fathoms data are being given away for free to aid, uh, agencies working in Myanmar at the moment. And that's being used to help them decide where to build things like, uh, refugee camps and aid stations so they're not in exposed parts of the world. So, um, that's a whistle stop tour. But again, anybody considering the flood risk can use the stuff that we have.

Speaker A: It's got to be super cool to know that like, like you said, two guys in a room. I mean, I'm sure there's way more than two guys now.

Speaker B: Oh yeah.

Speaker A: That kind of can have that kind of impact in places where it really matters, right? I mean, Thailand. So Bangkok is literally sitting at sea level. And you may Remember back in 2011, the flooding here was just insane. I mean it was everywhere. And when it rains here in Bangkok, I don't know what their flood controls are. Sometimes some days they're better than others. But like, I've literally walked down the street with the water up to my knees.

Speaker B: Michael, this, this people listening will think you planted this point because 2011 in Thailand was the whole reason we started our company. It was that event, actually. So that event happened. And yeah, uh, it was that event. And, uh, it was our knowledge of insurance companies. And they were saying to us at the time when we went to London, this is an unmodeled event. Which, uh, effectively meant we didn't see it coming. Like, nobody saw it coming. It cost nearly $20 billion in insured losses, not economic losses, in direct insured losses, still the biggest flood loss ever. Um, and we could see back then that our toy models, we'll call them that, and we're picking up these exposures. Um, there were some particular industrial estates outside of Bangkok that we could see were horrendously exposed and yet were insured with no knowledge of flood risk. So that was really the thing that set us off, um, and was the reason we started the company. Uh, but I will say, Michael, you mentioned I've said two people in a room. I built some of the early versions of the models. Our team is now so much better than I ever was at building these models. The company is now 60 people, and we have just. I don't want to sound too cheesy, but it's a real privilege, a true privilege to work with them because unbelievably innovative, smart, creative people building these models now, and they're far better models than I was ever able to build myself. So, um, yeah, absolutely. A, uh, much, much bigger team than, uh, a few of us in a

Speaker A: room now, do you think. And again, this is a really esoteric question as well, but do you think that great leadership does take a certain amount of humility? This idea of I'm amazing, but maybe that lady's more amaz than I am at this thing?

Speaker B: I think so, yeah. I mean, I think, um, honesty as well. Right. Knowing what you're good at and what you're not. I think, um, myself, Chris, and a few of our. Our early team were just really good at acknowledging and identifying what each of us was good at and then kind of being able to focus on that. Um, Yeah, I don't know if, uh, you're kind of a. Frankly, a bit of an egomaniac. I don't know how you really. And I know some people do succeed in this world being that way, but I don't know how we would have built what we' built being like that. Because the problem is so big, um, and the things you have to solve are so, um, varied and wide that you have to kind of, um, divide and conquer. You have to allow people to have autonomy and work on things themselves and not micromanage. Um, there's just no way you could do that with what we've built. So humility for sure, is a big, big part of why we've been successful.

Speaker A: Yeah, I think so, too. I do the same thing in my company. There's some stuff that I'm really good at, but there's a whole bunch of other stuff that I'm really bad at. And my ego is not big enough to say, like, I should be in charge of that thing, because I have a business partner who's way better at that stuff than I am, and he should be in charge of it. And he's half my age, so a lot of people look at it and think, like, oh, that's really weird. And I'm like, no, he's way better at this stuff. Okay, I want to have a little bit of a conversation with you about building a company from scratch and having it acquired and then still working at it. Do you know what I mean? But this sounds different to me and I'll tell you why. Because a lot of times you see guys, and again, I don't, I don't. Haven't spoken to somebody about this. But a lot of times you see people build companies, they exit the company, they make a decent amount of money and they're like, okay, I'll do the thing for a year or whatever, and then maybe I'll go do something else. Because I know how to build companies now, right? But it sounds to me. And, um, I'm. I don't even know when your company was acquired, right? So I don't know what the, the plans are, but it sounds like this is something. And you said this earlier when we were talking and I know this feeling, right? It's just like this is something you really care about. Like, you didn't just. Someone didn't assign you to your PhD thesis, right? This is something you cared about. You built it and then somebody happened to buy it. So I think your situation's slightly different. But maybe you can run me through like, some of the feelings around this when it was acquired, like why and then how you felt it, like all these little things. If you don't mind, because I'm super curious.

Speaker B: Yeah, sure. So the company was acquired at the end of 2023. Uh, we were acquired by Swiss Re, um, one of the biggest reinsurance companies in the world. Our aim in the end was we wanted to find a good home for the company, um, because we built this amazing team of people who I, who I love, just love working with, um, and I wanted to find, we wanted to find a good, a good home for everybody. It kind of, uh, got to the point where it didn't make sense for mainly two guys to be owning this company that all these people were a part of. So we wanted to find a good home for the company. We also knew that to really maximize the value of this technology that we've built, you'd get maximum value out of it by working with an organization like Swiss Re. Um, the reason for that is that kind of in an obvious way they can use the technologies to understand risk and exposure. But also I mentioned earlier, like calibration, so how you understand losses. So you go from physically based models to models that simulate losses. To do that, you need access to, um, information about losses that have been experienced in the past. So that was also a really, really big part of, of why we did it. Swiss RE is also a real science led organization. They have this thing called the Swiss RE Institute. Fathom is like, I say we're like a mini university. Um, so frankly, when, when, when that approach happened, um, it kind of felt like, okay, uh, this is kind of the right thing to do. The reason I work there now is you kind of hinted it. Like, I just, I really, I find this stuff really interesting and I still feel like there's a bunch of. The journey's not complete, right. The technology is not fully matured yet. We still haven't reached the pinnacle of what we can build. And it's really exciting right now being a part of this organization. Um, we want to build a new innovation center. So, um, yeah, we're still on the journey in my eyes. I might not own all the shares anymore, but I'm still on this journey. And again, I just love working with everybody. It's such an interesting, innovative, kind of exciting group of people that um, I can't imagine doing anything else right now.

Speaker A: Yeah, that's an awesome way to end. Um, Andrew, you've been amazing. I've loved this conversation. And this stuff is not boring if somebody tells you that it is. Yeah, I don't know, I find this really fascinating. I could have gone on, but I want to keep this to a certain length. I really appreciate your time. I really appreciate all the efforts you put, put into this. Um, and let's have this not be the last time you come on the show. I would say this to you. If new things come up, if new things get built, if some interesting things happen, frankly, if a big extreme event happens and you want to come on the show and talk about it and see what the results of it were, I'd love to learn more about the stuff that you're building and how it's advancing. I really appreciate your time. Thank you so much.

Speaker B: Awesome. Well, thank you for having me. And uh, be careful what you wish for because uh, as my friends will attest, I can about talk. Talk a lot. So, um, yeah, uh, we'll keep in touch and if uh, if interesting things happen in the world, absolutely happy to come on and chat through them. But again, thank you, uh, very much for having me.

Speaker A: It's my pleasure.