Bellwether presents to Canada 2020

On October 23, 2025, Bellwether’s Senior Program Manager Joshua Jeffrey spoke to Canada 2020 in Ottawa. Jeffrey’s presentation is now available:

Remarks to Canada 2020, October 2025

All right. Good morning everyone. My name is Josh. As Reva mentioned, I'm one of the founders and currently head of operations for Bellwether. As was mentioned, we're here because of the work that we've done in Wildfire Tech. So I'll give you a deep dive into that. Show you some of the other work that we've done in this geospatial machine learning world.

And as was mentioned, you're in for a real treat. You're gonna hear from Sonya right afterwards. So while we work on prediction, Sonya and Pano work on detection. So I have slides here we go. Some table setting bellwether is actually a project, a technology and a team at x, the moonshot factory, also known as Google X.

So we are part of the Google ecosystem. The mission of Google X is to invent breakthrough technologies that lead to new companies at the scale and ambition of Google, right? So trying to find the next Google we're probably most famous. For Waymo, which is a self-driving car company it's in a lot of different cities right now in the US [00:01:00] as well as Google Brain, which was one of the first AI projects in the Google ecosystem that really laid the foundation for all of the work that's happening in AI across Google at the moment.

And even though many of the technologies that come out of X seem very domain agnostic, we have self-driving cars, life sciences, balloons, up in the air, all this sort of thing, all of them, including bellwether, actually follow the exact same blueprint, which is very simple. We're just aiming for the middle of this Venn diagram, a huge global scale problem, a completely radical solution that's powered by a breakthrough technology.

So you saw on that first slide, Bellwether's. Moonshot is a prediction engine for the Earth and everything on it. We're talking natural world, human built environment. What does that mean? Practically? Practically what that means is bellwether is building tools that power, preparation, response, and resilience as our planet is changing.

So we're trying to build a platform that answers the following types of [00:02:00] questions. There's thousands of these, millions of these, but these are three that I'm gonna focus on today. Preparing for wildfire, five years ahead of an ignition quantifying financial damage immediately after a natural disaster and designing or simulating mitigations to the landscape before we even pick up a shovel or a chainsaw, let's say.

Okay, so at the core, the technology that we've created is a prediction engine. It's a prediction engine at the intersection of geospatial data. And machine learning. So our team has created a number of interesting new paradigms on how to work with vast quantities of geospatial data. Everything you can think of from overhead, satellite imagery to historical remote sensing data to models about hurricanes and all sorts of things in between to work with it at petabyte scale.

At tremendous speed with ease and grace so that we can point it very flexibly at a lot of different problems. The other thing that happens under the hood here [00:03:00] is that we've created a variety of geospatial agents. These are, you may have heard this phrase recently, in the AI space, but these are essentially components of an artificial intelligence system that can autonomously solve multi-stage complex problems.

So what you're seeing on the right here is just the. The top of an iceberg of the types of data that we have access to Isabel, whether inside of the Google ecosystem. So it's a lot of the more usual suspects of the overhead satellite imagery, like I described, but also other really interesting things like a geolocation tool the shape and direction of roofs on buildings multi-spectral imagery, a wide variety of things that power a lot of what comes out of our engine.

The last thing that I'll mention here too is that geospatial data. Is the fastest growing class of data on the planet with a number of small satellites that are going in the air. Everyone's phone in a way, is a geospatial remote sensing agent that can report back things like [00:04:00] temperature, these sorts of things.

This class of data is growing more rapidly than anything, and we heard from some of the previous speakers really that. Geospatial intelligence really is the foundation of not just natural national security, but safety, economic power, supply chains, all of these types of things. So this is where Bellwether is focused.

So what are some of the things we've made with this engine so far? Start with prediction. So I'm a resident in California of the Santa Cruz Mountains. I live in some beautiful redwoods and I also live a mile from a 2020 burn scar, California. In 2020s you might know massive firestorm across the state.

So this problem is actually. Very personal for me, but the team used that underlying technology. It's been about three years developing a suite of models that try and answer a number of questions around this in a new way. So it answers where will risk be about up to five years into the future. How much risk is there, not just low, medium, high, extreme, like what does that mean?

But actually physically, how much risk from [00:05:00] zero to 100% and what is driving the risk? So we don't want this to be a black box. We want a power action. Why is the risk there? Is it because of the trees? Is it the wind? Is it not enough rain? Across many factors and what we've ended up creating is something that actually computes over 600 layers of data per pixel on a map, and starts to understand over decades and decades of training, looking at fire starts, fire progressions, all of the nuances about wildfire completely from a machine learning perspective.

The sort of technical way to say this is that we have no physical chemical explicit. Equations baked into our models. We don't say, if a fire comes to a hill, it goes up the hill faster. It learns all of this implicitly. And the bonus of doing it this way when you have the ability to work with this much data, is you start to understand the nuances of how the peril itself is changing, right?

So for example, there. Not many tools [00:06:00] predicted that the recent Palisades fire in Los Angeles would've made it as far into the built environment as it ended up, as it ended up doing the accuracy metric here we're incredibly proud of. The model has shown for about 20 years of historical data that we're around 90% accurate at determining high risk pixels that end up showing up inside of a wildfire perimeter in the next five years, as well as pixels that we are certain will never show up inside of a fire perimeter.

So what does this look like? We took some data from the model around Jasper and looked at the predictions between 2025 and 2030. And there's a lot to look on the map here. You're seeing all of the colors around the landscape. Some things to point out. One is that we are producing absolute probabilities, as I mentioned before, so no more.

Low, medium, high, we can get very specific. In this case, we're looking at a 3.82% chance of risk, which feels very specific, but the most important thing about this is that we can [00:07:00] actually couch that risk with context. So in this case, we're comparing it to. The Western fiery states. This is you think from essentially Colorado West and it shows that Jasper, this specific pin at Jasper is around the 74% mark.

It's actually riskier than about 74% of properties across the American West. So this type of context is incredibly helpful when you're trying to quantify exactly how much 3.82% is. The other thing that I can show you here that's really interesting is that we can dive super deep into what is driving the risk.

So in this case, if we were to go into the model and go maybe let's say a kilometer or so Southwest, we can see that the tree species on that panel are actually driving risk at that pin up a certain percentage, right? These are the types of things that when you have enough geo data and you can work with it at speed, you're able to derive.

And then finally another view of risk. We're gonna look exactly at a specific parcel. We can [00:08:00] start to look at all of the other factors in that list of 600. So here we're looking at wind-driven risk. We can understand that there's actually things pulling the risk down. For example the aspect of the slope behind the home, or the fact that it's somewhat close to a fire station.

One of my, favorite anecdotes outta the system that I think is super interesting is that a lot of the early results that we were pulling were indexing very highly on this layer of data called road porosity. This is like how thick is the road or how gravel is the road, which we thought was some weird bug.

And it turns out when we brought in a bunch of fire experts, they said that makes complete sense because fire trucks can get in. Places further on concrete roads than they can dirt roads, right? So these sorts of small heuristics that are impossible to model at scale are now possible to do when you can bring this type of technology to bear.

So we're very proud of this work. The last thing I'll mention on Wild Far before I move on is some really late breaking work that has come out of [00:09:00] our partners at Google DeepMind. In fact, later this morning, actually in about an hour, there's gonna be a number of new announcements from Google in this geospatial AI space.

And one of them is around something called an embedding field model, which is like a very. Technical term, but it is essentially a new type of geospatial data that like dramatically compresses down all types of information that we have about the planet, satellite imagery, all sorts of things into this thing called a 64 dimensional pixel.

The way that I think about it is it's almost like Google Translate for the planet, and one of the things that our team did with early access to this is we created a new language for the translator, which we called. Tree species in Canada. Turns out when you're predicting wildfire risk, you need to know the tree species.

That has a huge impact on fire. And in the US we have a very different source of data that's rather detailed down to a 60 meter level about tree species. That unfortunately does not exist across all of [00:10:00] Canada, but. With this new model, we were able to infer with about 87% accuracy tree species across the entire country.

This is the first pass. We actually upped this to about 80 different specific species, and then we fed that into the wildfire model, and now the wildfire model is more accurate. The point is, again, this is like the meta theme here. The advancements that are happening in AI right now are incredibly dramatic.

But I think what is not obvious is that the specific subset of advancements in geospatial tech are like far and away happening much faster. And this is really critical, I think, for everyone in the room. Two other quick things from a response perspective, we've used the exact same technology and pointed it at disaster response.

So in this case, we're looking at some of the ice jams that happened up in Fort McMurray, and the same tool can ingest. Overhead aerial imagery and deduce very interesting things down the pipeline. Like for example, how much financial damage was done to that specific parcel. [00:11:00] This is something we have in production right now with the US federal government, specifically the US Air Force and the National Guard.

They deploy fixed winged aircraft after a tornado or a hurricane or something like this, and we can very quickly start to sum up all of the damage that was done in a given town to help unlock things like FEMA relief dollars. And then finally, I'll give you one more example here. This is from a resilience perspective back on wildfire for a moment.

But the same tool is actually being used right now by our team to simulate mitigations. So if we were to find a community somewhere that is at very high, wild, high wildfire risk, we can actually simulate what would happen if we did a canopy thinning effort. What if we did a controlled burn effort?

What if we did something more extreme? We can quantify that risk reduction. So that's my slide. I'm about to hand it over to Sonia. I think the last thing I would say is, it's a real honor to be here. I think right now, at this moment in time X has a long history of creating really interesting new [00:12:00] technologies, getting them out into the world for our partners, including nation states.

And we'd be really excited to figure out how we can get this technology in your hands and inspire you on how to solve problems with this new technology. So thank you very much.