AI to the Rescue: Battling Wildfires – Emrah Gultekin, Chooch.ai
Since the ever-expanding fire season is upon us, California utilities – and fire services – are swamped with as many as 2,000 false positives a week from existing wildfire detection systems. False alerts come from fog, rain, dust, bird poop and simple smudges on the deployed camera lenses. This is where generative AI comes in. Chooch’s generative AI tool gives firefighters in California’s Kern County a dashboard on their smartphones and PCs, populated in real time with alerts, so they can detect wildfires fast.
Transcript
This is Techstrong tv. Welcome to Techstrong tv. I'm Bonnie Schneider.
Today we are diving into the crucial intersection of artificial intelligence and emergency response, particularly in the context of increasing wildfires due to climate change. Joining me today is Emra Gokin, who is the co-founder and c e o of Chooch ai, a company at the forefront of the generative AI technology. Chooch is revolutionizing wildfire detection and potentially saving countless lives with its AI vision platform.
We are Emera. We're so happy to have you here with us. Thanks, Bonnie.
It's, it's great to be here. Thank you for doing this. Sure.
Can you share some, uh, details about your background and how you got to, uh, choose how you established this company? Yeah, It's, it's, it's been an amazing journey. So I, I was trained in the social sciences, uh, but had to learn, had to earn a living.
So I became an entrepreneur, uh, pretty quickly outta college. And I always kind of been at the intersection between, uh, grappling with realities with entrepreneurship, uh, living life as a regular person as all of us do, and trying to understand human evolution at the same time. So, uh, I've built several companies in the past and, uh, this one, however, chu, uh, is by far the most interesting and, uh, and challenging at the same time.
Uh, so it all started out back in, uh, 20 13, 20 14. And my co-founder, based on the research that he was doing, he was a basic practical applications of ai. Uh, my co-founder Han, he developed in the past, we wanted to see if there was a path, uh, to copy human visual cognition into machines.
Uh, so that was the basic thing. Can we copy human visual cognition into machines? And is that something that we can be part of, can be part of that, that stack?
And can we, can we do, do that internally as well as a, as a technology company? So intuitively we knew that that was gonna happen. It was, it was just ultimately gonna happen, but what we were looking at was where in that stack we would be.
And it's really ultimately about the journey of human evolution. Um, and that's one what I want to, I want to always come back to that because, you know, we're talking about ai, we talk about machines, but that we're actually talking about ourselves and how we're evolving as humans as well, like how we're perceive things, how we interact with ourselves, with other fellow humans, and also our surroundings. Um, so it's not ultimately about the machine's cognition, it's about our journey as well.
And that's why we, we, we started Chu was to be able to, to create that type of, uh, cognition and machines and copy it. How would you describe ChaCha's generative AI tool? And can you explain how it aids firefighters, uh, specifically with the example of in California's current county, um, in real time wildfire detection?
Yeah, so we're, we're currently doing processing about 1,020 streams simultaneously in, in, in from ground station in Kern County and all California in general. Uh, that is the extent of which the alert California has public cameras on the field. So, uh, the more cameras, obviously we can do more detections, but it is, you know, we're kind of limited by the amount of, uh, cameras that are out there.
So our foundational model, which is called Image Chat, that's which is generative, is able to recognize minute details and images and video streams, so such as smoke and fires and everything, and, and can abstract things as well. So wildfire smoke, uh, particularly is, is, is an interesting case because it can easily be confused with, uh, mist, fog, clouds, marine layers, uh, any type of smoke. Uh, so utilizing basic object detection, uh, to, to be able to detect wildfires, it just wasn't cutting the job.
And so, uh, the introduction of image chat became very, very important in that, in that journey because it's, it's a generative AI tool, uh, which, which is a combination of computer vision, ai, um, models l and LLMs, uh, juxtapose together. And, and it's, it's almost like, uh, the human sees with, with the eyes, but we think in language and the machine does the same thing. The cameras see whether its eyes, it does some basic object detection, but it really understands its context and what it's looking at only after adding the l l m with the computer vision AI models.
So that's it, it, it, it's a, it's a huge jump in, in, in computer vision AI that we haven't seen in about eight, nine years since the first, uh, deep learning frameworks came out. That's really fascinating. My background is in meteorology, and I know that, um, you can get false false positives even on radar, um, that something that could look like rain, but it's actually smoke or, or it could be dust or it could be something else.
So how does your, your platform handle these false positives by fog, rain, dust, or other disruptions? Yeah, so that, that, that's exactly the case, Bonnie, you know, it's, you get a lot of false positives. And before we deployed, um, the generative, um, model image chat, we were getting about 2000 false positives a week.
And across a thousand cameras, that may not seem like a lot, but in the, in the, in, in operationally, it becomes very, very difficult for firefighters to kind of look through all those images and, and try to, so this is something that sits on top of, uh, the regular object detectors and is able to, uh, um, is able to, uh, understand the context of what's happening in, in high detail. And that's where the merger of object detection, and I'm again, may, maybe I'm getting a little technical here, but it's important for I think, the audience to understand the differences between object detection, image image classification, LLMs, uh, the merger of these has become the important factor in reducing the false positives. Uh, so you really need ensemble models, uh, from the get go, uh, to be able to, uh, to, to, to decrease, uh, the false positives.
And also we receive a lot of feedback, uh, from the, from the, uh, from the field and, uh, the users on the field. And we constantly augment the, the, the data set, uh, with new information that's coming from the field. And that increases the model accuracy over time as well.
So it's a very iterative process. It's called reinforcement learning with human feedback. And that also keeps that, you know, it increases the accuracy of the models as we move forward.
So the, it's, it's a combination of the ensemble models, uh, that you create and also the feedback that you're getting from humans on the field. That makes sense. Um, can you provide more insight into how AI vision works at the nanosecond level to detect process and analyze visual elements in video streams?
Our, our audience is very technical, so feel free to, uh, take a deep dive into that. Alright, awesome. Yeah.
Uh, this is, this is the, this is the core of what we do. And, um, you know, we call, we, we, so where, when, when an AI makes a decision, uh, that, uh, that mechanism is called the inference engine, it's actually a, it's all, it's actually a very complex inference engine that sits, um, where the data is being generated or where it's being stored. And our inference engine generates predictions on video streams every six milliseconds.
Uh, so imagine you have thousands of streams and you have, you have raw data, which is the raw camera feed data going into the inference engine, and then it's being structured and then being sent back every six milliseconds. Uh, that's practically real time. Uh, and that's something that has, we've developed with our partners over the years, and it's also, it's on the edge and it can be self-hosted as well.
Uh, and this is probably one of the most difficult parts of creating a robust, um, infrastructure, is orchestrating the models, uh, the streams across devices. And that works seamlessly with the inference engine as well. And so these predictions detections that are raw detections that come in, so let's say you have the inference engine, uh, being fed all this, uh, all this data from video streams, it structures it to a certain degree on the, on the inference engine, and it sends out the six, the six millisecond, um, uh, inferences.
Those are detections that are happening, but they, they're also being fed into, uh, into analytics engine at the same time because it's impossible for a user to understand what's happening with six milliseconds, you know, types of detections. It's just too much information. So that goes into an analytics engine, and then that, that analytics engine, uh, creates consumable data for, for operators and humans to look at, like, uh, pie charts, uh, you know, detected, uh, smoke events, um, you know, distribution charts, alerts and so forth.
Uh, so that, that's the last piece of this. So if you look at it from a lifecycle, uh, point of view, you have the dataset generation part initially, uh, then the model training and the model ensemble, then you have the inference engine, and then you go into the analytics. So there are four parts to this that have to work in synchrony with each other for, for the solution to work for anybody.
That makes sense. Uh, with fire season expanding, I'm sure you're well aware of that in, in California, um, but it's obviously happening in ca in Canada, fire season is expanding due to climate change. So how does chu AI plan to evolve its technologies to meet this challenge?
So yeah, we definitely want to extend the offering. Uh, we're currently in California and also we're working with our partners in Canada, um, as well. Uh, we have recently deploying in Alabama, uh, too, and we're also working with our partners, uh, for more overhead coverage.
Uh, so we're working, uh, in initially with the ground stations, uh, but we're also working, uh, with partners on the drones, uh, geospatial, uh, and also balloons. Uh, and the third part would be satellites. The issue with satellites is, uh, because what we're trying to do is supposed to be real time.
The satellites cannot give us that real time information yet. Uh, so the drones and the, uh, and the overhead balloons, uh, make, make more sense. So we're trying to expand that coverage, not only in the US but also in other countries as well, uh, where, where, where, uh, we have these, uh, these problems due to climate change and, uh, and, and extreme events.
Yeah. Hmm. That makes sense.
Uh, do you have any examples of, um, success stories of how you've been able to improve this, uh, response time? Yeah, so, uh, current county and San Bernardino County, they're currently using our systems. Uh, and, uh, we beat the, uh, nine one one call on many occasions.
Mm-hmm. And that's, that's, that's critical for this, right? And, and they're able to review those, uh, images as they come in.
And so there's, there's, uh, you know, there's, there's, there's something going on where we're able to provide those, uh, those insights to, to, to people on the ground, firefighters on the ground immediately, uh, so that they can send the right type of response to. So we've had issues where the nine one one call may come in, but they don't know how big the, the fire is. But from our systems, they can actually, it, it, it, it, it provides the size of the fire as well, the smoke coming out, so they're able to send the right type of vehicle and, and resources to, to that, uh, to that location as well.
So beating the nine one one call is one, but also sending the right resources because, you know, the exact size of the, uh, of the fire is really important too. Yeah. And that's a really good description and example.
I mean, to beat the nine one one call and, and have them deploy the appropriate resources is really an, an amazing achievement. Um, I'm just wondering, so what, what else are you looking at going forward and, and ways to use AI to improve response times? Obviously there's a lot of different types of climate induced extreme weather events.
Um, could you see this technology being used and, and, and other, um, areas because, um, boy, it just seems so valuable. Yeah, so fire is, fire is a, fire is an important aspect of this. So we all live in California, and, uh, we started this journey in 2018 when the first, you know, the plume fires happened, uh, in Chino.
And we're like, we have to do something about this with the, with the technology that we have as, as vi as vision AI practitioners. So, but in time, what happened was, um, it's, it's not just about fires, it's about any type of, um, uh, emergency that might happen. It might be in smart cities, it might be in, in, in rural areas, agriculture.
So imit Chat itself is such an incredible model. Uh, you're able to create new bottles through single prompts. So yes, we're looking at floods, spills, flares, accidents, pattern of life changes, and, and many more things as we, as we move along.
So, uh, and, and you don't have to collect a lot of data to do that anymore. All you need to do is just create a prompt, uh, obviously with an ensemble model, and you're able to get that, get that, get those insights as well. Um, so it's, it's a really exciting time, uh, for all of us, I think, in, in terms of, uh, putting these ai, uh, based, um, solutions into practice.
And we we're able to do that and scale it as well. And, and, and the ultimate goal here is obviously, you know, saving lives, uh, saving property, uh, making communities safer, uh, you know, for ourselves, for our children, uh, for our friends and communities as, as at large as well. So, um, it's, it's, it's really an exciting time.
I think, you know, in, in terms of what we do. We're, we're in the business of making things more transparent, right? And, and that's kind of where we're leading with this.
And AI vision AI can solve a lot of our, a lot of our, uh, a lot of our emergency responses. That's really great. Um, and it's exciting to see what you're gonna accomplish in, in the years to come.
Uh, I really appreciate you joining us Emra, uh, gut Klin, uh, co-founder and c e o of Chu ai. Thank you so much for joining us on Techstrong tv, uh, really enlightening conversation. Thank you, Bonnie.
Thank you for having me. Wonderful. Alright, well stay with us.
We're gonna have a lot more on Textron tv coming up.
