Orchestrate all the Things - Managing disaster and disruption with AI, one tree at a time. Featuring AiDash co-founder and CEO Abhishek Singh

Episode Date: May 26, 2022

It sounds like a contradiction in terms, but disaster and disruption management is a thing. Disaster and disruption is precisely what ensues when catastrophic natural events occur, and unfortunately, the trajectory the world is on seems to be exacerbating the issue. In 2021 alone, the US experienced 15+ weather/climate disaster events with damages exceeding $1 billion. Previously, we have explored various aspects of the ways data science and machine learning intertwine with natural events - from weather prediction, to the impact of climate change on extreme phenomena and measuring the impact of disaster relief. AiDash, however, is aiming at something different: helping utility and energy companies as well as governments and cities manage the impact of natural disasters, including storms and wildfires. We connected with AiDash co-founder and CEO Abhishek Singh to learn more about its mission and approach, as well its newly released Disaster and Disruption Management System (DDMS) Article published on ZDNet

Welcome to the Orchestrate All the Things podcast. I'm George Amadiotis and we'll be connecting the dots together. It sounds like a contradiction in terms, but disaster and disruption management is a thing. Disaster and disruption is precisely what ensues when catastrophic natural events occur and unfortunately the trajectory the world is on seems to be exacerbating the issue. In 2021 alone, the US experienced more than 15 weather or climate disaster events with damages exceeding 1 billion. Previously, we have explored various aspects of the ways data science and machine learning intertwine with natural events,

from weather prediction to the impact of climate change on extreme phenomena and measuring the impact of disaster relief. AI-DAS, however, is aiming at something different, helping utility and energy companies, as well as governments and cities, manage the impact of natural disasters, including storms and wildfires. We've connected with AI-DAS co-founder and CEO Abhishek Singh to learn more about its mission and approach, as well as its newly released Disaster and Disruption Management System. I hope you will enjoy the podcast.

If you like my work, you can follow Link Data Orchestration on Twitter, LinkedIn, and Facebook. I'm Abhishek Singh, co-founder and CEO of AI-Dash. Before starting AI-Dash, I started one of the world's first mobile app development companies in 2005, and then an education tech company in 2011. After the merger of my mobile tech company with a system integrator, which was publicly listed, I moved to the US. And when I moved to the US, I realized that power outage is a problem in the US too, which was surprising because I was born in a small village in India

where having power for 24 hours was a news. But I did see that due to storm there used to be power outages and 2017 was a turning point when suddenly the wildfire became big news. So when I was thinking what to do as my next venture, a unique opportunity came where wildfire, storm, etc. were happening a lot and with increasing frequency. And in parallel, satellite as a technology was maturing. 2018 was a point of inflection for satellite technology to be good enough to solve many of these problems. So we jumped into that incorporating AI-DASH in January 2019.

So that is how AI-DASH came into being. Okay. Well, thank you. Thank you for the introduction. And I was wondering before we get any further, and there's a number of topics to discuss. So the the platform that you have been developing and also the latest release, the latest product, this platform that you're about to announce, I was wondering if you'd like to also share a little bit more about the company itself. So the founding team and where are you based and headcount and any general business metrics that you'd like to share. Yeah,

that's great. So AI Dash started in corporate in January 2019, based out of California. It's a developer company based out of California. Since 2019, we have been growing very fast. We started with one customer in 2019 to five customers in 2020, and ended with 40 customers in 2021. And most of these are in utility space, and we have been getting a lot of utility customers so fast, which is unique. Our founders include Rahul Saxena, who is our chief technology officer. He is a senior entrepreneur himself.

His previous company was in fintech space. And my third co-founder is Nitin Das, who is chief AI officer, and Antoine Tecky and an AI data scientist. His first foray into AI was way back in 2005. He started something because even when the AI was not mature at that time, he has been doing AI since then. So this is about our founding team. In terms of metrics, I mean, typically, as a typical Bay Area company, we have been growing pretty fast. Typically, our revenue has increased 5x year on year since we started. So our recurring revenue has increased 5x year on year since we incorporated.

Our number of employees has grown several times. I mean, we started with almost five employees in early 2019. And today we have 215 employees with global offices in US, where in US we have office in San Jose, Austin in Texas, and Washington DC metro area. Then we have office in London, which we incorporated last year. Then we have office in India. In terms of customers, we have customers in five of the seven continents, including 41 of the 50 US states and three of the Canada provinces. So, extensive set of customers across the globe. And this is a testament to the fact that the product demand is significant across the globe. Okay, great. Thank you. And as far as I've been able to tell, just looking up, doing a little bit of background research before the conversation,

it looks like you're basically developing a platform that you're using to first integrate different types of data and then develop different domain-specific applications, let's say, on top of this platform. So you already mentioned satellite data, but I think you also use other types of data. So I would like to ask you to explain a little bit the different sources that you use for your data and how exactly do you integrate all of those? Because as far as I've been able to tell, you also use data coming from sources such as sensors or perhaps even data from mobile devices and so on. So there's different provenance there and I wonder how you manage all of that. Okay, so our primary source of data is satellite imagery of various kinds, multispectral satellite

imagery and satellite imagery, which is available today. So we take that into our platform. And we combine that with additional sources of data, for example, weather data, which is very important for us. Third party or enterprise data within enterprise with are the assets, the location of the assets, for example, which we're monitoring if a storm happens. So those are very important data. Then we get data from, for example, for soil, which actually the predictions are made from satellite imagery itself. But satellite imagery, weather data and enterprise data are three core data sets which we use. These data go into our AI models.

We have various AI models for specific use cases, like encroachment model which detects how far the vegetation is from the conductor, asset health model, tree health model, outage prediction model, so various AI models which ingest these data and churn out insights. Where we differentiate from most companies, however, is that we don't just deliver insights, because we found that delivering insight is not a full solution, which traditional industries like. They want full fledged workflow solution.

So we pass on our insight to workflow applications which includes a mobile application and a web application. It is the mobile application and the web application that we hand over to our customer. Now this is easy for customer to use because they have an application they don't care about satellite they care about what to do, when to do, where to do and how to do. That is what we deliver. The advantage we get is because the mobile application is on the field, whatever predictions we are making kind of get validated, so our AI models get reinforced, so they keep improving, which is a significant advantage, which we have been a full-stack solution.

So this is our total platform. It takes satellite imagery and analytics, delivers those into workflow application which is handed over to the customers. Okay, I see. So I take it that well, you mentioned the fact that you use mobile applications and I imagine that these are mostly used by people who are operating on the field and need to get, for example, notifications for upcoming weather events and other events that may be of interest for them. I presume, however, that this is one class of users, let's say. I presume, however, that you must also have

a more detailed interface for analysts, for example, or even for data scientists, for people who want to dig in deeper in how you produce your insights and your predictions to be able to validate them or perhaps to go into the reasoning, let's say, that lies behind those to be able to understand better what's going on. So do you have such an interface? And what are the techniques that you use? And are they explainable and to what degree? Okay, interesting.

So let me give an example first that will set the context and then I can answer the specific technology question. Let's talk of storm prediction. That's a hot topic today. We have made an announcement today. So when we are making predictions on the storm, we are taking satellite imagery across the entire network of a power distribution company, for example.

We know each and every tree in the network. We know each and every asset in the network. We know their maintenance history from the enterprise data. We know each and every asset in the network. We know their maintenance history from the enterprise data. We know the health of the tree. Now when we supplement that with weather information, real time storm path, we can make predictions. The first prediction we make is 72 hours before the storm. What we tell is that, okay, when this storm hits,

when the landfall happens, these particular streets in these areas will see these many fallings, these many tree fallings and hence these many damages. To fix this problem you need these many people on field, you need these many extra people in your call center to handle extra demand and because there will be flood, we make flood predictions also, these areas may will not be accessible, these areas will be accessible. So you can deploy your crew on time even before the storm has hit and be prepared to do that. So yes, for the field crew there's the mobile application which goes, but for the planner, the storm planner, whoever is responsible for planning for storm, there are many such people,

they have the web dashboard in which they see real time status of how the storm is going to progress and where they need what. And they can make all of those plans, they can send communication to resources on field, they have to also plan for their housing, for the recommendation, their food,

significant amount of planning which they do, but they have all the information handy in a single web interface through which they can make those plans. So this is available. Now we don't sell our product to data analysts, we sell our product to end users. It's an end-to-end application. So we don't have any interface for data analysts, but for any compliance regulatory point of view from for explainability so for example if a wildfire happens it happened with the customer a wildfire happened happened the regulators asked the customer whether you are responsible or not

then from our data models we were able to submit data with proof that no the utility did the maintenance in time so they are not at fault. So those data are available for such compliance and regulatory requirements if required. But we don't expose the data layer to the analysts. We expose only the workflow layer. Okay. Yeah. Thanks for the insight.

And yeah, that's an interesting approach and well, a pragmatic one, I would say. So you are both compliant where you need to be, but you also don't expose much of the inner workings, let's say, of the platform. So let's come to what you're about to announce actually today. So a new application, let's call it, called Disaster and Disruption Management System. And as far as I've been able to tell, you already have been providing a number of applications.

So this is just the latest addition to your product line. So I would just like to ask you to say a few words about what it does exactly. And also a little bit about the pre-existing applications and where does that fill in in your product suite? Okay, perfect. And let me start with pre-existing application because then the technology progression would appear more clear. So the first application which we launched

was what we call intelligent vegetation management system. Vegetation management is a big problem in the US and globally. The US alone spends close to $10 billion every year on vegetation management. And still, this investment is not enough, which results in power outages, wildfire, etc. So our vegetation management system monitors power lines, oil and gas pipelines, etc. at scale using satellite imagery. So we know what is the clearance between vegetation and the power line today. We have growth rate model which tells how far the vegetation is growing. So we know what the clearance will be tomorrow. We have tree health model which tells which are unhealthy trees which can fall on power line if there's a wind.

So those trees can be removed. So vegetation can be better managed to avoid these wildfire and power outage challenges. This was our first product which has been deployed at over 50 utility customers. Then we launched our second product, which was sustainability management system at COP26 in Glasgow last year. That is again, looking at monitoring the land assets of utilities and other companies and giving them sustainability metric measurement in terms of what is a carbon sequestration, what is a carbon absorption, what is your biodiversity, and how you can meet your, what is the GHG emission, and how you can meet these goals in 2030 in terms of reductions targets which you have. That was our second product and it extends the technology from the first product where we were scanning the land at scale and identifying what, where are the trees, where are the grasses,

etc. So it extends well. The third product which we launched today, Disaster and Disruption Management System, that extends it further. We know each and every tree in your network. We supplement that with real-time storm data from various weather data companies which we have partnered with. And we get asset data from the customer. So we know each and every tree, each and every pole,

each and every location of conductor, et cetera. When we map that with weather information, we can make predictions for a storm. The prediction we make is before storm so that utilities can harden their infrastructure and during storm so that they can mitigate and help restore the power much sooner. That will result in saving of not only money, but also lives and discomfort.

And post-storm assessment of damages for government regulations, etc., as well as removal and clearance. The same system is also used on wildfire. Here, what we do in the tree health model uses the health of tree. So the drying trees, for example. We also have a model which tells us the moisture content and fuel load of vegetation. So for example, in California, a lot of wildfire is not caused by trees, but by dry grasses. So we can identify what is the dryness of the grass in the territory. For wildfire mitigation, we take sighted imagery twice a month, once every fortnight. Map that with real time weather information.

So we know, okay, the temperature is increasing, the grasses are drying. So these are areas where there could be more likelihood of getting, having a wildfire if something happens. So the utilities and other bodies which are responsible for wildfire mitigation can take preemptive action to reduce the chances of wildfire. So, so far the project is focused on storm and wildfire.

With time, we will extend it to other natural calamities like earthquake, flood. Flood to some extent is already included because we do flood mapping during a weather event to measure accessibility but a deeper deeper product around that function the development functionality around that will come as we progress the product. Okay I see. One thing I was wondering about is well how exactly do you approach making these predictions and feeding those models? Because from the sound of it and also based on what I'm a little, admittedly, I know about the domain.

However, I did have the chance to have an interesting conversation with people who combine data scientists and meteorology skills in the past. And what I've been able to gather from that conversation is that they seem to be applying a sort of mixed approach, let's say. So both using machine learning techniques, so using basically lots of data from the past to be able to develop models that should help them predict what's going on, but also injecting lots of specific domain knowledge into their models. So, for example, you just mentioned a fact that, well, in California, many wildfires start from dry grass, for example,

not necessarily from trees. So this is a type of domain-specific knowledge that is very beneficial to have and also to inject into the model. So I'm wondering if you could share a few words on a high level on how do you approach this modeling? This is a beautiful question. And this is a question which differentiates a product from a technology solution. So AI is good enough, but not good enough until it's deep domain. And domain becomes very important. company. We already have two certified arborists in our company who understand vegetation. We have a pipeline integrity expert in our company, and they have worked with utilities across several years for maintenance and operations activities. So we have these three main house, and their knowledge has been used in building these products. And more importantly, identifying what variables are more important than others.

Like one simple example, right? If there's a storm in the desert, it's not going to cause so much of damage because there are less trees. More than 50% of outages that happen during a storm is because of falling trees. Trees themselves don't generally fall, except if there's a two biggest storm

and a big tower which is structurally weak, that will fall. But generally it's the tree which fall on the wire and snap the wire and take off the pole and do all those damages. So that gave us the clue that understanding tree is more important than understanding weather. There are so many weather tech companies. In fact, we partner with them.

We don't compete with them. We take their weather data and we believe that their model of weather prediction, which itself also is a complicated model, which works. But then we supplement that with tree knowledge. And the second domain information we take is asset maintenance. Now, transformers, parks, breaks when there's lightning and all those things, right? So when was it device maintained? When was it last serviced? That is another domain information which we take into our model, right? So these localized domain information

is what makes our prediction granular. As I told earlier, we don't make a prediction like, okay, Texas will see this much damage. We make a prediction that this street in the city will see these many damages. So we're very granular. And granularity comes from deep domain knowledge

and the domain data, which is ingested into the AI model on top of the satellite and weather data, which is much more broader. Yeah, indeed. You did mention that the predictions you're able to produce are very fine-grained. And I was wondering how you're able to achieve that. And as part of that, I was also wondering, because you did also mention the fact that you utilize different types of data, so primarily satellite imagery, but also other data sets, historical data, local data, and so on. So I guess that means that the models that you're developing must be multi-modal and that sort of, on the one hand, that's kind of state of the art. And I guess it also helps you achieve that granularity that you talked about. Yeah, so if I understood the question correctly, so the models are actually, there are many models. So just for a storm prediction, for example,

at least seven, eight models are being used. And for wildfire, we have wildfire probability model, we have wildfire extent model, which tells, okay, if the wildfire happens, how much it will extend. That takes into account the distribution, vegetation distribution pattern, fuel load. So there's underlying model

on the fuel load assessment model from pre-Hong Kong. When we take satellite picture, we do a 3D elevation model of vegetation from which we compute the volume of vegetation and the fuel load and moisture. So there are many models which actually run together in symphony to make these predictions. That is true. I see. I think you must be probably familiar with the so-called model for analytics, which describes a linear progression in analytics.

So starting from descriptive, so analytics that tell you what the situation is, going to diagnostics, telling you why a certain outcome did come about, then predictive, the analytics that tell you what may happen to prescriptive, which is the end goal, let's say. So, analytics that are supposed to tell you how to achieve certain desired outcomes. On that scale, where would you say your product fits? Yes, typically, typically, in a progression, I'm not that time varying. So we have this another thing which is called, I mean, my AI scientists will say better about that, but it's called Kalman filter, which has time varying coefficients, and gives us much better prediction in a more time varying fashion, which is relevant to take

care of seasonality in predictions. So that is one thing which we use. There are some DNN models which are also used, especially on the image recognition side and identification side from the satellite imagery. So these are the two broad level of kind of model which we are using. Then many of these data are time series data where we see spatio-temporal change detection as to how, for example, for health of tree, right? Health of tree is a spatio-temporal change detection, which we see when monitoring the same tree across several years and seeing how the health is changing, very similar to how we do best X-ray determination using multiple X-ray measuring.

So these are some of the techniques which are used in our models. Okay, I see. And since I guess we're close to wrapping up, let's do that by asking you, what is your roadmap going forward? So we just unveiled a new product. What are your goals for the coming period in terms both of product development, but also imagine business development? For example, could this be of interest besides utilities, which you mentioned is your main source of clients to this point? Could it also be of interest for public authorities, for example? Yeah. So in terms of business first,

so in terms of business, as of now, we are selling to power utilities, gas utilities, and also energy companies, oil and gas companies. The same technology is very much useful and available for government cities municipalities etc and that is something which is in our roadmap going going next and in certain cases because there's a significant advantage of data right when we get data from somewhere in a certain region the same data can

be used to deliver solutions to different entities so for government entities some of these could also be given free of cost because i mean we don't have incremental cost so that is the direction which we may possibly take especially in a disaster disaster scenario so this is how we progress across industries across geographies we have customers in all geographies and we are expanding fast. Our goal is to get listed around 2025 and we are an accelerated growth path for the same and we will continue to grow accordingly.

In terms of product, these three lines of product which we have, one is the vegetation management, which is getting extended into what we are calling distribution asset management system, which will work for power line, gas pipeline, roads, railroads, et cetera. And sustainability management, which is utility, energy, water,

wastewater companies, construction, et cetera. And disaster management, which again, starting from utility, will also go into public, government and regulatory bodies, et cetera. So these are three products which will expand across different industries. And these are pretty complicated challenges and pretty big market to focus on.

And we are a deep domain company, so we will remain vertically focused on these three products, which I think is a significantly huge market for us to play for the next few years. I hope you enjoyed the podcast. If you like my work, you can follow Link Data Orchestration on Twitter, LinkedIn, and Facebook.