Episode 2: How Aurora Solar expanded its business with Machine Learning
TIGRAN PETROSYAN: Hello, everyone. I'm very excited to welcome you to our next episode of the Real-Life AI podcast series. I'm your host, Tigran Petrosyan. I'm the co-founder and CEO at SuperAnnotate. We're building an end-to-end data infrastructure for various machine learning applications.
I'm also particularly excited to have Chris [Christopher Hopper], the co-founder and CEO at Aurora Solar. As the name may suggest, Aurora Solar is a solar industry software, and it's making the solar installation processes very quick and easy. Of course, Chris will tell more about this. The company has already raised over half a billion dollar. Hey Chris, great to have you. How are you doing?
CHRISTOPHER HOPPER: I'm great. Thanks for having me, Tigran.
TIGRAN PETROSYAN: Excited about this conversation. Can you share more about what's your story and how you ended up founding Aurora Solar?
CHRISTOPHER HOPPER: Yeah, happy to. We started Aurora almost ten years ago, at this point, in 2013. My co-founder, Sam and I, met in business school at Stanford, and we became friends early on in the program. And as we were talking about what we might do after school and where our interest lie, I shared some of my background with Sam, which was the off-grid electrification in space. So I spent a couple of years trying to figure out how to sustainably electrify off-grid communities in developing countries. And out of that discussion came the idea of starting actually a larger scale solar installation business, again focused on emerging markets. So building medium-sized residential and commercial solar installations, in particular in East Africa. And so, to show that we can do that, we decided to put together pilot projects. During nights and weekends, we would meet in the library and dorm rooms and design a system. We found suppliers, we raised a loan, and we put the whole thing together. And then, ultimately, we flew to Kenya and installed the system. So that was back in 2012 while we were still in school.
And then what started happening is other people started reaching out to us and asking us, "Hey, does solar make sense for me, for my home, for my business?" Because they saw that the system worked, what a math happened. The school had no more power cuts to save them a bill every month. We got sort of inbound interest, and the question was always the same, "What can solar do for me? Does it make sense for me to go solar?" And our answer was, "Well, we don't know. It depends". Every system is different. Every building has a different roof structure, different environments, and different energy consumption profiles. And so, what ended up happening is that you go through a design process that's new every time.
That's how we realized that to scale this as a business, we really needed software to streamline the process of going solar. And then we looked into all the details, the markets, and so forth. We saw, "Whoa, there's actually a lot of costs associated with the process of going solar!" In fact, in the US, more than half the cost of a solar installation is not the equipment, it's the soft cost that sits on top. And so it became clear to us that to take solar to where it needed to go, there needed to be a software platform to power this transition to a future that's in large part powered by solar energy.
TIGRAN PETROSYAN: Amazing. If you remember these early years, what do you think was the most difficult part of the building, basically, the MVP, the early product?
CHRISTOPHER HOPPER: The product was… well, what we do is a lot of deep technology. So that's actually one thing that differentiates us in our market, is that we take a pretty rigorous approach to everything we do.
So we're just a series of engines, if you will, that do different things. We have a simulation engine that simulates the energy production of a system. We have a ray tracing engine that calculates how much sun which part of the roof in the system receives. We have an optimization engine that designs optimal solar installation. We have, of course, a machine learning engine as well. We have a sequence of engines. Each one of them was a significant development effort. So I guess that was a challenge initially. Well, it obviously takes a lot of work to put all that together.
And also, it requires a lot of people with very specialized knowledge. So also building the team that can build that, that was also one of the early challenges. Although, to be fair, we actually had maybe an easier time with it because we had a great mission that appealed to people. So we found really talented folks who wanted to join us and help make this the future that's fired by solar become a reality.
TIGRAN PETROSYAN: Yeah, really, really amazing! How long did it take to really build that? Because I can imagine building that sophisticated system. It's not like many software products building an MVP and testing, [it] may be very quick, but in your case, it sounds pretty sophisticated.
CHRISTOPHER HOPPER: (laughs) In a way, ten years. But the first version probably took a year and a half, something like that. It's constant iteration, you get feedback from the user, and you add features and repeat.
TIGRAN PETROSYAN: Do you remember the time when you felt like, "Oh, my God, it's working, it's going to scale. There's nothing like this in the world"?
CHRISTOPHER HOPPER: Yeah, there were multiple of those, in a way. The first one was when we received the first check, the first payment for our software. And that was a big thing. It was just going from, "OK, we're putting together this cool project" to "Someone is willing to pay us for this," which was pretty amazing. It was only $159. We still have a physical check. We had no billing system. So we literally got a check in the mail. I walked to the Chase Bank down the street, and I cashed it. First revenue. That was huge. Not because of the amount but what it meant.
Then the second was actually when we hit 1 million in ARR, which was probably a year after we launched a product or so. That was pretty amazing, too. The million dollars, you know, was a large number for what I was used to. And so the fact that people are paying us that sort of money was quite special.
Also, to note, maybe in tying it back to your question about challenges, early days fundraising wasn't that easy for us. Now, we've raised a lot of money, but back in 2013, it was a very different environment. Not necessarily in terms of the macro, but solar, in particular, was not a very hot space. And so we ended up bootstrapping for five years, and so revenue was particularly significant.
TIGRAN PETROSYAN: Wow, bootstrapping for five years and keeping the team together. I guess you really need that strong mission and vision. Everyone has to be super driven during the time to really have that belief that it's going to make it work, right?
CHRISTOPHER HOPPER: Yeah, it was that. Also, we could see it working. So the product was working. The customers love the product. We can see one key metric we track is how many solar projects we touch in our software. And that number just kept going up and up and up. We had a lot of indications that things were working and the mission that we all believed in.
TIGRAN PETROSYAN: Awesome. I know that you've started using computer vision to further optimize your processes. Can you share more about how this idea came? And what is that about, how it helps to further improve your processes?
CHRISTOPHER HOPPER: So the design process… our software has these engines that do different parts of the design process. If you can think of the process of getting to an optimal solar installation, there are multiple steps, right?
There's first, understanding the energy consumption of the buildings. What's the utility bill? How much do you pay before you go solar, which you can reduce because you now start producing your own energy with solar? So how much energy do you consume? During what times of the day do you consume energy because of how utility rates work? So that's one step.
The second step is understanding the environment. So what's the roof like? Other neighboring buildings, mountain ranges, and trees casting shade. What's the weather like? That's step two.
Step three is the system design, so configuring how many panels fit, where to put them, and so forth.
Step four, what's the energy production going to be like based on what you put together?
Step five, what are the savings like?
And then step six, the finance the system. After everything is said and done, how much do you save on your system or your bill every month?
If you think about it, sort of this sequence of steps. We started working on optimizing and automating as many as possible, each one at a time. The big one that was outstanding, that was really difficult to solve, is the step of creating a 3D representation of the building. We have our own CAD engine if you will. You can 3D model your roof. That's one of the things where our product really shines. But it was a manual process. And so we were trying to figure out, "How do we automate that part?" That's a necessary part because residential solar goes on roofs. There are 100 plus million roofs. How do we design solar at scale? So ideally, the dream is, "Can I just snap my finger and design an optimal solar installation for every roof in the US?" Or even in the world, right? That was sort of, in the limit, the vision we wanted to get to – to be able to do that. And the necessary part there was to automate the roof modeling and create a 3D model of the building. Six years ago, probably, we started our first efforts in computer vision.
TIGRAN PETROSYAN: Perfect. So I'm wondering, it takes quite some time until you get from an idea to some working product around machine learning. Can you share how long it took from the first idea until you realized that, okay, this AI application, AI installation into your system, is working? Because according to many estimates, over 80% of ML projects fail. And when is the right time you realize, okay, it's working, it's not working? Or what it took you to get it working?
CHRISTOPHER HOPPER: It took a while. It took several years until we had something. And one thing was also state of the art – that shifted so much.
When we started, our initial approaches were still very much based on sort traditional computer vision rather than machine learning. Then all these advances started rolling in. We shifted our approach towards a machine learning-based approach. We were sort of catching up to and then, at the cutting edge of what was happening, which was very, very exciting, but it took quite a while.
The other thing is, it's not the challenge whether it's working. Because something was always working. The question is, sort of, how you map that to the needs of the market. What are the use cases, and how do you meet that threshold? It's a bit of a blurry thing, and especially as we're sort of bumping against state of the art and to our knowledge nobody has done what we've done and has the same, you know, there's not a textbook solution for it. That was a bit of uncertainty I guess you're waiting into and faith you have to have.
So it took us a couple of years, but it's something we also knew and a risk we were willing to take. To be fair, it wasn't the core of our business. We had a working business that was there. This was sort of the next generation and something where we believed in where we could take that technology and bring machine learning to renewables, to clean tech. We thought this would be an amazing thing that's super impactful if we could make it work. Recently it's been terribly exciting because when you see it come together, it's, I mean it's magic, right? You hit a button, and everything happens. Now it's even on many dimensions outperforming humans, right? The bar for humans is also not perfection, right? Humans need to be trained and make errors and mistakes, and so to be able to augment them and then give humans that sort of leverage, it's pretty powerful.
TIGRAN PETROSYAN: This is such a great example of how machine learning projects work. Generally takes time, energy, and commitment for a long time. And if you spend some time, let's say half a year or a year or, at some point, it's so easy to give up because it requires so much data, it requires so much iteration, so much building. Iterating and understanding what works and what doesn't work. And this is such a great example of persistence that when you have a goal and then you're persistent in iterating well, it will work out, especially in data science and computer vision projects. And this is what we're trying to push forward for every other client who's building machine learning and, of course, helping that to be successful from our side. Such a great story there.
CHRISTOPHER HOPPER: It was definitely something that's risky, but we can see the payoff on the other side – that's a huge game changer.
TIGRAN PETROSYAN: Absolutely.
CHRISTOPHER HOPPER: That helped us persevere.
TIGRAN PETROSYAN: Absolutely. If I switch the gears a little bit, you have raised so much great fundraising grounds, over half a billion across several times into our development. I think the last one was 200 - 250,000,000 within a short period of time in a row. What's next for Aurora Solar? How are you utilizing these funds to bring up Aurora Solar to the next level?
CHRISTOPHER HOPPER: Our mission is to create a world that's powered by solar energy, to make it so that everyone can benefit from clean, renewable energy. Yeah, that's why we get out of bed every day.
Now, the way how we do that is by streamlining the process of going solar to bring down the cost of solar to the end consumer and to make it easier to design these systems at scale. That's always been what this has been about.
Increasingly what we're doing, though, is we're also touching more and more of the process of going solar. So we started with a design solution. You can think of it a little bit like not perfect, but AutoCAD for solar, maybe. And then we expanded from there. So now we have sales tools. We have a machine learning-based product called Lead Capture AI that lets homeowners self-qualify and basically design a system for them. We're interactively building more tools, a toolkit for solar professionals.
Also, there's a big world out there outside the US. So we're also tuning eyes to that. So that's going to be an exciting new chapter for us as well.
TIGRAN PETROSYAN: Perfect. Is there already a public number? How many houses or homes have been installed with solar panels through your system?
CHRISTOPHER HOPPER: We don't track installs per se, but first of the metric is the design because we also used it early in the funnel for evaluation and quoting. But more than 10 million buildings have been designed in Aurora. So it's a huge, huge milestone for us. And it's incredible to see that number go up every day, really, every week and every day.
TIGRAN PETROSYAN: Wow, that's so exciting! I can't wait to see what's next for Aurora Solar and you guys. I'm such a big fan and, especially from us, seeing every day, in every industry, technology starting to adapt to new AI applications and new processes around AI.
Surprisingly, you know, there are industries that I would never think machine learning is coming to, and seeing the solar industries also being implemented with machine learning applications, of course, that's adding even further fascination for me. So I want to wish all the best of luck to you and Aurora Solar.
For the listeners, I just want to remind our guest today is Chris. Chris is the co-founder and CEO of Aurora Solar. They're building the future of solar energy software to make it much easier and quicker, and more understandable where to install solar, how much savings you'll get, what it will take, and so on. So really excited to have you with me, Chris, today. Thanks for your time.
CHRISTOPHER HOPPER: Thanks for having me, Tigran.