Name: AI in action: Scaling accuracy, automation, and API-driven solar workflows
Uploaded: 2026-03-26T19:16:09.096Z
Duration: 36 min
Description: AI in action: Scaling accuracy, automation, and API-driven solar workflows

Transcript for "AI in action: Scaling accuracy, automation, and API-driven solar workflows": Hi, everyone. Welcome. Really glad you're here. I'll just wait a few more seconds while folks trickle in. Alrighty. I will introduce myself. My name is Elia. I am a product manager here at Aurora, and I manage the site design and AI roof products. So everything from modeling your roof, trees, shading, LIDAR, imagery, weather data, basically, all of the tools that you need to get an accurate site design. And today, I will be guiding you through the very exciting world of AI and solar. We're going to cover a lot of ground, so I wanna make sure it all kind of hangs together as one story. We'll start by zooming out. We'll take a look at how AI is being used across the solar industry right now, not just in design software, but in manufacturing, on the grid, everywhere. Then we'll zoom back in and talk specifically about how Aurora uses AI for site and system design. We'll walk through some of the improvements we made over the last year on our AIRoof product. From there, we'll look beyond AIRoof. How else is Aurora using AI across the product and across the company? We'll close with a look at what's coming next on our road map. And I'll drop the seed now, but there will be a live poll at the end. And in this case, your feedback will actually shape our priorities and what we wanna build. And there will be some time at the end for q and a. So as we go through all of this, please feel free to drop whatever questions you have in the q and a window. It's right there next to your chat window, and we'll get to as many as we can at the end. Oh, also, fret not. This webinar will be recorded, and it will be sent to you via email once it's all over. Alright. Let's get into it. Let's talk about the market. So I'm not going to sugarcoat it. The market is really tough right now. With the 25 d tax credit that was eliminated at the end of last year, OM Analytics is forecasting roughly a 20% contraction in the residential solar market in 2026. And, honestly, most of us in this room don't need a forecast to know that. We've been feeling it. The good news is, if there is good news, is that analysts anticipate that h one of this year is that low point, and volume is expected to start climbing again in the second half of the year. So that means we're likely in the trough right now. When you zoom out a little bit, the picture looks a lot better. Both Ohm and Wood Mackenzie are forecasting that this residential solar market will return to steady growth over the next several years. You can see that chart sort of going up into the right. Ohm is predicting 17% growth next year alone. And within five to ten years, we could see volume back at the installation levels we saw in 2023. So all this to say the industry isn't going away, far from it. The underlying demand drivers that we know of, rising energy costs, grid anxiety, energy independence, those are not going away either. The question isn't whether solar comes back. The question is, what do you do between now and then to make sure you're in the best possible position when it does? And that brings me to what I think is the most important topic in this industry section, section, which is soft costs, everything that isn't the hardware. Soft costs still represent about 65% of the total cost to go solar. That's customer acquisition, sales, design, permitting, installation labor, all of it. Now a lot of those factors that influence those costs are outside of our control, like policy, regulation, interest rates. We can't do much about those. But efficiency, that is the lever that we can, like, actually pull on. Moving projects through faster, reducing rework and change orders, eliminating manual steps. Right now, the compounding effect of getting more efficient is huge. So the way we see it, every project you move through more cleanly in a down market makes you, like, stronger in a stronger position when that volume starts to come back. So I like to think of this period of transition as an opportunity, an opportunity to think differently, adopt new tools, and just come out of this on the other side leaner and stronger. And that's where AI comes in. So let's talk about AI. That's why we're here after all. I wanna be clear here. AI is a word that gets thrown around a lot right now. I personally live in San Francisco, and I can't turn my head without seeing a billboard about AI. So rather than just, you know, talking about it abstractly, I wanna show you what this actually looks like in practice and at different levels of our industry. So let's zoom all the way out. So AI and solar isn't just about design software. It's really, like, reshaping our industry in a lot of ways. At the manufacturing level, MIT and Stanford researchers are using machine learning to optimize their sort of solar panel cell production research. I'm I'm I know I'm just gonna butcher this name, but perovskite panels. Basically, these are the panels that will eventually replace silicon altogether. And the manufacturing process of these are really complex. They involve juggling, like, 12 or more variables at the same time. And AI is able to compress what took years of lab work into something that's much faster. At the system design and sales level, which is obviously a lot closer to home for us, AI is compressing the design to quote process. We'll spend most of our time here today, so I'm not gonna get into it right now. But if we look at the grid level, there's some new research published just this month in the IEEE journal, the journal of open access journal of power energy, it's called. And they're showing that AI can actually run grid level stress tests. So that means, like, modeling extreme weather conditions, sudden demand spikes, equipment failures, all of these things. And they're able to do it, they're saying, 79 times faster than conventional methods. And we know that Duke Energy is already using cloud based AI to run hundreds of millions of simulations overnight to plan infrastructure investments for the future. And finally, looking at the virtual power plant level, the US Department of Energy projects that AI coordinated virtual power plants could cover 10 to 20% of US peak grid demand by 2030. So that's, like, your customer's home battery, your neighbor's solar panels, all of these things working together as a fleet. That's huge. So the point I wanna make with all this is that solar isn't just, like, a product that AI helps to sell. It's an industry that AI is really helping to reinvent at every level, not unlike what we're seeing with under other industries right now. Okay. Let's zoom in on where it matters most for your day to day business. Next, system design and sales. So we know that solar is not a one size fits all solution, and that's actually part of what makes it really complex to sell and design. So every single site is different. The roof structure, the shading, the trees, the neighboring buildings, the location, the weather patterns, the AHJ requirements, all of that all of that will determine, like, where your panels can go and how the system should be laid out. And on top of that, you've got the homeowners utility rates, their consumption patterns, their preferences. It's a lot of inputs. But AI can help to process so many of these inputs and produce an accurate three d model, and we're looking at that happening in about ten seconds right now. So something that might have taken you hours if you're doing it manually is now, you know, down to just an a number of seconds. And once you've got that model, you can, you know, design a complete system and proposal in minutes. The goal here ultimately is to make this feel almost automatic, and we're getting a lot closer to that every year. And here's what it could look like at its most automated. Some of our customers now are triggering the AIRoof run directly in their CRM. So the moment a lead is created, the process starts. The model like, once the model is finished, you can trigger fire pathways based on your local HJ requirements. You can run auto designer to lay out the system based on your preferred components. And then a radiance analysis runs, performance simulations run, and all of that data syncs back to the CRM. So with just a few inputs from the homeowner, you can get a fully completed site and system design in under a minute using this process. And then you're doing all of that without ever opening Aurora. And then with your proposal template and pricing already configured, the proposal's ready to go the moment you need it. And this is kind of a picture of what AI looks like when it's really embedded in your workflow rather than just sitting on top of it. And before I get into what's new, I just wanna spend a minute on something very foundational. Because AI is such a loaded term right now, it's worth being precise. When you hear about ChatGPT or Claude, those are LLMs or large language models. They're trained on text, and they generate text. AIRoof from Aurora is completely different. This is a computer vision model. So this means it's trained to make predictions and decisions based on structured data. So in our case, it's trained to look at aerial imagery, LIDAR data, parcel data, three d models, other inputs. And this is something that took years to build, and it's trained on millions of real solar designs, and it's validated by experts. So this model was basically built for one job, and it's getting the roof model right. And that actually leads me to the competitive landscape because the I know a lot of you might have seen other software out there making some rather bold claims. Some are saying that they can make a model in as little as five seconds with near 100% accuracy. So what's how is our model different? It comes down to two things, accuracy and speed, and most critically, not sacrificing one over the other. On the accuracy front, Aurora has been setting the standard here for design accuracy for over thirteen years. Our AI model has been trained on over 4,000,000 runs, 500,000 expert validated designs, and it goes through a very rigorous QA process before anything change or anything shifts. So when you hear other companies citing accuracy studies, I just I urge you to read them carefully. A lot of those studies are comparing production estimates to an actual installed system output, which sounds great. Right? But that assumes that the site was actually designed correctly in the first place, and that's a big assumption if you're working off of a two d model with no guardrails. Today, Aurora is the only solution that I know of that validates the three d model itself and not just whether the panels fit on top of it. So that's on accuracy. On the speed front, we worked incredibly hard to get to a ten second model generation without cutting corners on quality and accuracy. We'll talk more about those improvements when we get there in a little bit. But what I what I really wanna drive home here is that it's truly the combination of the two things, fast and accurate, not one or the other, that makes our AI model different and really reliable in the real world. Alright. Let's take a pulse check on how AIRoof is actually doing and what we've built lately. Let's start by looking at the numbers first. In 2025, so last year, over half of all Aurora projects started with an AI model. We've crossed nearly 5,000,000 total runs since its inception. And our acceptance rate, which means, like, designers are actually keeping the model rather than replacing it or starting fresh, is at 96.5%. And this has been steadily climbing over the last few years. That last number is the one I'm the most proud of because it's the true signal of quality. Right? It's not a marketing stat. It's people working voting with their workflows, choosing this tool again and again. So let's take a look at what's actually driving that acceptance rate higher. One of the biggest updates over the past year is that AIRoof is now available through an API. I know we talked about this a little bit earlier, but the workflow before this existed was someone had to manually trigger the AI run inside of Aurora, which worked fine, but it introduces a sort of manual step and causes friction and slows you down. So with the AIRoof API, you can generate this model automatically through a simple API call. So the moment of projects created in your system, you fire off the request, get back a model, and you never have to open Aurora at all. And this is what makes that fully automated workflow I described a few slides ago actually possible. So this turns Aurora from, like, a tool you run manually into the infrastructure that can power this solar workflow. We've also made significant improvements to the speed of our smart roof generation over the course of the last year. Some improvements to the underlying infrastructural and the the model execution pipeline, we've been able to reduce our run time by roughly 30%. So that takes us from about fifteen seconds last year to about ten seconds now. And I know this might sound like a very small delta, but when you're running dozens or hundreds of designs a day, that really adds up. Our goal has always been to make this feel almost instantaneous, automagical, if you will. And this is something, like, that you would trigger and move on from, and we're getting a lot closer to that. What else is new? So the third area, like, of major improvement we've done has been the quality and diversity of the data powering our models. We spoke a little bit earlier about, like, the two levers you can pull on speed and accuracy. So this is on the accuracy front. AI is really only as good as the data that it's trained on, or as we say, garbage in, garbage out. So we've made a lot of meaningful investments over the last year in this zone. We've expanded our training datasets to cover more geographies. That includes additional data from markets like Germany. We've improved the types of input data that our systems can use. So that's higher quality LiDAR data support for raw DSM data point cloud data, if you will. And I know it's been a long time coming, but Aurora now supports importing and processing drone data natively. I'm gonna double click on this one a little bit. So satellite imagery and LiDAR are great. We know this. But there's times where they have limitations. They can be outdated. They can be lower resolution in some markets. And for complex roofs, they don't always give you the full picture. Drone data starts to solve a lot of those problems. So when a sales rep or an installer flies a drone on-site, they're capturing the most up to date, high resolution imagery, and it reflects exactly where the site is today. Now we can take that drone data, those images, and process process it into, like, a background map that's made directly from your images on-site and a LIDAR model. From there, you can run a roof AI roof directly on top of that. So the result is basically exactly what you'd expect. Better inputs, better outputs, the more accurate roof geometry, better obstruction detection, and, ultimately, the better your designs will hold up in the field if your input data is stronger and more accurate. I mean, just, like, look at the difference between these two models at the same site. The drone model, as you can see, is a lot crisper, a lot fresher. Okay. I think at this point, we know AIRoof pretty well. I would be remiss if I did not spend some time on what's happening beyond it. Just because AIRoof is our most visible AI product, that doesn't mean it's the only one we're working on today. Let's talk about the AI chatbot for homeowners. I'm very excited about this one. So the solar journey, as we know, it starts long before a rep even picks up the phone or knocks on someone's door. A homeowner might be curious for solar about solar for, like, weeks or months before they ever fill out any forms. And in that window of weeks or months, they might have a lot of questions, and sometimes there's no one there to answer them. So we've been experimenting with a conversational, like, AI assistant, if you will, that acts like a solar adviser. So it can answer questions about solar, explain savings. It can guide homeowners through the early stages of evaluating whether or not solar makes sense for them. And my favorite part is it uses Aurora's own solar content as the source of truth so you can trust that the answers are accurate and grounded and AI is not hallucinating. And as it does that, you you're capturing key information. Right? So the homeowner is talking to you. They're telling you their location, their electricity usage, the details of their roof, and those inputs will flow directly back into Aurora and into your sales workflows. So there's a big difference between a homeowner who shows up to a sales conversation already having answered those questions versus one that has not. So this is still in beta, but it's one of those areas that we're really, you know, excited about exploring. And that same investment that we're sort of asking you, urging you to take into your workflow, we're doing it internally too. Our engineering team uses AI to shift faster, to catch issues earlier so we can spend less time on this, like, repetitive stuff and more time building features that actually matter to you. There's some examples of that here on the screen. And it's not just engineering either. You actually caught us in the middle of our company wide AI weeks right now. I swear this webinar was not planned that way. So right now, every team at Aurora has dedicated time to rethinking about how we work with AI. We wanna stay on the forefront of technology. We know these things are moving so rapidly. Tools that you may have used, you know, last year are completely different now. We wanna move quicker. We wanna enable, like, ourselves to deliver value to you a lot faster, And that's actually paying off. We're seeing faster releases, better tooling, a product team that can spend less time on repetitive things and more time solving problems that actually matter to your business. Alright. Let's take some time to look forward. We talked a lot about what's here today, what's improved. Let me show you a little bit about what's coming. Here's a little look at our short and long term road map. I want to be transparent with you about how we're framing this. Some of what I'm about to share is in active development, meaning, you know, we're we're ready and starting to build these things now. Some of it's a lot further out. So there's things that we're actively exploring, but we haven't committed to yet. So let's take this as a bit of a sneak peek, a look behind the curtain, if you will, and I'll call out the differences between these things as we go down. And and part of why I'm sharing some of this, like, exploratory things for the future is that we genuinely want your feedback. So I'm setting the stage here a little bit, leading the witness. We will have a poll at the end, and would love to hear what your ideas are and what you think would be you know, move the needle the most for your business. Existing panel detection is one thing that we're actively exploring. So battery add ons. We know battery add ons are one of the fastest growing parts of the market right now. And almost always, that involves an existing solar system on the roof. So the problem is that in today's workflow in Aurora, you're dealing dealing with that can be kind of painful and cumbersome. If you've ever tried this, you'll know what I'm talking about. You're jumping between screens. You're, you know, manually entering production data, and the whole thing isn't really accessible in sales mode. We are aware of this, and we think that AI can start to help solve some of these problems. So if we're able to look at the roof and know what's already there, basically, detecting existing panels, prepopulating the system specs, and bringing this workflow into sales mode can start to help modeling battery add ons a lot easier. And this is one of the things I'll be pulling you all on today, so please keep it in mind. Another thing we can talk about is tree detection using AI. Anyone who's worked on-site with, like, significant vegetation knows how tedious this can be today. If you're not using LIDAR shading, you're hand modeling the trees one by one. You're estimating the pipe, the canopy spread. We're building an AI that can do that automatically. So we can detect trees from aerial imagery and lidar data with some other inputs and generate a full three d model with location, height, canopy spread, no manual input required. So we can go from, like, five to ten minutes of manually modeling and adding trees per site down to a tech target of about fifteen seconds. This is an active development, and we're currently training our tree model, but it's another one of the features I'll be asking you about a little later. And proposal assistant. This one is firmly in what in the, like, what could be category. We're not building this tomorrow, but it's worth planting the seed here. So the idea is, like, what if a rep could walk out of an appointment with a proposal that's already done and if editing that thing was just as simple as asking. So we're exploring what that could look like, and we'd love to know if this is something that resonates with you. Essentially, you know, improving the creation and editing of proposals using AI. Okay. Without further ado, let's get into it. Of the three things we just talked about a little bit earlier, tree detection, panel detection, proposal assistant, which of these do you think would be the most impactful for your workflow? And if you have another idea or something that we haven't even thought of or talked about, you know, in this time, please type it in the chat. We'd love to hear your ideas here. The you can find the polls tab that just popped up right there next to your chat window. I'll give you all a few seconds just to answer that. I see some votes coming in. Thank you all so much for participating. Alright. Just a few more seconds, and let's see what the results are. Wonderful. Okay. We've got tree detection as a a resounding winner here with 60% of the votes. And it looks like the proposal assistant is another one of those things that you all think would be really impactful. Looking at the chat, I see a recommendation for better tooling for going back and adjusting roofs and array spacing. Oh, and then some improvements for EagleView without needing to create a new project. Wonderful. Thank you so much. And that brings us to the end of this presentation. I will slow down now. Thank you all for spending this time with us. I hope it gave you a window into where Aurora AI is and where we're going and how we're thinking about the market while we're navigating all of this together. Now I wanna open it up to you all. What questions do you have? If there's anything you have, please drop it in the in the q and a window, and I'll work through as many as I can. Okay. I see a question here around battery only stand alone proposals. Are they going to be offered in the near future? I I can't tell you exactly when they'll be ready, but this is something that we're actively working on. We are prioritizing this quite highly, especially because at the same time, we are focusing on improving the way we model batteries in design mode in Aurora. So bringing batteries into the workflow and modeling them more accurately will allow us to unlock battery only proposals right after that. I see a question here on the chatbot. The question is, is the information from the chatbot ours, or does Aurora keep it to the market and sell it? So, ultimately, the idea for this one is that your information would be completely private. This is something that, you know, is is currently being explored. So, you know, nothing is live yet, and nothing is going to be sort of shared out. But we're highly aware of any of the, you know, privacy concerns and want to keep all of your private information private. I have a question here around how hard is it to set up the drone imaging in LiDAR. So this is something that is currently in beta, and everyone should already be able to see it and have access to this feature. If you open up a project in the same place where you might look for your you know, to order an expert design model, you'll see a new tab there for importing your drone data. Right now, anyone can use it. It's completely free, and there's some, you know, added instructions there for clarity. But, ultimately, it's as easy as go to the site and take, you know, 20 to a 100 pictures with your drone and then import them to Aurora. It takes about twenty or so minutes depending on how many images you upload for that to process, And then you'll receive the LiDAR in the in the background map image as a result of that process. I see a question here on how accurate is the AI at detecting the height of obstruction, the roof vents, etcetera, on a flat roof. Ultimately, we hold our models to a very high standard. So when we're looking at accuracy values, we're actually comparing what a, you know, human being could model with the same data and holding our AI to that standard. So I I I wish I could answer this better as, like, how accurate is it? It's as accurate as a human could do almost 85% of the time, if that makes sense. So we're basically targeting a goal of, like, if the data is the the data that you're receiving has enough information about the height of those obstructions, they will be reflected in your AI model. I have a question here on AI for commercial. So you can run AI on commercial properties. However, our model is trained primarily on residential properties. So commercial is something that is, you know, on our radar, and we want to improve in the future. But we have found that it's much more effective on on smaller properties where the data is more up to date. Any other questions? Please feel free to drop them in the chat. If not, we can wind down now, and I really, you know, appreciate everyone spending a part of your afternoon with me here today. Alright. Thank you very much. Thank you for joining. And as a reminder, you will get a recording of this sent to your email inbox afterwards. Have a nice day.