Visions Host Jim Tatum interviews Christopher Van Dyke, CEO and co-founder of Overview AI, about the company’s mission to bring automotive-grade AI perception into factory inspection. Overview AI offers an easy-to-deploy smart camera that detects variable defects, trains in hours, and logs real-time results on the edge.

The conversation covers why they leaned into AI early, challenges around sparse defect data and synthetic augmentation, integration advice for manufacturers and integrators, use-case limits (mix/volume and speed), and how the technology simplifies and improves accuracy for visual quality checks.

Visions: A Machine Vision and Automation Solutions Podcast, is the podcast for engineers, designers, integrators, and end users who want to keep an informed eye on the imaging and machine vision industry. Every Tuesday we will explore the latest in imaging trends, developments and solutions. Here you will find interesting, useful insights and observations from expert interviews, solo episodes, even the occasional panel discussion, all of which aim to expand your knowledge on imaging and machine vision. 

Related: Visions Podcast: Embedded AI and Smart Cameras

Related: Visions Podcast: Cognex Talks AI Vision on the Factory Floor

Related: AI Based Machine Vision Boosting Distribution Yard Efficiency

Transcript

Well, hello and welcome to "Visions: A Machine Vision and Automation Solutions Podcast." I'm your host, Jim Tatum, senior editor of Vision Systems design and Visions is an Endeavor Business Media production from your friends at Vision Systems Design. Here you'll find the latest on everything from end user machine vision solutions to trends, developments, and perspectives on all things machine vision and imaging. Whether you've been working in the industry for a while or you're just starting to take a closer look at it, this podcast is designed to grow your knowledge and bring greater focus to your understanding of the imaging and machine vision industry. And now on to our show.

Well, hi everyone, and welcome to Visions. Today we're going to take a closer look at a startup that leaned into AI pretty early on. And now they're reaping the rewards for making those bets. Overview AI is a company that was founded on a simple idea: bring the kind of AI powered perception that helped self-driving cars understand the world into the factory floor. With two founders from Tesla and two from outside that company, the team saw firsthand how traditional machine vision often struggled with variability, set up complexity, and changing production conditions. Their answer was a new generation of AI driven defect detection package in an easy to deploy smart camera that can be trained in hours, not weeks. Our guest today is Christopher van Dyke, CEO and co-founder of Overview AI, who was kind enough to spend a few minutes with us to tell us about Overview's, origins, why it bet early on AI inspection, and how it's helping manufacturers and integrators make defect detection faster, simpler, and more reliable. Thank you, Chris, for being with us today. First, if you would, please give us a little bit about your backstory. My understanding was everybody came from Tesla and formed this company. Is that right?

Uh, half of the founders, two founders from Tesla and two founders from outside Tesla.

And what kind of pushed all together to do this? Yeah. So speaking for myself, I was there at Tesla. I was working in automation. Tesla had very aggressive plans to automate the battery lines and the EV, the model three EV, as quickly as possible and with as much robotics as possible. And it was just hard. It's hard to do it. And you learned a lot through that. And vision seemed like a very useful tool in the kind of needs of a factory. Tesla was also working on the self-driving car technology for the vehicle, and their cameras were receiving the road, and they were able to figure out lots of information about the road. That same technology was not available in the factory, even though Tesla, we had really big buying power. And so it seemed like there was an opportunity to take this new technology and put it into the manufacturing space. I really liked manufacturing and felt like it was a very rewarding and exciting industry, but one that actually didn't get as much attention from the tech base as some others. And so, um, that was kind of a common theme. Again, I'm speaking for myself, but I think all the founders had some of these same realizations. My other founder from Tesla was deep in the manufacturing development work with me. And the software engineers that we linked with also really liked manufacturing. So, um, we were kind of really drawn into that, um, trying to bring the new technology into the industrial space.

Well, to go back a little bit, I know we kind of had the flyby view of what was going on at Automate, but just if you can get a little bit more into the products and the solutions, especially the ones y'all were doing there, kind of get get into that a little bit. Tell us a bit more.

Overview is really focused on defect detection. And we often say AI defect detection. So that's where the defect has some variability associated with it. Or potentially the manufacturing um parts and environment have variability. But in general the AI techniques are much better at dealing with disruption and change than the older vision systems were. So vision systems have been powerful for a while, but they often can be very hard to set up and, and finicky and get kind of confused. The newer generation of technology solves a lot of that. It's much more judgment based, has a much more ability to deal with uncertainty and change. And so then we've really leaned into, you know, trying to get good, make products for catching defects, um, maybe just assembly verification and any kind of visual confirmation of what's being built, but a lot of the ability to look at something that's being built over and over again and find the errors, we sell it in the smart camera form. So that means get a piece of hardware that has the camera, the computer, and all the software. At one in one, you buy the device and it's quite easy to set up. Within a couple hours, you can have it set up and accurately spotting defects and communicating with other factory computers, you know, PLCs or data systems. And so that that's really like our driving force is trying to make defect detection really, really easy and really accurate and trying to make it very easy to get the device integrated into your line.

Okay. Is it all, um, real time data or are you using a lot of visual twins, that kind of thing.

All the computation is done on the computer and it's which is right on the camera. Um, and it's based on, you know, the training. So what it's seen from the configuration setup that the user shows it some examples and gives it a kind of guidance as to what's good and bad. It also has a kind of pre-built brain that's tuned for manufacturing defects. Um, but yeah, it sees a picture, makes a judgment and spits it back out and then logs all the results and stuff. But it's a real time judgment. That's the key.

Okay. Um, well, since AI does require a lot of data, why does defect data seem to be so hard to collect at times?

Yeah, that's a problem. AI requires data for sure. And, you know, we, we work really hard to get that to be as low as possible. And, you know, we can get really accurate results with maybe five to ten. Examples of defects and good parts are often easier to find. But that is a kind of key part is trying to get it low. So I think it used to be people might need hundreds of examples and it's quite a bit lower than that. But you still need good data. Um, it can be hard with defects because sometimes defects are very rare. Um, sometimes there's confusion on your own team about what's good or bad. That can be really detrimental. If somebody comes in and, and says something's good and you train the model that way, and then somebody else thinks it's bad, then you know, you're just not going to get good results. So when data is confusing or when you have kind of outlying corner cases, then it can be a little harder. Uh, we have a few ways around that. I can get into those if you want. Uh, we without giving away your secrets, but yeah, like, sure, we have a synthetic data generating tool. So that's a way to build up your, when you're doing the configuration step, the training step. You can create some defects through a tool that is like Photoshop or something, where you're just sort of making the defects up through, through. You do it through prompting and explaining natural language, and then you end up with extra defects. So that can be a way to kind of manifest unusual defects. We also log all the data that the camera sees up to twenty thousand images. And then we have a tool that shifts through those sifts, through those images, and can flag ones that are unusual. So you might miss it the first time, but we can kind of find those unusual defects, um, surface them for you. So you don't have to be looking, you know, constantly observing your system or looking through hundreds of images or thousands of images. We pull them out and then you can put them in your data set. So for very high accuracy use cases, you do have some of this more iterative process to kind of really tune the accuracy, but a lot of use cases that you don't need to do that.

Okay. To backtrack just a little bit, you guys were pretty much leaning into the AI early on. So what kind of pushed you that direction and what are kind of really the pros and cons of it? I mean, but yeah, between traditional machine vision, you know?

Yeah. AI is a big field and it's got a lot of different techniques that are kind of called AI. It obviously became very prevalent in everyone's mind when like ChatGPT came out and the big language models added this huge new capability. We were sort of the AI was still kind of an exciting thing to be involved in a few years before that. That was the CNN models and the kind of there was a whole set of development and compute was coming down quite a bit. Nvidia was still driving down the cost of GPUs, and there were new techniques to do these judgments that that was really driving the the automotive focus on self-driving cars, autonomous driving. Um, there were still there was other, um, visual like security companies and there was a lot of focus on vision in the kind of twenty fifteen to twenty twenty era, even before the kind of more the recent explosion of AI. So we weren't like way out on a limb by being involved in AI. It was, it was definitely like good, you know, interesting stuff. Where we were more unique was like within manufacturing and manufacturing defect detection, there were not nearly as many AI tools. So big companies hadn't done AI tools yet. And there were maybe some other startups starting around the same time as us. But there weren't. It was it was definitely like a early time to be plugging those two things together.

Yeah. Kind of build your presence. And now anybody coming into it is kind of catching up.

Yeah. Yeah. That's true. And then all the customers are much more sort of familiar in some ways. That's good. Some ways it's bad, but the customers know a lot more about AI. They're they're more interested. There's a lot of people who sort of have an abstract idea that they should be using it more. Um, so that, that, that's been helpful too.

Well, with that in mind, are there instances where you might tell somebody, look, this really isn't for you?

Yeah. We really try hard actually, to make sure you don't want to sell anyone something if they're not going to have a good outcome with it. Um, where maybe the most common pushback is like, if you have high mix, low volume, if you're not really automating other parts of the manufacturing process, then automating the vision usually is challenging. There's some new techniques that kind of span parts. And if there's common features, but, you know, some mix that the AI can actually kind of handle that. But that's one where there's setup time and we want to make sure somebody knows how much it takes to set up and how much you get out of it. And if, if you're not making that many of something, you know, sometimes the setup time is more expensive than the savings from the operational savings. The other would maybe be, uh, but one thing high speed, we max out at about three seconds. So there's kind of chunk of we have a newer tool that can go faster. But, um, that's something where we're often working with customers to make sure we can do it at the rate they need. Uh, and then yeah. Um. Those are the biggest ones that the kind of, you know, is it, is it high enough volume that you should be automating it? Um, I guess if it's not visible to the, um, you know, we do 2D visual inspection. So if something is occluded or it's an unusual, like somehow you can't see it from a certain angle or at the surface, then you're going to need another technique.

Uh, what's an example of that?

Well, certainly like something where there might be an internal crack you'd have to use like an X-ray or some kind of an internal thing. But sometimes it's more actually just like the defect can be underneath the other component, and a human can easily lift it up and look around, whereas a camera just has one angle, or you have to put in a bunch of cameras so that, um, if you're kind of just saying, oh, I've got a lot of inspectors I want to start to automate, it does kind of matter what the inspectors are doing.

Okay. If I'm an integrator, then what do I need to know before I actually get into adopting something like this? Especially when you're dealing with, say, synthetic data or something like that, what are what what overarching piece of advice you would give?

Yeah. One thing I think for integrators is some of these technologies are kind of new and they do work really well. And I think that if you're kind of reference is vision from twenty twenty or before, you're going to be very skeptical of some of the new techniques, but they're worth trying. They also are pretty easy to try. Our tool in particular, you can just spend a few hours and you get a really good sense of whether or not it's going to work for something. So sometimes integrators don't have that time. They need to just pick their product. But you can end up saving a lot of time if you pick the right thing. We've really strived to make it easy to use. A lot of our early customers were end users, OEMs, but that is also, you know, it makes it it's important that it's very easy to integrate it into existing lines. So we've set up a lot of easy comms and easy interface points. So yeah, I think it's kind of similar to when we warn people not to use it. If an integrator is working on a moderately high volume, five thousand a year kind of build, you know, automating, um, this kind of inspection is definitely cost effective and usually really adds to the capabilities. Again, the kind of Defects that are variable scratches, um, welds, laser welds, solder joints, things where glue, where the defect can be kind of anywhere. Those used to be really, really annoying to do and now are quite easy to do. So I think probably knowing it's easier than maybe you anticipate integrators in particular, who have a long history with automation, I think are even more skeptical maybe than some in-house engineers of new technology. Um, then pick somebody who's got a good presence already integrators maybe have less sort of forgiveness. If something doesn't work, it's totally on them to fix it. So they do have to be conservative. So when picking a technology, make sure you pick one where there's a proven track record. But you know, for AI inspection that exists now, do you lose any effectiveness at all with the, uh, trading offer, the ability to set up and go so fast. Um, you can always get something more accurate if you spend more time and really get your data set right, collect more data, that kind of care and labeling it. But, but sometimes those are very marginal gains and that you can get to the, you know, very high level of performance that you need to quickly. And, um, it just kind of depends on the customer and the use case. We have some high performance customers where they're what laser welds need to be, you know, it's not just about having a connection, but it's about having the same impedance through the joint. So they need to really inspect the weld very cleanly and get it exactly right. And so they've pushed really high on accuracy. Other use cases, it's maybe just the connector needs to be fully Connected, and it's quite easy to tell if it's even if it's partially connected or disconnected. Uh, you know, there's, there's enough data, so it kind of depends on the use case, the failure mode as to, you know, how much time you need to spend on the accuracy. So the going fast will be not as good performance as being incredibly careful, but we've set it up so you can go fast and, and still get really great results on a lot of use cases.

Yeah. Okay. Well, um, any other thoughts or reflections you might add to that?

Um, no, I mean, I kind of hit the point earlier, but I think there's a little less of this now that AI has become popular. But we've been doing this for several years. And I always want to encourage, we actually like working with customers that know vision well, have a deep history doing automation. Our tool is meant for just a generalist manufacturing engineer. You know, you don't need to be a vision expert, but vision experts often are, you know, quite good at setting systems up. So, so we're excited to work with them. But then they're often the ones that are, you know, the most skeptical of, of sort of newer techniques and newer technology. And so I just encourage people to give it a try.

Well, that's a wrap for this episode of Visions, produced by Endeavor Business Media, a division of Endeavor B2B. Thanks very much for tuning in. If you enjoyed today's show, be sure to subscribe to the podcast and share this episode with a colleague who would find it helpful. Until our next episode, you can find us at vision dash systems dot com or on LinkedIn, Facebook, or for more insights, updates, and breaking news to keep you in the know. Thanks for tuning in. Until next time, stay focused on your visions.