In The Arena by TechArena - Data Insights: A.J. Camber on Computer Vision & AI Scale

Episode Date: July 14, 2026

In this episode of Data Insights, Allyson Klein and Jeniece Wnorowski speak with A.J. Camber about how AI is evolving from specialized tools into more accessible systems that can be used by a broader range of professionals.

Welcome to Tech Arena, featuring authentic discussions between tech's leading innovators and our host, Alison Klein. Now, let's step into the arena. Welcome to the arena. My name is Allison Klein. This is a Data Insights episode, which means I'm here with Janice Norowski. Hey, Janice, how are you doing? Hi, Allison. I'm doing well. How are you? I'm really fantastic. I know that we have a really cool topic today, and it's all about solid I am for a change. So why don't you, tell us what we're talking about and who you brought with you. Not always about Saladine, that's right. But when it comes to AI and it comes to infrastructure and now software,

I'm excited that I brought a fabulous guest. Today I have AJ Tamber with me. Welcome to the program, AJ. Hi, Janice. Thanks for having me. Of course. And AJ is our resident VPGM of AI software at Saladime. So we're excited to talk a little bit about a new. technology that was just launched, oh gosh, AJ, what, T-minus three weeks ago, a program called Lucetta.

So we're going to deep dive in today and talk about how you can use this new technology and what a game changer it is. So thank you for joining us. Yeah, thanks for having me. To start off, could you provide a brief background on your role at Solidine? Yeah, thanks for asking that. So again, I currently lead the AI software group for Solidine. So most recently at Solidime, the last two years, I actually led our strategy team. And in our strategy office, we were focused on long-term growth.

And this was actually one of the incubation projects that we had. And so in December, we decided to make it its own business unit based on the progress that we had made. So, AJ, from your vantage point, what's changing in how most organizations are actually putting AI to work? What's your thought on that? Yeah, you know, from Solidine, we are. a leader in data storage, and specifically we have some leadership QLC drives that give us a different perspective maybe than what you might have in the normal industry. And what we saw there is this kind of explosion of computer vision data specifically. And so the data sizes specifically related

to cameras are huge, right? Cameras can generate terabytes a video. And that's just like with a single inspection point. And specifically as we go into physical AI and other areas, we see this area of computer vision just really is the place for innovation right now. I've been thinking that a core theme in your work is making AI significantly easier to consume. Why does accessibility become such a critical barrier in enterprise AI adoption? Yeah, you hit on, Allison. Thanks for bringing that up. So with Lucetta, again, we're focused on democratizing AI for everyone. And the reason why we do that is because they're this scarcity of data science, right?

So most of the tools out there, they're built for developers or even computer scientists that maybe aren't data scientists yet. And there aren't enough people that know how to use these tools to effectively use them. And it's just not easy to play with AI. And so what we thought is that the engineers working in the factory environments or working in the warehouse, these are the ones that understand the defect. or the quality inspection items that you're looking for. And this is like an underserved community, right?

So even though data science jobs are growing, right? A few years ago is a quarter million in the world. And I think the latest estimate for 26 was maybe a million. And the demand for those data scientists over 11 million in 26. And to make matters worse, all of those data scientists are the vast majority are concentrated in like 20 companies. And so the probability that you're going to be able to have enough of this talent kind of in your company just as a manufacturer is probably difficult.

So we thought that we needed to help out and do something, try to figure out a way to get more people involved. I love that. And on the surface, AJ, training models on your own data sounds pretty straightforward. But as you and I have discussed and pointed out, there's a lot happening under the hood. So where do organizations usually and typically underestimate the complexity of running these models? Yeah, good question.

So all of these models, they use data, right? So generating the quality data sets, usually the first area of focus. That's another thing that our Lucida product does that helps you get to the point where you can click data very quickly. Another nuance of this is that a lot of times we hear a lot about pilots are easy to generate. It's quick to get some results that maybe work under common or perfect conditions, but we kind of all know that that's not really real life, right? So there's a lot of balance that you need in the data set. So you need a certain amount of diversification to kind of compensate for all the complexities that kind of real life brings, right?

But you don't always just want more data, right? More data is not always better because it can cause like bias in the models or overfitting where you unintentionally miss like outliers, for example. Or it could be like low quality or repetitive data that's not useful. And if you generally add more data to models, which is conventional wisdom, right? The models actually get slower and more expensive to run and it's bad for the environment. So there's a lot of bad things. It's even more difficult to troubleshoot, actually. So the trick is just getting this just enough kind of mentality. And to do that, you have to be able to identify the right amount of data. And so most of the time when we think about Lucetta, what we're really doing is we're just trying to help people sort through massive amounts of data and find the most useful data so that they can generate these models the way they want. You referenced the idea of AI generating AI.

How is that shifting the way models are built, especially for teams that may not have deep machine learning expertise? Yeah. Yeah. So like I just said, the software basically automates a lot of steps for you. So for example, Alison, that was a good question. I just talked about data diversification. So maybe let's talk about that a little bit more. So we would estimate that if you did have a data scientist, that someone skilled in the art to do this,

would take maybe about 20 hours if you're roughly making one model a month on a fairly large data set. And so again, if you think about what you're paying in terms of a data scientist's salary, 20 hours out of months, that's roughly an eighth of a month, right? And if you're paying anywhere from 500 to 750 bucks an hour, that's a lot of money, right? $10,000. So with us, we automate that. So again, with our tool, we aim to kind of democratize and enable industrial engineers and mechanical engineers to do these sorts of things. And in that data diversification example, I just mentioned, it's zero for us because we do that in the background for you.

That's one of the advantages. So AJ, Saladine's Lyssena AI platform is really focused on enabling vision-based use cases, right? What is it specifically about visual data computer vision that makes this area particularly impactful right now? Yeah, I think the world's kind of headed to all of this robotics and really physically. And so all of that starts with cameras. And, you know, for us, visual inspection is really just one of the first, most obvious use cases, really in manufacturing rights, where we see it first.

But there's a lot of other use cases too, right? There's like looking for personal protective equipment, like in safety and compliance application or in warehouse environments, looking, doing county applications to do reconciliation to orders. But there's tons of quality inspection things too, like everything from textiles or wealth quality to orange peel and paint, the applications are kind of endless. And so from our point of view, one of the reasons why it's really interesting is because cameras, again, are just more consistent than the people doing the job. They don't need to take breaks. They don't get tired.

And so you have an opportunity really to take advantage of that by employing kind of these computer vision applications. When you think about making these capabilities usable across different technical skills. What does a well-designed AI platform actually need to get right? Yeah, that's a really good question. This to us philosophically is something that we believe where we should allow people familiar with the domain to be transparently guided by a tool that can iterate and can iteratively improve the model output. So we don't think that our customers will be as successful as they can be if they're just using like pre-trained models, for example. So we want you to use your own data and to iterate on that tool so that not only

is it tailored to your use case, but it builds trust with the people using it. And we see that that trust of the tools from the people working there is extremely important to get adoption going. And so again, that's a little bit of a philosophical difference. And it has to do with the way we've put together kind of RIA pipeline in order to converge models very quickly on a very small amount of data. Yeah, AJ, there's always a trade-off between simplifying the experience and kind of maintaining transparency overall. But how do you think about balancing abstraction with the need for control and trust? Yeah, we think that the best approach actually is one that assists the people working on the line again. We're not trying to replace anybody, right? It turns out the models actually

work better when they get the benefit of the judgment from the people or human counterpart. if you will, and by allowing the person to see, you know, in a systematic way, this kind of collaborative improvement towards key APIs, we actually get better results than you might have otherwise gotten if you tried it with maybe a different approach a few years ago. For organizations applying AI to real-world operational environments, what separates projects that scale successfully from those that never move beyond pilots? I think in this case, I would say that it comes back really to. that quality and variation in the data set. So like I was seeing earlier, there's a lot of complexity

in the real world. And one of the implications of that is that a model that's perfect today is really not perfect tomorrow. And we should be really encouraging teams to iterate with their own data, make sure that they have enough of that variation to be successful because perfect conditions are not going to get you the results you want. And this flexibility, too, is a little different, right? So traditional optical inspections are very rigid, the kind that aren't really AI-based, but one of the benefits of using an AI-based approach here is that you achieve this flexibility, which enables the model to adapt. Right.

And so, again, like, not all scratches are identical and not every dent's going to be in the same part. And by using AI training techniques that allow for this adaptation, I think, again, you'll be more successful than if you weren't doing that. So as tools make it easier to build and deploy models, how does that change the role of domain experts versus, say, traditional science teams? Yeah, I think it just means that if you have a tool like Lucetta, you are less dependent on just a few individuals like I was mentioning earlier, which means you should be able to get more done, right? But by the way, it's great to have a data scientist be part of the team and leverage their expertise on where to go forward. but if you have a tool that can automate some of the more rudimentary aspects of the job, then their efficiency can be better to you.

So I think our goal is to improve the efficiency by raising the level of contribution from everyone. Now, looking ahead, how do you see enterprise AI software evolving as expectations shift towards Saster deployment, broader accessibility, and even more measurable outcomes? Yeah, you know, we're still at the beginning of this, but I would say that I could simplify it down to just a few things. I mean, for our customers, our application expectations are pretty straightforward. They're just looking to save time. And it's easy to dollarize what that time's worth, whether it's a data scientist or mechanical engineer or people annotating specific amounts of data. Or in the quality inspection space, what they're looking for is actually improve their yields, right? So just throw away

less product, make sure that quality defects are caught as early as possible before they leave the factory, which is also easy to dollarize. So again, our customers definitely are looking for ROI. And I think most cases boil down those two items, or at least that's what we've seen so far. This episode was fantastic. I know that people are going to be interested in talking to you more. So for listeners who want to learn about Solidimes AI software

and the work you're doing, should they go to engage your team? Yeah, the easiest way is probably just to reach out to me or send us an email. So email address is Industrial AISW. at solidime.com or you can also follow the link I think that we just dropped in the chat here at Solanine. Thanks for joining Tech Arena.

Subscribe and engage at our website, Techorina.ai. All content is copyright by Tech Arena.