FIR 135: Interview - Can AI See Better Than Humans ??

Financial Investing Radio

In this episode we look at can AI help me see better in a cost effective way!

Grant Everybody, welcome to another episode of click AI radio. Okay, I have in the house today with me, someone I've been very excited to talk with. He and his organization reached out to me and I was quite surprised when I saw the cool AI solution that they have been bringing to the market. And Carlos has been giving me a little background on this. And I think you'll be excited to hear what it is he's putting together. But first and foremost, welcome, Carlos Anchia. You got Yeah. All right. There you go. Carlos, please, welcome and introduce yourself.

Carlos Hey, Grant. Thanks a lot for having us on. Like you said, my name is Carlos Anchia. I'm the CEO of Plainsight AI. And we're bringing to market an end to end computer vision AI platform. I'm really, really happy to be here love talking about AI, computer vision, and how we can get more people to use it.

Grant So okay, so tell me a little bit about what got you going down here. As you and I were just chatting a moment ago, there's so many components to AI, or it's such a broad range of technologies there. What got you thinking about the CV or the computer vision space? What problem? What How did you get started there?

Carlos Yeah, that's a really good question. So like you said, AI, the breadth of AI is huge, you have deep learning, you have machine learning, forecasting, prediction, computer vision. And these are just a few. There's a lot of different applications for AI and places you can go down and succeed in. From our respect, we really, we really focus in on computer vision, specifically how to apply that to imagery and video. Today, there's a huge amounts of data going throughout the internet and in enterprise storage classes, where you can't really extract the value of that data unless you actually perform some sort of computer vision machine learning on that type of data. So we're really extracting the value of the picture or the video. So it can be understood by machines. So think of a dog and a cat in a in a picture, right? Those cases, the machine doesn't know it's a dog and a cat, you have to train it. And that's where computer vision comes in. And really, we got into it because we were pulled in by customers, customers of ours wanted to start doing more computer vision and leveraging our platform that we had around high throughput, ingestion, and event driven pipelines. So these customers came to us and hey, you know, this is great, we'd love to really use this for computer vision. And the more and more that kept happening, we kept retooling around the platform. And finally, the platform from end to end is purpose built to do computer vision technology. And it really allows us to focus in on on what we're good at today. Right? And that's really delivering value within the computer vision space.

Grant So I remember the first time I wrote some of the OpenCV framework code, right. And that was my first introduced introduction to it. This is a number of years ago. And I started thinking, Oh, this is so cool. So I'm writing all this Python code, right, building this stuff out. And then I'm thinking, how many people you know, are actually leveraging this platform and look at even though open CV is cool, and it's got a lot of capability, it still takes a lot, you know, to get everything out of there. So can you talk about how you relate to that open CV? And what is it that you're doing relative to that? And how much easier do you guys make this?

Carlos Yeah, so I mean, you hit the nail on the head there, right? So from a developer perspective, it's really around, I need to learn open CV, I need to learn Python, I need to learn containerization I need to learn deployments. There's a variety of different companies that, you know, they're all great in their own right, right. Every one of those companies that we just talked about organizations are contributing tremendously to AI. But from a developer's perspective, you really have to learn a little bit of everything to be able to orchestrate a solution. And finally, when you get to, hey, I use AI. Let's pretend we're looking at strawberries. Hey, look, I built a model that the Texas strawberry that is your over the moon excited, but the very next thing is around, okay, how do I take that and deploy it 1000 times over in a field across the world and understand how to make that in an operational fashion where you know, it can be supported, maintained update, and that's really where we have this this crux of an organization where it's really different building something on a on a desk for a one time use. And and there's a lot of wins through that process. But then taking that and operationalizing for business driving revenues driving corporate goals around, why would that feature is being implemented, that's really where we come in, we want to be able to take off that that single single path of workflow where it's a little bit of everything to orchestrate a solution, and provide a centralized place where other people, including developers can go and help build that workflow in a meaningful way where it's complete.

Grant So operationalizing, those models, I find, that's one of the biggest, or the most challenging aspects to this, it's one thing as you know, to, to build out enough to sort of prove something out and get some of the initial feedback, but to actually get it into production. I think I saw MIT not long ago, maybe this a year ago, now, they had come out with this report, it was through the Boston Consulting Group as well, they'd mentioned something about, hey, you know, 10% of organizations doing AI are getting return on their investment. And, and, of course, when you look at all of the investment of the takes for the business to really stand up all the data scientists and all the ML work. And you can see why the numbers translate that way. So to me, it feels like not only doing this in the area of CV, but the problem you're really trying to solve it feels like is you're attacking that ROI problem, which is you could take this kind of capability into business say you don't have to stand up all of these deep technical capabilities. Rather, you can achieve ROI sooner than rather than laters. Is that Is that accurate?

Carlos That's correct. And I think it's really through the adoption of technology and you hit a you hit a really strong point for us there around the the difference between it works and it's operational. That's that's really the path of your your, you're less there in the CV world and more there with the DevOps ml ops portion of it, getting machines running consistently, with the right versions of deployment strategy, that latter half of it, just as important as the model building pieces of it. But even after you get to that piece, you need a way to improve, and improvement in the model is very costly if it's not automated. Because I mean, you can just look at the the loss for a simple detector, like a strawberry, where, you know, if the model starts to perform poorly, you're not pulling as many strawberries out of the field. So you need a way to be able to update that model, quickly, get training data into a platform and push a new model back out. And it's really around how fast you can go end to end with that workflow. Again, and again. And again. And again, this is that continuous improvement that we have born into us from previous software development life, but really in in machine learning and computer vision, your ability to train, retrain, and redeploy. That is where you really get the benefit out of your workflow.

Grant Well, that really confirms my experience with AI will. Typically I'll refer to the term that we've been using, as we call it a SmartStep. It's that notion that I need to be able to refactor my models and take in consideration that changed context around me, whether it comes in from the world or from the customer, or whatever that means, some level of adjustments taken place that begins to invalidate my previous AI model. And I need to be able to quickly make those adjustments. That's fascinating. How long typically does it take for you to do that kind of refactoring of your models? Is that Is it a day? Is it a year, a month? Or the answer is? Well, it depends.

Carlos So it's twofold, right? So it's, it's hours to do that. But it really depends on the complexity of the model, and how long you have to train. But in an automated workflow, you're you're continuously adding data to your training set, that are lower quality predictions, where you can retrain automatically when you hit a certain threshold, and then validate the model and push that back out into your production alized environment. So it's it when you go to develop these sort of workflows. You really have to start with whatever I build, I know I have to improve on later. So that improvement cycle ends up costing a lot if it's not part of the initial discussion around how do we count strawberries, right? So it's evident you and I can nerd out on this. Let me shift the focus a little bit and ask about it from say your your customer Write from their perspective, what is it that they need to be able to do to be successful with your solution? What skills or capabilities do they have to bring to the table?

Yeah, and I think it's, I think I have this conversation a lot with our clients. And it's really less about them having technology around data science and building model, and more around a collaborative environment, where organizations, you know, they have a culture of success. But that culture of success is really borne by holding hands through the fire, it's, it's being able to commit and lean in when the organization sees something that's really important to them, either either from a technical perspective or revenue perspective. And it's these companies and these types of people that get to rally around a centralized platform where they can build and collaborate with machine learning computer vision applications. And, you know, it's it's a, it's really interesting to see the companies that succeed here, because it's really based on a culture of winning, right, where the wind doesn't have to be the hardest, most technical, logically difficult problem, because complexity really drives timelines. And if you're looking to change from an organization's perspective, start getting the little wins, get the little wins, start having some adoption within the company around, wow, computer vision is working. We've identified these problems in a few hours, we have a solution deployed, you start building this sense of confidence in the organization where you can take on those larger tasks. But you have to start with a build up, you can't just go right to the highest ROI problem. No one starts at human genome sequencing.

Grant Have you, or do we got a problem? Yeah. Back up, back up. So So all right. So it means to me it sounds like as an organization to succeed with this getting my problem definition, understood or crisply put together first, what would be an obvious thing to do? But how long does it take for me to iterate? Before I know that I've got value, that I've pursued the right level of the problem? You You made an interesting comment a minute ago, you're like, oh, within a couple hours, I could potentially retrain the model and have that back operationally. That means if I can fail fast, right, if I can pick my problem space, get something out there operation, try it fail fast, and then continue to iterate with AI as my helper that that's really, really quite powerful is that the model that the your person?

Carlos It is the model? That's exactly right. And it's not just hours to retrain, it takes hours to start, right. And just to highlight you kind of started with, we have to define that problem set first. So even after we define that problem set, a lot of times we have to go back and redefine that problem set, and really the piece around failing fast. It's it's experimentation. And do we have the right cameras? Do we have the right vantage point is the model correct, you want to be able to cycle as fast as you can through that experimentation phase. And sometimes you have to go back and redefine that problem set. Because you're learning more as you go, right? And you're evolving into okay, I now understand the corner cases a bit better. And with the platform, you really can cycle that quickly. I mean, machine learning at scale is really how fast you can iterate through improvement.

Grant It's quite, I think it's quite a testament to how the AI just world in general is improving. I know that you and I were talking earlier, years ago, you know, when I first started writing some TensorFlow code and Keras code, you know, the, the time it took to fail was much longer, right. And then the cycles were huge and, and getting this down to a matter of hours or even a few days, you know, for an enterprise's is massive. What's the, from what you've seen terms of different industries? Are there certain industries that tend to be leaning into this and adopting it or the is there no pattern yet?

Carlos No, there's a there's a definite pattern. And in 2022, we'll all see kind of what that what that's looking like. And it's really an industry that has traditionally not been able to go through that digital transformation. So think of think of think of a piece that's a very manual piece, right, like physical inspection, where humans would look at something, they'd write down their notes on a piece of paper, then that that item would go through either pass or fail or some criteria for rework. That's all possible. Now with computer vision years ago, that was impossible the accuracy wasn't high enough, plain and simple. It just took it wasn't, it wasn't better than a human. Now we have models that are better than a human for visual inspection. And these industries are digitizing their workflow. So it's not only the feature of computer vision, but it's also now I have a digital record of all the transactions, I have extracted video information, it makes auditing easy for those sectors that have a lot of regulatory compliance, those that require proactive compliance to audit requirements, as well as visibility. Visibility has always been an it's funny when we talk about visibility, but it's computer vision, but like a lot of human processes, that there's zero visibility in it there, it's really difficult to audit, you know, why is this working better or not? So having that, that digitization of the flow with the feature of computer vision allows us to extract the value. And industries like agriculture are going I mean, agriculture has been a leader in technology for a while, but now you're really seeing adoption at livestock and row crop with drone technology. It's a very rich image environment, medical space, medical space for computer vision in 2022 to 2028, is estimated to be billions of dollars just with medical imaging. And that's not that's not the the total addressable market for the hardware, it's just the imaging piece. So we see we see a lot of growth in sectors that are going through this digital transformation that are adopting technologies that are now getting to the point where they can get pushed down into the masses instead of just the top five companies in the world.

Grant Excellent, it seems to you part of your comment earlier made me think about process optimization for organizations, and the ability to extract processes, you're familiar with process mining, right? The ability to extract, you know, out of logs of these organizations and doing something like that where you can produce this visual representation of that, and then building models against that, to optimize your process might be an interesting use case. Yeah, that's fascinating.

Carlos That's a really good point, right? Because that's, that's a different portion of AI that can be applied to like just log analysis, that then would allow you to go back and Okay, now that we have the process mind, where can we improve along the process?

Grant Yeah, yeah. Amazing. So many uses and use cases around around this CV area, for sure. So let's say that someone listening to this wanted to learn more about it, where would they go? How would they? How would they find more about your organization?

Carlos You can find us everywhere, right? We have a website, plainsight.ai. We're all over LinkedIn, we have Twitter, we're on Reddit, we have a Medium blog, there's a Slack channel where we geek out around computer vision use cases and how we can improve the world through computer vision. We're really we're really out there and feel free if you have questions come reach out to us. We have amazing staff that are looking to empower people in AI. So if it's through just just a question around how does this thing work, we'd love to talk to you if it's Hey, we're kind of stuck in our journey. We need some help reach out to us we can help you.

Grant That's awesome. Carlos I can't thank you enough for reaching out to me and for a listening to click AI radio, but also for reaching out and sharing what it is you are you and your organization are bringing to the market think you're solving some awesome problems.

Carlos Thanks a lot, Grant. Appreciate it. always appreciated talking about computer vision and AI and thank you to you and your listeners and really appreciate what you're doing to the AI space.

Grant Alright, thanks again, Carlos. And again, everybody. Thanks for joining and until next time, go get some computer vision from Plainsight.