Unaligned #12: Industrial Vision Recognition with Always AI Founder Marty Beard

Who are you?

Who am I? I am Marty Beard. I'm the co founder and CEO of Always AI. We're a Vision AI company and my background is heavy in enterprise software and mobility. Many years with companies like Oracle and Sybase, SAP, BlackBerry, et cetera. But currently in the Vision AI space.

Very cool. And you're right on point for unaligned because Grant and I, when we started this company, we wanted to get into businesses and help business people use AI. And vision's a big deal. Let's just start with the basics. Where is vision good to be used in a business?

Yeah, it's a great question. So I think just, let's just start high level. So vision, obviously the most powerful or one of the most powerful human senses, vision. And I think there's widespread understanding now that it's going to have a major impact on the tech stack. Sensors, cameras, drones, et cetera, we're all experiencing these. And with deep learning AI, those sensors are getting really smart. So the question is, where should I use them? We've been doing this for several years and it turns out it's a lot of really what I would call operationally oriented applications. Right. How do I get information in real time that helps me run my restaurant, run my manufacturing facility, my construction site, my mine, and we can get into all those. But a lot of it is just about productivity and trying to understand how to optimize things faster and smarter.

I'm happy to announce that I am now taking sponsors for unaligned. I'm proud to have AI Top Tools as our first official sponsor. One of the most important things for me running my podcast is staying on top of the industry, knowing all the players and what they can do. I made a complete list of companies in the AI industry on x.com but no one has as comprehensive a library of tools as AI Top Tools. They break down products by use case like productivity, media, chatbots, customer service, and more. A lot of companies are looking for the exact right tools for their needs and there's no better place to find it than AI Top Tools. Yeah, there's so many multimodal models coming out of the woodwork lately. I've seen about 30 of them so far and probably missing about 10.

So yeah, let's set up this because.

These multimodal models, they can chat. GPT has vision, right? So you can use your iPhone, camera, aim it at something, it'll tell you about it. You're doing a different kind of vision that's really aimed at manufacturer.

Yeah, yeah, this is well let's, let's describe it. So it's very different from generative AI and so forth. So computer vision is all about. You need to build an application, a computer vision application, which is very unique. Right. The AI models are all about detecting people and objects and what's happening in my environment. Right. This is not getting on a browser and asking a question to a chatbot. Right. And the chatbot's probably looking backwards in time and gathering all this information and then giving you an answer. Computer vision's like real time. It's stuff that's happening literally right now. So it's, you have to create the data set and then you have to create the model around that, and then you have to get it ready to go out on an actual device like a camera and that, that process. And we can walk through the complexities of that, but that process is pretty sophisticated. And. But when you get it going, it's in Chrome, right? Because it's. Oh, wow. This is not one of those black cameras sitting up on the ceiling of a Starbucks that is passively just recording. This is something that is literally giving you information, more intelligent information. How many people are in my store? What, what kind of people are in my store? What zones are they going into? What brands are they attracted to that kind of information that up to this point we've been using humans to try to intuit or figure it out in hindsight.

Yeah. And you just mentioned there's lots of different things vision capabilities could be used for. In a company I just visited, Lumafield, which makes a CAT scan and gives you a 3D data field of your product coming off the manufacturing line, do all sorts of quality control kind of things. Are you, your system used for that as well as customer watching in a store?

Yeah, let me, yeah, let me, let me give you some examples. It turned and it actually turns out some of the most sophisticated early adopters of vision AI, like edge AI, are industrial companies. And so we've done a lot of work in mining, for example. I would not have picked that at the beginning of the company. But what do mines try to do? They're trying to optimize the process. Right. They run very tight chips. They're trying to extract more yield out of what they're doing. There's a huge amount of demand coming from the tech industry and electric vehicles and everything to try to get those precious metals. So we can all wear our vision pros and so forth. But it's a really complex and a very expensive. So what can cameras do? Cameras can tell you how efficient that process actually is, where you might be making mistakes, where somebody might be not safe, and you want to catch something before something bad happens. So that's an example. But 180 degrees from that. We've done work with what are called fast casual restaurants, companies like Mendocino Farms, things like that, where they really want to optimize the. What, how many, how efficient are we location by location in literally making a sandwich or making an order from end to end? And we want to understand literally what's happening in each zone. And we want to know on average, how long does that take and why do we want to know that? Because store A is not performing store B. But we don't really know why, what's going on. So it allows you to get in there and diagnose and figure out. Those are just two examples. I can give you many more. But at the end of the day, it's cameras that are starting to do a lot of work for us.

Yeah, these, these multimodal models that are raining out of the sky. Chat GPT is probably the best one to talk about for consumers. It recognizes things like sandwiches, knives, forks, cutting board. The standard stuff. Yeah, but you're trying to use it in a new way in a business which might not have training done on the model, on the commercial models. Tell me what makes you different from the other multimodal models that are out there.

Yeah, look, the way this works is let's go back to the examples. For example, the restaurant example, I want to identify. I want to know how many people are coming in, when they're coming in, where they're going. And then I want to understand behind the counter how efficient is my process. Yeah, you're right. There are a lot of models out in, out there that can get you going. But really what's. And let's say that maybe you can get 50% of the way there where the model identifies spaces and identifies even bread and things like that. But then you're going to have to really optimize it specific to your business case and specific to your environment. And you might have various things you need to do. For example, you want to blur faces and protect privacy. You have strange lighting conditions that are impacting what the cameras can see. You've got more constrained space. There's a variety of things. That's where all the APIs that we bring to bear come into play. So we use all these APIs to really build an application that's optimized specifically for that store's use. Case. But you're right, you get started really quickly with some of the models that are already out there. But then you need a platform to help you really customize it for that long tail of needs that people have. Then you're not done right. So this is where it gets a lot harder than generative AI and some of the like large language model stuff. And now you have an application to identify people and determine how efficient is my sandwich making capability. You gotta get that thing packaged and running on a camera on the edge or maybe a hundred cameras so that now you're dealing in the world of physics. It's okay, I got an actual thing. I have to get this running on some metal out there that is waiting for this deep learning algorithm to inference on. So that's a whole process where we're really strong at that. So we're, we understand edge environments, we understand how to get AI working on the edge and we understand how to get that goldmine of data coming off and then present it to people in a really nice dashboard that helps them make decisions. It's a long answer, but those are some of the areas that you have to work through when you're doing vision AI and some of the places where we're really, we really shine.

Yeah. Do you split up your customers into people who are using it on a manufacturing line versus somebody customers who are using it in a retail store? It's like what you were just saying of watching workers or making sandwiches.

Yeah, good question. I think, look, if I had to pick two areas and each one's huge, but two areas where computer vision on the edge actually is really happening in a sophisticated way that's in the industrial space. So in that I'm going to include like manufacturing and warehousing and industries like mining. And it's all about I'm in these big, complex, dangerous spaces and I want to know what's going on so I can catch problems before they happen. Or I can identify anomalies, for example, on a piece of wood before it gets cut. Or I can see like how many buckets of ore am I actually getting out of this mine. I thought it was 10, but the camera's telling me only six. Right. So it's that type of stuff. So that's in the industrial space. Then the other one is in retail and that's where you get more into quick service restaurants. We've done work for example at Burger King and others, or you get into specialty retailers, but at the end of the day that's customer analytics. What people Call zone analytics. Where do you go? Think about the last time you went into a store. There are probably brands that you were interested in, but other areas you were ignoring. So the store wants to know, is everybody doing that? Why am I wasting space over here with brands that people aren't attracted to? So retail's the other segment. We do segment those two. There's common things happening in industrial and common things happening in retail. So we'll be able to service a lot more efficiently to those customers.

I was just thinking about visiting Disneyland at some point. Disneyland closes parts of the park when they're too crowded. Could a system like yours be watching flows of people buy it and know, okay, there's now enough people in that side of the park. Let's close that off and push people to the other part of the park? Right.

Yeah, you're absolutely right. That's a great example in a lot of venues. Think of sports venues. Yeah, people want to understand crowds. And mostly it's again, it's almost like manufacturing sandwiches efficiently. It's like I want to get people in and out of the stadium as efficiently as, as possible. Everybody's gathering by gate 14 when all these other gates are available to let people in. So, yeah, people are absolutely using cameras to do kind of density analysis, crowd density analysis. Cameras are really good at that. Just again, trying to improve the customer experience.

Yeah. I toured a museum in Washington, D.C. with the. The CEO of the Guy who runs the museum. And he knew the dwell times in each room. Right. He knew how long people hung out in one room. And that was. Yeah, that's the kind of thing you could also.

That is. Yeah, the work we did that I mentioned with Burger King was really interesting where it's literally what you were talking about. It was what they wanted to know is from the moment somebody walked in all the way to, they ordered, they waited, they got their bag, and they left. What was the dwell time? How long did it take for that process to happen? However, this is work we were doing in Europe. Totally need to anonymize that individual, protect their privacy, et cetera. So really, literally what happens is you just see a number walk in that's not identifiable as an individual, and then you can calculate exactly how many seconds it took. And why do they want to know that? They want to know how efficient are we in serving people and how many people are we getting through that process as quickly as possible? Again, cameras, it's harder than it sounds because the camera A has to identify human beings walked in. B, timestamp okay, they've walked into that zone. Now they're in this zone. Now they're in that zone. What's the amount of time for that specific individual? So you have to have zone analytics and timestamps and before and after views and so forth. But all this magic is. Is re. It's. Look, two years ago, you have cameras that are just. We're all used to seeing cameras. It's mostly for security. It's mostly what are called passive. They're just. Somebody looks at them once a week or once a month or something happened, maybe they look and they want to find out what happened. This is more active. This is active cameras. Right? These are cameras that are proactively identifying information and providing that information to managers. Different space. And again, deep learning is what allowed that to happen.

Yeah, yeah. I would assume that it's giving a database some data. Right. Customer five had dwell time in front of the milk of 45 seconds. Tell me about the infrastructure of business needs to put in place to really start modernizing itself and thinking about these new workloads.

Which is a great. Which is a great question. And again, this just for the audience. This is really different than ChatGPT, where we all know now that these massive models that are being constantly trained and updated. When you get into the world of computer vision, I'm trying to shrink down the intelligence onto something that can run out on the edge, right. Much smaller data sets, but in some ways super powerful data sets. So let's go backwards. So the way it works is there are cameras. And again, it could be two cameras or it could be 100. Those are really the sensors. They're collecting the visual information. The actual magic is going to be happening on what's called an edge device. And this is literally out in the location. So we all know about Nvidia. Nvidia has a, as an example, has a group of edge devices called Jetsons, right? So they're about yay big, right. And they can sit out in a restaurant in a manufacturing facility. That's where the inference, the AI inference is actually happening. That's where it's going. Is that a person? Is that a person? Are they in a zone? What's happening that's running on the edge device?

That is the data never needs to leave your store.

It does not. It's literally, it's 247 just happening on that device. So in some ways, this is much more secure because if that gets hacked. Okay, I exposed one element, but not the whole cloud. Right. This is one thing that's running out there in the real world. The other major difference is it's free. So the inference 24. 7, it's just. There's no charge for that. Whereas if you're doing the inference out of the cloud, then obviously if, let's say I'm using aws, that's cha ching, cha ching, cha ching. It's running. And we do have customers that want the inference happening in the cloud, but just to back up. So then you've got the Edge device, then the always AI platform where you made the application and you're actually managing and pushing out the application. That's. We built all that's on the cloud. Right. So our platform and the way you access it is on the cloud. But once you press the button and deploy it out on the edge, it's running out on the edge. But we have the capability to see all those devices update those devices, you know, make whatever changes you need to over time. So it's a long answer. Edge is really a different world because you're dealing with kind of smaller data sets that are really powerful running 24. 7. And that's the magic of Edge AI.

Yeah. And low latency. Right. And you don't have to worry about the Internet being up if your Internet goes away.

That's right.

For whatever reason your works. Let's talk about manufacturing rather than tracking humans going through something. Although that's fun too. I was just in Las Vegas. You start thinking about a casino and how many. Yeah, they have hundreds and hundreds of camera. Just for one. Correct.

Is.

Is that kind of business something that you're going after or do they have more specialized. They have already systems that are running and tracking people and doing the they or would they coming to you cameras. But we need to do a new workload that we haven't been able to do before.

Yeah. The thing about AI is you can always build a new model which drives a new application and people have ideas about, oh, if we can do that, can we do this? Can we do this? Yeah, no, absolutely. So we've had some really fascinating interactions with casinos where one example I'll give you is, okay, we have a priority set of customers. When that customer comes in, we want to give them a total 360 view of their experience. For example, these customers say, yes, it's okay that you track me on your cameras. And yes, I would like to know that Nobu is too busy right now, so I'm not going to go waste my time and stand in line or what have you. So it's giving that experience, like hey, now would be a good time to go to the pool or did you. So it's trying to give information to that customer. So they're getting a live fee of their experience. And that's AI. Those are computer vision applications and they need help. The other thing about AI, which I know is it's unlike other software cycles that we've done through this is really active and by that models constantly change. You need constant updates. You have constant new cameras, new things. So it's a very interactive, iterative process. It's not like you're ever done when you deploy an app. And so yeah, these casinos need help. Right. Because we're not experts in, we're experts in cameras and we've got really sophisticated infrastructure. But constantly deploying new apps and staying on top of that, we need some help.

I visited a pistachio farmer down in Fresno one time and he has a huge factory that processes pistachios. Let's say he wanted to put together a vision system. Could he watch a high quality quantity of pistachios going by the system with you? That would like.

Yeah.

The size of the pistachios maybe because 10 different kinds of sizes of pistachios and they come in the factory off of the tree and they're all different sizes. So he has to build a system to sort those. Right. That's hard for that or is it more specialized for everything?

Yeah, let's talk about like the way that you would do that and I don't want to make it sound like hey, it can do everything etc because.

I'm just trying to figure out is it good for one. Like humans move slower than going by on a conveyor belt factories.

That's true. Yeah, that's true. Yeah. So I think in this case what you would be looking for are anomalies. So you wouldn't train it to look at every single pistachio going through. That would be difficult. And how useful is that? But if you had a good data set for the anomaly and I don't know what it would be, I'm not a pistachio expert.

Rotting a pistachio. Yeah.

Or yeah, then cameras are really good at that, they're good at identifying that. So it would be constantly looking for that anomaly and trying to catch that out and on. Now the next problem is the speed. So you're saying, okay, this is going. These pistachios are going by really quickly. Okay, that now you're getting into. Now you probably need a more sophisticated camera and you're probably getting this into some more complexity there. We did have a situation of a. A rope, a cable. And again, this is back to mining. So you have these mining cages are like the size of a house and you've got these steel cables going down. You do not want to have a problem with the cable. Nope. Right. So can you. Can you see that? Yeah. You don't. This is not a good thing. So you want to identify any problem with. But it's going at, I don't know, 15 miles an hour or something like that. Very difficult to do. So people are working on things like this. But I don't want to make it sound like it's really easy.

No.

But yeah, it's.

Use cables to pull cars up the hill. Right. Cable cars. And Right. There's a machine that's running at a constant speed. You could have a camera watching the cable to make sure it's not fraying.

Correct.

Or being thinner. That kind of thing.

Yeah. We've done lumber. We work with a lumber manufacturer up in Oregon. And it doesn't run as fast as what you and I were just talking about, but these massive beams are coming in and they want to catch knots and anomalies before it hits the saw. So again, that's where you just. It's all about garbage and garbage out. You need really good data. If you have really good data and you can build a good model. If you have a good model, then you can get it running. It'll do what you want. Right. So you just have to collect enough information about what does an anomaly look like. And this is where Back to Generative AI is really super fascinating right now is if I can build. Help build a data set that helps me identify, but not only from real images of problems, but from synthetic data. Right. I'm literally just making up a data set from digital synthetic data. And we're getting closer to being able to do that, which is game changer and super interesting and helpful in our business for building those data sets.

That's cool. How else do a business person who's just about to buy to start an AI team and start to do some of these initiatives. Maybe they haven't done it before.

Yeah.

How do they compare? How do they pick you versus one of the other companies? Because there's a bunch of companies that are doing various computer visions.

I think the first question is, why would you do computer. And one thing, I just want to mention one thing that we've seen over the last year and a half is a couple years People got really interested in models, and then there was a massive investment in data scientists. So now corporate. Every company of any size has data scientists sitting at corporate, spending a lot of time on models. And I think what's happening now is people are realizing, okay, I invested in model creation, model analysis. It hasn't gotten anywhere. Right. It's, I need to get it out of this kind of corporate silo and get it out running and doing something. And this is where computer visions get easier because you're not dealing with such massive models. The cameras these days are like 70 bucks. Right? You can have cameras that do a really good job for maybe 70 bucks. The Edge device that I mentioned is that's okay, that's probably several hundred dollars. But now you're ready to do AI practically for real in your business. So I think people are starting to realize, oh, why did I spend so much time on this sophisticated model that I'm not really getting it out into my manufacturing floor? And that's a whole different set of skills that I don't have. So that's where always AI comes in. We're like, hey, we can practically get that model or something smaller and easier to work with working for you, literally, like in hours or a day to show you some information. And I think people are realizing, oh, okay, maybe that helps me. And I think the fact that we put in those years of work across that whole life cycle I was telling you about building a data set and training a model and everything involved, we're just really good at that and that's helped us. All right, so it's a long answer, but we're going from the world of data scientists just playing around with models. So now let's practically get stuff deployed. And that's a good trend for always AI.

Yeah. I just interviewed a guy, a company that's building robots for landscaping business, and they're using computers.

There you go. There you go.

See the edge of the graph and obstacles, like not hit the tree or whatever. And so they're building a whole new business out of computer vision, which is very different than what you're building. There's a lot of things you can do in computer vision, isn't there?

Yeah. And I think when you back up from a macro level and you think about all the labor constraints that people are feeling and just a lot of supply chain problems, and it's where cameras can practically. The common view is like, oh, creepy, and you're just tracking me. It's actually that's not as interesting to companies as Anomaly Detection on their steel or their wood or these kind of practical applications really are the ones that are getting the attention. So it's a cool space. And I do think the focus on ChatGPT and generative AI and so forth is let everybody experience the potential practicality of AI that's helped. And like I said earlier, this whole concept of synthetic data. Geeky Silicon Valley term. Right. But it's an interesting thing that's happening where we're taking unreal to then build something that works in real. And it's helpful.

When I made Apple II motherboards back in the 70s, my mom got a job to make a metal nice. And then later I worked on a manufacturing light at hp. Could it take a picture of something like a motherboard and make sure that every part is in the correct place at the right time? Is that the kind of thing that you could do?

Yeah, we just.

Is that better for a different.

Yeah, I'll give. No abs. I'll give you an example. It's not as small as what you're talking about, but very similar. So a supplier to car manufacturers, and they deal with panels, steel panels. But the steel panels also hold. Cars are basically just rolling computers. Right. So the steel panel on the inside has all these sophisticated instruments and attachments and electronic components. And it's exactly what you were talking about. They wanted cameras to look at it, but all they're looking for is something's not right. Something's not right. And if they can catch it. And before they literally have human beings looking at these things and doing a good job, but they miss things and the entire line has to stop back up. That's bad. That costs money. So again, it's what you're talking about, where a camera is looking at it, taking that picture, as you said, and then saying, ah, that doesn't look right. That one's bent. Something's wrong with it. So that was interesting. But yeah, computer vision, again, it's all about training the data. You have to select the right AI model architecture to do this type stuff. Computer vision is crazy.

I have a friend who built a ripcord, which is a robot company, to scan paper really fast. So he. You can put a foot of paper in there and go. But at the first step it takes, it flashes off four lights. And the computer vision system looks for shadows on the paper. That's a staple.

Oh, interesting. Right. So it's a state that's interesting.

It's just looking for staples or paper clips or stuff like that. Right. Stuff that's not Paper so that a robot can pull the staple out and then scan the.

Interesting. Okay, that makes sense. I mean that's, that's. I had not heard that.

Yeah, that would flash. Boom, boom. Right. You would so fast your eye can't see it. Right.

But the computer, we haven't.

Right.

And. And back to your comment about multimodal. That's multimodal. Right. So now you're using. Okay, I have a light source as well as the camera and what the camera's looking at. We're starting for me, multimodal. What that really means is vision is one mode and then another mode would be like sound or voice or text. And so an example in a mine, like in a mining environment, I've got cameras that are looking at things happening, but there's also large loud sounds. Some are good sounds, some are troubling sounds. Right. So can you merge those modes into one AI model that helps give even more information. So that's what multimodal means to me.

It's how we know if a machine starts sounding bad. Right. Because it's making a lot of racket. All of a sudden we know that needs to probably be taken a machine shop taken apart and fixed. Right. Jet airplanes.

Absolutely. Right.

They listen to the sound of the jets of the engines going around and if they start getting out of bounds, they go in for main repair.

Yeah, absolutely. Yeah, yeah. Humans were multi. We're the ultimate multimodal immersive machines. Right. And we're trying to find how far can we go to mimicking that. In some cases it's like a miserable failure. Like a 6 year old would go. That's a computer cord, not a snake. Right, Right. We're working through these things and trying to find what are the best sweet spots for machine vision. Right. Versus kind of humans. We have just incredible immersive capabilities. Right.

Computers are really good at counting people. So back to the original situation. Cloud controls of watching people's behaviors, of watching your own employees be. All right, how good are they at. Employee a makes sandwiches 30% faster than employee B. Why? Right. Figure out if employee can teach employee B to be faster.

Yeah. Or when in. And this is where it gets into that issue of, okay, so a construction site. So there's supposedly 18 people that worked on the site and they charge their time, but the camera said there were only 13 people. What. Why do we get charged for 18 people? So these type things are starting to emerge in certain environments. And then obviously you get into all kinds of issues of okay, how, how does management Deal with that. And obviously you want to know truth, but you also don't want to spook your entire employee base where they just feel like, okay, because people will put a towel over the camera or they'll do whatever they're going to do to try to stop that. But you just, you made me think that. Yeah. There are certain conversations we said where the company's interested in, we're getting billed for X, but X was not there. Right. Another example would be like, supplies were delivered, so we got charged for a certain number of wooden boards. The camera says, okay, there's six feet of boards. And then we got charged for 12ft of boards. So this, I think there's a certain amount of visual auditing going on and some of that makes sense. Right. It's okay. What's actually happening.

You can look for skimming behavior.

Right.

All sorts of things. You're giving me lots of good ideas of, like, how to look at my business and oh, a camera system might be good there.

Right? Right. Yeah. Or like contactless checkout. Right. We're all, we've all been, we're all getting experienced on. Okay, I want to be able to do that. But is there a more efficient way to do it besides you hold it up to a barcode and then you hold it up and then it says, okay, that's an avocado. And is there a faster, faster way of doing. There is, but it's. It's using some of the sophisticated computer vision.

When I talked to Volkswagen's team that does augmented reality and other things, this kind of bleeding edge tech, I asked him, when are you going to put a camera on your factory floor to start watching the factory floor and look for things that could be improved for the worker?

Yeah.

And they said, oh, we're in Germany. We're not allowed to put a camera on the workers.

Interesting.

They don't have that problem. So in China, they're doing that already.

How should we think about this?

We have a little manufacturing plant like Louisville Slugger. Could we put cameras on Louisville Slugger's little manufacturing plant, look for ways to rearrange the factory floor to make it more efficient or look for things. Problems?

Yeah, it's great. This is, it's a great issue. And I'm sympathetic to both sides on this one. And the way that people look at it, I think there has to be some upside for the employees in that factory floor. Right. So one thing that we've seen what's interesting, some companies are very punitive and some are more what I would call just innovative. Right. And the innovative side. Oh, we've run across people that are like, look, if location a is making 25% more sandwiches, we're willing to bonus everybody in that store for that over performance. So they start gamifying it. So it's interesting. It's like there's incentive. Right. So the camera suddenly is, hey, we're better. Right? We're better at this. And they were surprised that the employee base actually embraced that. That's very different than just punitive use of a camera where you just, okay, I think we're going to see some people use it more innovatively. But. But there's undoubtedly going to be the other side, which is some tough discussions with unions to say. Okay, but hey, we're not making as much money in this construction site as we could, which would be good for everybody. We got these tight margins and part of it is we're getting slippage on materials and we're getting overcharged on labor. Can we work together to try to get there? So I'm not trying to be Pollyannish about it, but I think I've seen both. I've seen the struggles with the union side as an example and then I've seen the innovative use where you're like, oh, that makes sense. One of the mines had two shifts, two 12 hour shifts. And they really start to understand the productivity of one shift over another shift. And they're using that as upside. They're like, okay, these guys adopt their best practices and everybody can benefit. So that's interesting.

Particularly if they own some stock in the company.

Right.

It's an employee owned company.

Exactly. That's a good point.

They should be interested in saving money and making more money. Right. Because that ends up somewhere.

Everybody. That's right. We're all owners. Right. Yeah.

It's a wild world. A new world of using it is. And as cameras continue to come down in price because of the cell phone industry. Right. They're bringing cameras more and more capable every year. This is just gonna feel that gets more and more used over time.

It's Moore's law. Right. Display is just keeps. It's playing out on the edge for sure.

What should we budget for some of the simpler things we've talked about today, like watching employees making sandwiches, you know, with five can.

Yeah.

I think how do we start thinking about, okay, this is going to cost us a million dollars to put in each store then?

No. Yeah. We always, we always try to back up from the business case of the customer. Right. For Example, if sandwich making processes improve by 5%, what does that mean for your business? Or by 10%, what does that mean for your business? And then we can price where it's, it's, look, we need, we need to make some money because we're, we're enabling the magic, but the magic's happening. Right. And so forth. And then that also depends on how many locations are we dealing with. In the case of mining, you have one massive sophisticated location. In the case of a quick service restaurant, you might have a thousand locations right in all of this. But no, this is not millions dollars stuff. This is if you've got, if you've got a bunch of locations and you know that an investment of 75 to 100k or something like that is going to potentially give you a payoff of a million bucks or more every year and maybe escalating. You're willing to make that, you're willing to make that investment, right? In the case of a mine, they're talking about if they make a mistake, it's $100,000 a day for them, penalty to them. They want to avoid that. So that's a much bigger payoff if it's done well. So obviously we negotiate around that, but really it's just backing up. But long answer to your question was you're looking, this is more like a, for a medium sized business that's trying to do something real to impact their business, this is like 100k type all in. That's the hardware, that's the software that's getting it implemented and getting it going.

Sounds like you're offering a little bit of consulting with buying the vision system as well to get it set up and get and figure out what the business need is and all that.

Yeah, yeah, I would say and you're right because in almost every case, even if I've already seen what somebody's asking and I know that I can get them 85% of the way there because I've already done this in restaurants and so forth. I know that last 15% is really specific to what they need right now. When I say they, some of them have IT resources and some don't. Right. That's just the reality. By the way, it doesn't cut by the size of the company. Some really big companies are like, no, I don't have any resource to do this. Or like I said earlier, they got their data scientists but they're not really the practical science.

Yeah.

So what we've noticed is in the pilot phase like getting something into production and working, let's say that takes about three months. You are going to work really closely with my team just because we're computer vision experts. We've seen everything around lighting and shadows and you just going to want to leverage that. And then the baton gets fast because now they really understand how to use the platform and so forth and go for it. Go create as many applications as you want. And it's really thrilling watching people use it because it's a vision. So the cool thing is you build something and you see it on your computer right away. We have this streaming capability where you can immediately see what you just built and people go, oh my God. That's literally what's happening.

Yes.

And it can work through that. So it's a three to six month process of you're going to work closely with us in the beginning and then you're going to independently move your own way over time.

You, you've mentioned cultural changes needed in some of these places. Do you work on the cultural change part of it? Getting the employees to see the benefit of this and not get out about it and start doing nasty things or whatever?

Yeah, we're not HR consultants, but. But we are. But we do. Look, when I started working with a company, if their business unit people are involved, I'm psyched because those are the people ultimately responsible for making it happen. Right. If it's all corporate people and let's say a lot of nothing against data scientists, but then it's okay. Are you guys, is this actually going to get deployed? Who's the business unit that you're supporting? So I think to answer your question, we are starting to see a bit of a shift from kind of corporate centric AI work to more business unit. Hey, what is it? I read about this stuff. What does it do for me? And that's a good shift, right? Because that's ultimately where the rubber meets the road. So we're always encouraging very early on, hey, if they're not already there, who's the business unit? Let's get them involved. Let's get them. No, they don't have to be tech experts, like who runs the stores? Who's in charge of that lumber cutting machine? Let's get them, let's get them involved.

I think that'll naturally occur in technology is one thing. Getting the humans to do it and fund it and build it and adopt it and use it and feel good about it, it's a really hard job. Tell me a little bit about your company. How big is it? How was it Funded. How was it started? Tell me a little bit.

Yeah, we're. Yeah, we're down in, we're down in San Diego, which is somewhat unusual for an AI software company. But we.

You're right next to quality.

I was going to say we, we like it because people that know San Diego, it's a neuroscience global leader in neuroscience brain research. So when you think about AI, it's like I'm trying to mimic the human brain and I'm putting things out on things. So that's Qualcomm, that's like mobile and all the mobile expertise and I'm dealing with tons of data. So that's like biotech and Illumina and companies like that. So turns out that down here all the elements from AI and we've been able to hire just talent which I my whole career I was traveling out of San Diego setting down here. Yeah. So we're just a group of engineers here in San Diego. We've been doing this about five years. We were originally started through VC funding company called Bluerun Ventures, which famous for one of the earliest investors in PayPal and companies like Waze and a whole variety of mobile companies and so forth. Big believers in vision and kind of edge AI and Yeah, so that got us running. We're really product centric. We just have a view that unless the product's really high quality, if you're going to be in the enterprise focus, it's got to be enterprise quality. Right. So that's my background and that's my co founder's background and that's. We're always focused on that.

Yeah. Since you work with so many different companies, are you seeing trends that we haven't talked about in, in companies today? Is this early days for computer vision or. Because computer vision's been used on factory.

Floors for a while. It's early.

It's early.

Yeah, it has, yeah. Machine vision, like so called machine vision. Machines that are looking like all this. You know, you're in a restroom, you're washing your hands. That's computer vision. Right. The water comes on when it sees your. That's like the lowest level of. Right. It's been, this has been around for a long time, but what happened was when you got deep learning AI and now all of a sudden you're able to do more predictive work and now that's really powerful. And oh by the way, you can put it into a software container, quote unquote and get it out on a camera. Wow.

Okay.

So that's been the dramatic change really in the last five years or so, it's still early. When you look at all those industries I talked about, it's still in that. Okay, I'm experimenting. Okay. Now I'm starting to go into pilot production, but I'm not yet at the point where I'm going to just full on, go full on with this. Some companies are, but you can see it moving that way. And I've been through this in my career. It's. People need to see the value. Right. What's the value of this? And so I think we're in that kind of early innings. Right. I think we'll be here for a little bit longer. I think chatgpt and all that, it's like this halo of marketing about AI and some practical uses of AI that we're all obsessing over now. And I think that wakes up people like, oh, maybe I should try this. And computer vision Sounds complicated, but maybe it's not as complicated as I thought. It's way cheaper than I thought. Right. So that's an experience we get with our prospects. They're like, oh, like you said earlier, they're expecting the million bucks. The cameras aren't that expensive, the edge devices aren't. They're all getting more powerful. It's the software that's the magic. Right, Right. So we're early, but it's a fascinating space. And synthetic. I'm going to come back to. Synthetic data is like a really interesting phenomena that's playing out right now in front of us. The. You saw the SORA announcement. Yeah.

Doing videos, prompting to a video engine and having a kick out of video is really cool. You've mentioned synthetic data several times. Why is that such a big deal for what you're doing?

Yeah. Because everything in AI, not just computer vision, is based off. You start with a data set. Right. So you've got a data set. In our case, it might be pictures of bread and condiments and meat for a restaurant, let's say. So there's a data set, but I need to start. Everything starts with a data set. And then you annotate the data set and then you create an AI model and then you do head on your way. That data set used to require only real images. So you'd have to go out and take pictures or video clips of bread and condiments and meat, real bread and condiments of meat. Now if I can get on in tech, say give me 50 images of bread inside of a slightly fast casual restaurant and it generates those images, they might be Useful where I can actually annotate those and help me create that ultimate that I need to create. So it's giving us a fast and cheap way to augment data sets, which is major because now I can move way faster. I can say, okay, because before you might have had to go out and collect it. And it's, it could be difficult. And so we're still a ways away. Is that really high quality? And how useful is that? But it's coming, right? And that's a wonderful thing for our industry. I think it's going to explode the number of applications that people can build and make that a lot easier and faster.

I was just thinking about robotics, which also uses computer vision, right. To see the world. In 10 years, we're going to have humanoid robots working in our stores, in our manufacturing, stuff like that. Does your systems help prepare the business for that world? And I, I help you start thinking about, okay, what am I going to use computer vision for? That will help you think about what you would use a robot for in the future.

That's well put. Yeah, that's well put. I, I think of a robot, by the way, a robot could be something incredibly simple or an Isaac Asimov robot. Right? So it's like every and everything in between humanoids, right.

I saw a landscaping robot that had two motors, right. That simple.

I just think of that as. I just think of it as a device. It's a, it's like a drone or like a camera. It's just a sophisticated device that is doing AI and yeah, there's absolutely room and need for computer vision embedded capabilities in that robot. Whether it's a sophisticated thing that just does one thing on a manufacturing line or somebody who makes your latte, right. Those. That's. It's going to need computer vision capability to be able to do the thing it does. We did some work with a. It was just an arm on a manufacturing site. It was spatial AI. Like, how far does it go? It should only go 2 meters, no more. And it does this one thing, and that's simple computer vision helping it do that. But, yeah, we're going to head into a world where it's obviously much more sophisticated. And it sounds like we're all going to be robots because we're going to have the Vision Pro in front of us as we walk down the road going like this. That's a whole. Yeah, yeah, I think I saw your. Yeah, I think I saw you doing some of that.

Which, by the way, it helps you think through computer vision, right? It's like, oh yeah, you're. You have six cameras on the front of this thing. You walk around the factory floor, a retail store. It starts, right. You start going, oh, oh, a camera can see this. So maybe we should.

Right.

The camera here studying thing. Right.

It's. It's almost like a car, like a Tesla with all the cameras and the sensors and the. That's been shrunk down because that's doing the same thing. It's like in an environment. It needs to understand the environment. Yeah. No, it's super fascinating stuff today.

It's big and ugly. Right. Five years from now it's going to be a pair of glasses that's lightweight with.

Right. It'll be total bling. Yeah. It'll be like. Yep. You can see a worker we all.

Know we're heading which is wearing a pair of glasses and using your computer vision system to validate they made their sandwich for a properly. Right. Boom. You made your samples. Right. You get another point.

That's right. Or you get. You've got the real feed of the dashboard. Like how well are we doing? And yeah, so absolutely 100%. That's a new. It's a whole new paradigm and. But really it's just pushing more and more out what we're already doing. It's just a new way of presenting that information.

What do you need from the AI industry? These smart people at OpenAI or micro. Any company that's making I think models, I think core research.

Yeah.

Are you hoping shows up in the next couple of years to make your business better?

I think that's a great, that's a great question for me because we're so edge focused. Like all the magic really happens out in the real world on the edge. And I think getting more powerful sensors, more powerful cameras, getting a wide variety of edge devices. Nvidia, hats off to them. It's amazing what they've done. But let's have a variety of options, right. For edge devices and so Moore's Law can play out and the customer can benefit from just lower cost edge devices that continue to provide more power. So I think there's a lot of investment going into edge devices and edge capability and sensors and chips and I think that's wonderful and that'll just make that edge even more powerful for what we're doing all the way on the other side with the model architecture stuff coming out of academia as well as obviously OpenAI and everything going on is amazing. I feel like that is almost in some ways ahead of everybody. So that seems like that is just happening and there's so much money going into it and we benefit from the ones that really make sense for our vision world. So that feels really good. Yeah. But if I had to pick one area, it'd be that sort of hardware on the edge.

Very cool. Where do we learn more about you?

Yeah, we'd love people. We pride ourselves in having a very thoughtful and information rich website. Alwaysai co is where to go. You'll find just tons of information there, not only about our platform, but about solutions that we've got driven out to market. Yeah, we're just, we're very active on that. Last thing is if somebody really wants to just use it and, and get going on a freemium version and check it out, we'd love that. So just come to our website. You'll see a link. Just go ahead and ask for that and we'll provide it.

Very cool. Thank. Thank you so much. This was a great hour.

Yeah, it's fun.

Caught up on how computer vision is changing business.

So awesome, man. Thanks. Thank you.