Computer Vision Deployment Disasters: Your Playbook On How To Avoid Them | Ask An Expert

Foreign.

Hello, and welcome to another exciting and elucidating episode of the Omnitalk Ask An Expert series.

I'm one of your co hosts for today's interview, Anne Mazinga.

And I'm one of your other co hosts, Chris Walton.

And we are the founders of omnitalk, the fast growing retail media organization that is all about the companies, the technologies, and the people that are coming together to shape the future of retail.

And now we spent all fall at a plethora of retail conferences, as I'm sure many of you did, from grocery shop to shop talk to NRF Paris, you name it.

And one of the most talked about technologies at those conferences that's being deployed and tested today is computer vision.

But between the fanfare and the vendor promises, it's still a very difficult technology to deploy.

Even Amazon, one of the kings of the industry, has had several issues with some of their computer vision technology.

So now that you've started your investigation, you've walked through the halls, you've met your vendors, you got your projects greenlit.

We wanted to provide you all with an opportunity to learn from two experts who developed a playbook for how to approach these computer vision projects successfully and to learn from some of the mistakes that some, some people have made already.

So it is with that that I'd like to introduce Joe Serrano, the global managing partner for retail and CPG, and Dan Horton, the VP of engineering and delivery at HTech.

Joe, Dan, welcome to the show.

Joe, you're a longtime friend of ours.

We're excited to have you on Omni Talk Retail.

How are you feeling about today's episode?

Are you ready to give us your playbook?

Of course I am, man.

Yes.

It's been great to know.

What, have we known each other eight or nine years?

Probably back when you were building that store of the future at Target and I was a highly caffeinated founder trying to do a pilot with you guys.

And we've been friends ever since and been following along.

So I'm excited to talk about, you know, computer vision here at htec.

Awesome.

And Dan, you're a first timer on the show.

We're excited to have you as well.

You've got the perfect podcast studio set up already, so I know it's going to be a successful event.

Yeah, I'm looking forward to talking to you too.

I just think this is a great topic and I really appreciate you having us on.

Yeah, it's a really interesting topic and it's a really great intro too.

Ed, I loved how you did that.

So before we get started, just a quick reminder, for those watching live on LinkedIn, please remember to ask your questions of Joe and Dan at any time via the chat session window in LinkedIn, just to the right hand side of your screen.

All right, before we dive into the questions, Joe and Dan, I'd love for each of you to give us a little bit about your backgrounds and also to tell us about your roles at htec.

Joe, why don't you go first?

You bet.

I mean, I've been on both sides of the table a lot of my career was a startup founder pitching retailers.

Then my wife said it's time to go collect a paycheck after the Target debacle.

We don't need to go into that right A little bit.

But I joined Best Buy where I led innovation, emerging technology partnerships and the digital strategy for stores of the future.

And then later joined Macy's where I led the launch and scale of their third party E commerce marketplace.

A lot of playing in the space of full Omni Channel for me.

So I've really seen like between stores and how those connective tissues work together.

I'm at htech now because I found, finally found a firm that can move fast in innovation and can build things right the first time and that builds software and hardware and embedded systems to address these issues that I've always wanted to tackle and harmonize across stores.

So I'm really excited to be leading a retail at htec, to be able to bring that to life now.

Yeah.

Joe, you're one of my go tos based on your steep, your, your very deep and steep background in retail in terms of, you know, how you think about things.

I always like to query you whenever there's a question on my mind in terms of how to approach the situation.

So I'm excited to talk to you today.

Dan, how about you?

What's your background and what's your role?

I was just thinking as you were saying that and screwed myself up just because you, you basically said he's your go to guy.

Dan Horton, it's a pleasure to meet you two here.

I am the vice president of IND at hte and from my background I've spent about three decades working at the intersection of retail.

As I say because I've been a business owner, I've been an architect, both software solution and enterprise as well as a consultant for that 30 something years now.

Getting old, it's getting gray and I had the opportunity and I'm very grateful of it over those years to help modernize major retailers in different ways, whether it was legacy modernization that we've all touched on or, you know, store operations, vendor management, pricing and promo.

MO Took years of my life doing some of that, but it's kind of cool.

I've looked back and realized that I've probably touched pretty much every system that's in a retail environment, and I've either built one from scratch, worked off of one, or tried to repair one.

And it gives a very unique perspective.

The really cool thing now is I have the opportunity, working here at htec with the group and our teams, to reimagine what the physical store looks like in the future.

So where are we going with this and where it's going to be?

So you're talking edge tech, you're talking AR VR, you're talking robotics, you're talking IoT and then computer vision, like we're talking about today.

The thing I think is really cool about that is we forget that the physical store is still 80% of the revenue for most retailers.

Right.

HTech in itself, I think, is interesting because we come from a background, like Joe said, of embedded engineering hardware and software.

So it leads us in a very unique place to look at this, this space with computer vision and store of the future.

And because we can look at everything from the physical devices, like the camera systems and all the things we're going to talk about all the way up to the software that runs it, the edge tech that managed it, the AI models that need to make it smart.

And it's really cool.

It all comes together.

So it's.

It's a nice synergy as we move forward.

Yeah, we've got a good quartet here, really, when you think about it in terms of the experience, you're bringing on the engineering side, Joe's experience in retail and then.

And, and my experience too, because I think between everything, we've got pretty much every aspect of retail covered in our experience at some point, it sounds like.

So.

All right, Joe, well, let's get to this then.

So, you know, I'm curious.

You know, we talked about computer vision.

That's what this is about.

Like what, what type of clients are you at htech working with on computer integration strategies and concepts like set the landscape for us?

Yeah.

Well, obviously htek is built across the value chain in retail, but we're getting really, really good at this computer vision space because of what Dan said and what we mentioned.

We can build from the silicon up.

So we're building, you know, smart stores, building, you know, DVR virtualization Connections to, you know, edge networks, cloud networks and those satellites that make it bring it back to headquarters.

So we're really, we're getting really good at that.

We have some great case studies with some, with some live with some live work in mostly Europe right now.

And we're really building our offerings around it.

I think it's mostly because we looked at, okay, htech are these super geeky engineers, mostly based where the airport's named after Nikola Tesla.

So everybody wants to be an amazing engineer and they're really good at building like we're building autonomous vehicle technology, you know, across the board.

So I'm, you know, you know, I'm crazy enough to think about this, right?

Like we're looking at if we can build autonomous driving vehicles, how do we apply that to retail?

How do we make an autonomous driving store?

The store is the robot itself.

So that's essentially what we're looking at piecing together.

And we've got some good start with you know, we're building, you know, everything from you know, shelf availability, inventory shelf on computer vision to looking at the point of sale checkout to multi channel camera tracking across the store in aisles.

We're even looking at things inside the warehouse itself and like helping to pack, helping to create better processes there.

But I think we're also seeing that, you know, there's been a lot of point solutions and this is where we're at I think in the platform shift with AI in particular computer vision is if you look across all the store operations, shopper insights, operations, retail media, lots of point solutions across the board over the last probably five years.

And I think we're at the, just starting to look at how do we harmonize those together to those were the unbundling and now we're coming back to how do we bundle these together to have a single pane of glass to really start driving AI forward as the models get better and better and better.

So that's kind of where we're at now.

We're early in the process, but we're getting really good traction because of what we bring to the table from cross sector pollination.

And Joe, I'm curious too.

You guys have both mentioned the store side of computer vision implementations pretty extensively thus far.

Are you also looking or helping clients or retail customers with the computer vision side of things on the, you know, e commerce operations side of the equation too?

Is that part of it as well?

Well, I think we're, we're, we're sort of niching down into the store because of what Dan said before is obviously lots of people are still in store.

But yes, I think, I know that you and I think a lot alike in this as we have all eventually have to harmonize across it all because it's not just about.

Well, the stores are becoming micro fulfillment centers.

So how are you going to have the buy online pickup in store?

How are you doing the picking?

So all those elements eventually have to come together.

But right now we're focused on the store.

We certainly are looking at how do product catalogs, how do the product catalogs play in and how do we have inventory across both E commerce and store.

And I think those are going to be elements that add to the equation in the very near future.

But right now we're focused right now on store and just nailing.

Right, right.

By working on one, you're effectively working on the other too.

You almost have to if you're building micro fulfillment centers.

Right.

Well, Joe, I think you did a really good job there of identifying kind of the entire landscape of where computer vision is being used.

It's definitely the backbone technology for a lot of the digitization of today's retail stores.

But, you know, all of us have been there.

You mentioned a bunch of things that people are exploring and testing right now in their own retail operations.

You see all of these demos on the floors at conferences.

But, Dan, I want to go to you first.

I'm curious from your perspective now that you're working with all these retailers, you're trying to figure out the right computer vision deployments.

Where are you seeing, like the biggest gap for the partners that you bring in between what they thought they were going to build or what they were sold on the floor and what they're actually experiencing.

Once they start to deploy this in their own retail operations, everything in a.

Demo is going to look like it's 100% perfect.

So I guess I would say when I'm looking at the question from I'm going to a demo, I'm going to see something from a vendor believe that if it says it's 90 or 100% accurate in the store, it's probably going to be 60.

Yeah.

And it's so kind of knowing that.

But I think the things I would look at would be, I guess, two angles.

If I was going to ask questions on it immediately, I'm asking of where have you really run this?

How is it run?

Can I see it running in multiple stores with real world data?

Or at least do you have, you know, digital twins set up with the mockups of actual stores so that I can see the chaos running correctly.

Like I need to know what this is going to do in the real world because you know, it's the chaos in the store of all the edge cases that's going to show you whether whether this is going to be successful or not.

We can all create a perfect demo.

I think that's really where that comes in.

I think there's secondary parts and that's where I think the gaps come in, which is a demo is going to have pristine data, but I don't have controllable pristine data in my store still.

So I should have questions of how is this working and handling around those data sets?

How do I plug in data sets when I'm in a store?

How easily can I modify the data sets I'm working with so this thing can get smarter and better?

So we have data sets.

What is the camera types that I can work with?

Can you handle legacy cameras?

Can you do analog cameras?

Because a digital camera of a 1080p or better is going to give me a certain recognition because it can do more fine grain visibility where you know an older 720.

Right.

It's just not going to see anything.

So if I'm trying to do theft and I'm using old cameras, it's probably not going to work.

So then I need to do that.

But then there's also the really important thing is if you're looking at products, whether built like we do, custom built, or you're looking at out of the box solutions, you also want to know do they require more like rip and replace or can I augment existing components?

Because every use case you go through doesn't require the same level of complexity or the same level of hardware.

So you don't have to waste money trying to buy all brand new cameras.

They've found in a lot of the research that's been done through surveys that a lot of retailers can leverage 60 to 80% of their existing environments with minor augmentation of the environment, the systems.

Itself, which we're doing right now too.

Right.

You can build an augmented component, special component that you can put on top of some of the old legacy cameras to one move AI closers, you do closer processing, which is huge for network latency issues and things, or you can just extend the life of that camera a little longer to at least allow you to validate your current use cases before you get to more complex cases.

So just interesting things to think of as you go through it.

The other point that you didn't bring up that.

I think I know how you're going to answer this, but I want to make sure we bring it up for the audience, too.

As you look for you look through, you know, what aspects of the needle one needs to pass through as you're evaluating a demo is, is also like, what impacts the customer.

Right?

Some of these computer vision processes and implementations can or cannot affect the customer.

And when I, when I hear statistics like you're saying like, 60, it's probably going to be 60%, you know, accurate to what you need it to do.

I don't want that touching my customer at all.

So that makes me think, like, particularly with the smart carts, like, you know, you got to get that working perfectly in an innovation lab setting first.

Then you're probably trying it in the real world with just your employees or people just testing it before you're actually even going to pilot in a lot of stores.

And I think when I look back on the past year, two, three years of all this smart cart talk in the media, I don't feel like that's what's happening.

I feel like people are just jumping in, piloting these things in store, and they're probably uncovering the fact that, oh, man, these things aren't anywhere near the act, don't.

Don't have anywhere near the accuracy we need them to.

So.

So that's got to be a piece of this equation, right?

Is like, what am I seeing and what's the likelihood it's going to impact my customer or not?

So I think that's beyond the questions for the demo part of it.

I think that's really getting into some of the other conversations too, we were having today.

Because that's really the critical question that should be asked, I think, above the demo, because a lack of ROI and a lack of focus, of actual value from this is going to negate many other things.

So to what you just said with smart carts.

Joe and I were just talking earlier, and in the conversation, it was smart cart was one of our topics and the realization that, like the Amazon example I had, you couldn't take the cart out of the store.

So you want me to take it from a smart cart and you want me to put my bags in a dumb cart so now I can walk out to my car.

How much time did you actually save me?

You didn't, right?

So your question is perfect because there's a loss of, there's a misplacement of the technical value we're adding to the customer satisfaction that we're hoping to give.

Right.

It's a loss of trust, you being that customer.

Me, I know myself, I would not be happy walking that out to my car after I did that.

So now I would be less apt to use the service altogether when I came back.

Right.

And I think there's other components like that too, which is why one of the concerns and the questions we should ask on these two is that privacy, data collection, roi, usability.

Because there's also parts of computer vision where you're watching what people are doing, right.

And then how are you using that?

Because if you're incorrectly losing that, that also drops trust.

So all of those things to what you just asked kind of drop trust and then customers don't feel comfortable to shop with you anymore.

And I want to get to that later, later too.

So.

But before we do that, Joe, I'm curious too, like what, what role does like an organization's perceptual AI readiness have?

Like how does that, that concept come into play?

I love that term, perceptual AI.

I don't.

Maybe it'll, maybe it'll take off as, as like where this goes eventually, right?

Like perceptual AI, right.

Well, I mean every retailer is different, right?

Like you got to grow how they perceive it.

Yeah, it's about like a little bit about what Dan said beyond the demo itself.

Like are you edge ready?

Are you going to choke on 4K?

Because you've got a 15 year old network that's in the, in the mop closet, you know, and that's going to be a problem for you to set up the edge networking problem possibilities for you.

And I think just being a startup person that's in a startup and then moving to, you know, leading innovation or, or stores of the future inside big organizations, you have to have buy in across the board so you should really bring them along early.

So you got to bring your, your store ops team, your district managers along for the ride here.

Like what is the biggest problem that you had?

Is it, is it inventory availability?

Is it you're losing vegetables to like going bad.

So I think it's going to be retailer by retailer and I think that's where like we really focus on because we, we build custom solutions, not point solutions that can tie all these together is like where are you really starting from and what do you want to work with?

Because ultimately AI should be creating enterprise value itself.

So is it on shelf availability is a planogram compliance that you want to focus on and really diving into?

Where do you can you Derive the most ROI and then get realistic about what you can do with AI.

And I think it's going to be most important, especially now based on where the AI models are at today, is to really start building the foundations and building think platform versus point solutions in the future is one foot in platform, one point and one foot in hey, where are we going to start this today?

And maybe it's inventory tracking, maybe it's operations and labor productivity, you know, maybe it's dwell time or queues or whatever the retailer might think is the most important roi and doing that estimate beforehand and then making sure that you're getting buy in across all those stakeholders, not just for the point solution but for the platform strategy that you have for the future of this to drive enterprise value.

But Joe, if I play devil's advocate for a second, like, you know, I could see, I could see the other side of the coin, which is the retailer being like, you know, it's probably less risky for me to attack it from a point to point solution because if I invest in the platform design and I get that wrong, then I'm really left holding the bag.

So, so how do you, how do you think about that dichotomy?

Well, I think you get.

Well, I mean again, you're going to have to run a, you know, a demo that works, a pilot that works in crawl walk run fashion.

I think that's where I take from my startup experience is like I'm always telling you, I advise startups.

I know you guys do too.

And then you got to have the discipline to scale down on that one, that one use case.

It might be painful for how small it is, but being able to nail that first to show it and then move forward from there and crawl walk run fashion, it's really hard to do that.

I mean inside every large organization I've been in, they make it work in one store and they want to scale it to you all 200 in the next six to nine months and then you've got different silos.

I think that's the other challenge.

We have these holdover of Taylorism, right?

Everybody's in a silo, everybody is incentivized for different things.

You have to get them all aligned.

We are going to have to move to new different models of operating with AI, the better and better it gets.

So I think you have to consider those things for the future.

And that's a change management issue and that's probably the bigger challenge with all of this, even over AI itself is how do you align change Management to make this work well.

So assuming Joe, that they have all of those things in order and they've sorted through that, that mess to get to their pilots.

Dan, I want to bring you back in here to kind of COVID off on some of the points of failure.

One being one that Chris and Joe were just talking about where you have so many point to point solutions that you have and you have different teams operating those or initiating contracts with those partners who are bringing us in.

What are some of like the more common points of failure that you're seeing with the computer vision deployment and what are some ways that the audience can avoid them?

I think you actually even from the conversation we were just having, I think it's worth noting that in general they say that only 25% of retail AI pilots, right.

Any kind of, for these pilots we're talking about really scale past proof of concept.

It's more about learning from those things each time when you're deploying.

Right.

But that was 23, 24 and 25 now, right.

All the executive teams turned around and said okay, now I want some value for my money I spent.

So show me how AI works, show me how I use this in the stores.

And the funny thing is when you talk about the platform and building the single platform and everything else, a lot of retailers are still fighting with data.

As an example, we still have problems with data environments, we still have problem with integration.

That's why point solutions come up, Point solutions come up to try to work around the pains of today and let me kind of find a crack so I can try to provide some value to the organization and get away from all of the other, you know, people process tech problems I might be dealing with today.

So there is a real complexity in that we, we're not going to deploy computer vision or anything else on top of this store without understanding what we're build, what we're deploying it to.

Right.

How we're deploying it into that environment.

So it's going to be things from understanding if you have outdated hardware, if you have bad bandwidth.

Right.

If we don't have the computer systems we need.

So I think understanding where your possible failure points are going to be in the environment is one of those important things that will be a consistent failure point for trying to run this kind of a project.

And then you mentioned privacy and GDPR too.

I'd love for you to talk about that.

We didn't get to dive into that too deeply earlier.

But like why is that something that's become more important the more that we start to see computer vision deployed in stores.

And how do you avoid that from becoming a disaster?

This one is interesting because it's a mix of.

Everyone you talk to is going to tell you it's a technical security problem.

I think they're going to tell you it's gdpr, it's ccpa, it's, you know, we're another foreign country.

We have double the rules you do in the U.S. you know, privacy is most important.

You have to.

We have to have this.

And yes, you do.

Like, this is.

This is table stakes.

You need to have that security.

But as we were talking earlier, there's a bigger part here that keeps getting missed.

And, you know, how comfortable are you when you're in a store?

If you felt like you were being monitored, even as an employee, if you felt like you were being watched for your day job 24 7.

Right.

I really don't like that.

And you guys, you know, so I think there's a bigger part here for privacy and consent more than just the security.

Table stakes.

I need to have gdpr, there's shopping for something.

What if you're buying products in a store that, you know, you really want a little privacy to go buy, you know, and now you feel like there's cameras everywhere, you know, staring at you.

Right.

I think that's the bigger concern that people have, at least including employees and everyone else that we are maybe not taking into account.

I also think putting.

Personally, I think just putting signs up telling me that you have cameras in your store is not solving the problem.

And that seems like something, Dan, that's along the lines of the smart cart that you were just talking about too, where it likely wasn't an issue until someone brought it to their attention.

Like, the pilot was like, let's go.

We're just no big deal.

We're going to have cameras everywhere, and it's going to provide all these benefits.

But then you have the issue of what all are you watching?

What do I.

What freedoms are you taking from me as a consumer?

Do I imagine it's the first time the case study goes a little bit out of bounds.

Right.

So I did a test a long time ago, playing around with saying the word for chicken coop.

And we were testing a scenario, and it showed up as advertisements on Facebook for me the couple hours later.

And we were testing the fact that, you know, certain things are listening to you.

And think about the first time the camera's in the store and, you know, they're on and it watches you shop and then it starts giving you very specific recommendations for products that it knows and watched and shows you that it watched you walk the entire store.

How comfortable would you be to go back and shop?

Right.

If you didn't opt in for that and you didn't ask for that, you would now feel violated almost.

And I think that's where the.

That's where the privacy and the trust is going to come in.

Because trust is hard to gain.

Right.

It's easy to lose.

Hard to gain, they say.

So if you start losing customer trust in what you're doing, then no matter how great this solution is, it's not actually going to increase revenue.

Yeah.

You got me thinking about something I've never thought about really, which is there's the whole mantra.

I think Facebook originated it.

Go fast, break things.

But when you start talking about computer vision, AI deployments, the edge cases that you may not understand yet are proportionally probably more important than some of those edge cases in that traditional mantra that were.

That are probably encapsulated in what that means.

And that's something I never thought about.

So, Joe, I want to hold your feet to the fire then, based on something you said before, which I think we would.

Ann and I would fundamentally agree with you, which is you have to decide what you're going to go after.

And so if from your perspective, knowing what you know, if you were a retailer and you say, look, we think computer vision can do something for us, which is probably the wrong mindset to actually approach the problem with to begin with.

It's really, what should I be using it for?

So what do you think is the biggest problem out there that you think computer vision can help retailers solve or get their arms around?

I mean, you got to start.

We're seeing better.

Saves you the most money right now.

Right.

So I think in my perspective, that kind of goes down the path of inventory visibility.

Right.

You got to have things that are available that your customers want, want.

Right.

You can use that to.

In more predictive nature, probably more so in the future a little.

The near future around, hey, where's.

Where's the dwell time happening?

And.

And where are we out of stock more?

And then how do we.

How do we make actions out of that eventually?

Right.

So I think starting with inventory visibility and, and stocks is a great place to start, particularly if you've got micro fulfillment in the back tied to your E Commerce as well.

How does that all play out?

Shrink control.

It's a big issue.

What are we at 1.6% on average?

@ stores and people keep changing their self, self checkouts or monitor checkouts or whatever.

So how does that play in?

I mean I know there's various types of shrink, but you know, I think that's a good place to start too.

If you're looking at security and inventory control, I think those are high roi, low risk loops for you to start and then, then from there you can start looking at that personalization piece which starts to get that, you know, that creep factor a little bit.

But like you got, we're going to have to figure it out.

Like, you know, these, I'll say these autonomous driving stores are going to be inevitable.

So like pick your, pick where you think you can win now and where you can save the most and build the ROI case and build your foundations and you can go from there.

I think then you're going to get into retail media, the whole retail media thing.

I think that gets into where what's really pushing these smart carts to some degree too.

Chris.

Right.

And then it's like, I don't know, how do you play that the right way?

I know the whole, I was always on my, my soapbox when I was leading digital strategy for Best Buy store in the future.

How do you do, how do you do in store what you can do online?

But does it really play out that way?

And then, and then what are the cause?

There's a, you're going to have fundamentally different problems in the physical world than you are in the online world.

So.

Well, what are people really willing to accept?

And you got to be real about that and place your bets, I guess.

Right?

Yeah.

I mean, yeah, you're kind of backing up the thesis which is like I'd be focused on the operational side of these of or solving the operational problems before I start trying to solve the consumer facing problems.

Because you're going to get operational problems as you focus on the consumer facing problems.

But Dan, what do you think here?

Same question to you.

I like where Joe was going with all of a course because we work together.

But in all seriousness he brings up a very good point and it's like the shelf, the out of stock type of detection.

Those areas allow for you to empower and almost augment the store manager and the employees in a way that you can make the store run better.

Thus the experience for the customer gets better.

But from a technical standpoint, the part that I think is really important that I don't think we should overshadow is those tolerate imperfections so they allow us to provide faster or roi because real time is not necessary day one.

Yep.

And perfection is not necessary day one.

And I think that's important because if we looked at the checkout like we're doing, and we're looking at shrinkage, we're looking at theft, we're looking at that type of thing, then everything needs to be super fast, top notch, you know, we can't make mistakes.

I can check a shelf, have time to process it, and have an end of day report that shows where we're tracking or not tracking.

I can manage to tell you that there's crates in the back room of the store that haven't been put out yet within an hour.

And then make sure that the system is also written in a way that I'm providing real roi.

So don't create me an end to day report.

The AI in the system you're creating should actually tell the closest employee to where those boxes are, that here's your next task, go back, grab that stuff, put it on a shelf.

Like it needs to be a smart system, not reports.

We need to move away from reporting.

Right.

And I think that's the part, the imperfection part I think is really a key.

I think that's where we're at in AI right now.

There's the hype cycle that was, I don't know, we're already at in the hype cycle for AI.

I'm not sure exactly where we're at, but I think there is the hope of what it can be and I think it eventually will get there, but it's still a little sloppy.

Right.

So you're seeing this in the agentic commerce world a little bit.

Right.

They're talking about moving from SEO to GEO to something new.

I mean, there's some sort of new way to optimize every month now with that, but at the end of the day it's all going to be powered by product catalogs that have to be absolutely precise.

But that's only just, you know, having an inkling of being possible today.

That's why people are more focused on being optimized to an LLM, you know, engine for search, which is a little bit more sloppy than optimizing to a taxonomy to a specific retailer or a marketplace, for example.

Right.

It's not quite there yet, it's almost there.

So I think that's kind of where we're at.

So you gotta like, you know, build the foundation still to have that data ready, have the edge ready so that when the LLMs can pick up, you can do some of these Things more in real time.

I think we've, we've had some success operationally with warehouse robotics and are saving one one of our clients like $6 million a year using some level of computer vision just so the robots are on a more efficient path.

For example, that's a low hanging fruit you can tackle.

Yeah, but Joe and Dan, both of you, Dan, I'll go to you first because you mentioned it earlier in our conversation, but how do you then kind of look at and evaluate when you're making the decision about where to deploy computer vision, AI?

How do you evaluate what you do with your current hardware versus investing in some of the, the warehouse robotics like you're talking about, Joe?

Because I, I think part of me as a retail exe executive is thinking I already invested in shelf edge cameras or something.

How can I make them work harder?

Or I theft cameras or something like how can I get more with that current investment that I have?

And how do you make the decision to say okay, but then there's these, you know, these camera or these robots that we can use in the back room that will make these operations more efficiently.

Like where's the build versus buy rubric that you'd use?

Dan, build versus buy is.

Yeah, the one of those quintessential forever questions.

Right.

Which one's the right answer to what.

But you bring up an interesting point is most large retailers that we would go and speak with and then we do speak with are never greenfield, nobody starting from scratch.

Like we have tons of stuff.

That's a good point.

So I think there's, there's finite rules that you can kind of use.

So like you said, if you were doing that assessment.

I think it starts with an assessment.

You have to assess the current hardware and the infrastructure in the place where we're playing because there's what is available and then how old is it?

You know, we use the word legacy in general, but there's true like old end of life.

Like we're no longer should even have this, we're going to replace it anyway.

So that already answers your question.

Right.

If it's truly that old and I'm going to be getting rid of it soon, then just get rid of it.

Like don't waste brain cycles trying to figure out, you know, if you're going.

To save, can I reuse it?

Yeah, right, right, right.

It's a matter of looking at all those components.

So you're talking about everything from cameras and sensors because they might have existing sensors today.

If you put in Iot and You're running MAPK systems and you're teaching it to get smarter around all of your sensors in your refrigerators.

And maybe you're doing shelf waiting today and you're doing a lot of other things.

You can leverage that with the cameras.

To me, the cameras, and I've been having fun with the title, is the Eyes of AI.

So, right.

The camera system.

Oh, I like that.

Into the store, right?

Yeah.

It's going to catch those edge cases that you couldn't do with the other sensors.

So we want that data, but you got to do a current state.

Where are me?

Where am I with what I have?

Can I use it?

What am I actually missing?

Because the other interesting thing is computer vision might just be taking care of one gap you had with your existing systems.

So then the question changes.

I'm not ripping, replacing, or changing anything.

I'm finally doing what Joe said, and I'm bringing it all together to now create a unified architecture, a unified system that can be made from multiple products.

But I'm bringing them together to finally work as one thing, and this might be the thing that does it.

So I would look at what I had in the environment itself.

I know I said it before, too.

The interesting thing is a lot of conversations with retailers.

You find that still 60 to 70% of what's in the environment today is still reusable with augmentation.

So this really isn't the kind of a rip and replace kind of thing like, you know where you want to go and, you know, what you have.

Now we can logically say, okay, how can we get there?

Right.

You might find that the case study that you want to run because of the roi, you know, you want to get only touches a third of your cameras.

So you only have to modify a third of your cameras.

You might find that your network is okay, but maybe doesn't support the bandwidth you need.

But if I do edge tech, or if I actually move AI right up to the camera, which is available today, then I can do processing and, you know, more near real time right at the camera itself, instead of waiting and having network latency.

So I can adjust my architecture in a way that I manage my.

My limits that I have in the stores so you can get creative with what you have and the things that are available today.

And tech's running, tech is getting smaller and faster.

And even like we're doing at HTek, we're putting AI on chipsets, so we're moving AI right up to the edge into custom devices that we build in Retail and other domains.

And this is making a difference.

It's reducing the cost to fix the problem and it's allowing us to be more creative to fix problems.

Do you no longer have to do the old big box, you know, software solutions like we did a lifetime ago either?

Like I don't have to rebuild everything.

I love the idea that computer vision is the eyes of AI.

And then I step back from this entire conversation.

What's interesting to me is smart carts were here long before generative AI.

Right.

We were talking about those five, six years ago.

And now I think almost like we're getting smarter as a collective industry now in terms of how to evaluate what technologies and which technologies should go before others.

At least I feel more equipped to do that now, having talked to both of you.

So thank you so much for your time.

Dan and Joe, if people want to get in touch with you, Joe, you or Dan, what's the best way for them to do that?

Well, you can look us up@htech.com we also have HTAC AI you can check out for our AI capabilities.

But if you want to get a hold of myself or Dan, my email is Joe Serrano, techgroup.com and Dan's is dan.hortontechgroup.com and you also have a webinar.

Coming up in January where you're going to go even deeper on this topic.

Right?

It's hard to believe, but you're going to go even deeper.

Tell us about that.

Even deeper.

Yeah.

So session was about why the projects for computer vision fail.

So we're going to go into how to, how to actually do it right in the next one.

So in January we got a.

We're hosting a smart vision in action, Turning Retail Cameras into Profit engines.

It's a deep dive into how retailers are transforming their current cameras, like Dan talked about and if they already have, into real time engines for profit and efficiency and a better customer experience.

So check us out in January.

All right, I think I might do that because, God, I learned a ton from talking to both of you today.

So that wraps us up.

Thank you again to both Dan and and Joel for joining us.

Thanks to everyone that joined us live and who also might be listening in later.

Today's podcast was of course produced with the help and support of Ella Siriort.

And as always, on behalf of all of us here at Omnitok, be careful out there.