Hello and welcome to the NVIDIA AI podcast.
I'm your host, Noah Kravitz. quick programming note or sort of more of a housekeeping note.
If you haven't already subscribed, subscribe.
You're listening to the podcast. You love the podcast.
So let it come to you. You can find us on Apple Podcasts, Google Play, SoundCloud, Stitcher, CastBox.
Or if you want to binge, go to the NVIDIA AI podcast page on the NVIDIA blog.
We've got every episode up there. order.
You can go in reverse order. You can skip around.
You can get crazy with all the podcast content.
They say the customer is always right, but what makes customers happy?
Fully stocked shelves, fast and easy checkout, and friendly and dedicated staff, according to our guest.
Francois Chopard is CEO of Focal Systems, a startup focused on changing the brick-and-mortar retail experience with deep learning.
Francoise here at GCC19 and took the time to join the show.
Thanks for stopping by and welcome to the AI podcast.
Thank you for having me. So on your website, front and center, the slogan, AI to make customers happy.
How's Focal doing that? So I think we chose that slogan and that mission really because it's not about automating brick and mortar retail.
And you can do that just by taking, cutting, slashing half the labor in the store.
And some retailers are just doing that as labor costs increase.
What happens? You get longer, longer lines if you've been a Safeway in the last ever, right?
The last five years, just, you know. And it's even worse, because it says, ring the bell when the line gets three people long.
I'm the sixth person in line, and they're not ringing the bell and not increasing the lanes.
Just don't even tell me you're going to do that if you're not going to do it.
And then you get more out of stocks because you've got just less people.
That just pisses off customers. It's basically giving me a coupon to go to Amazon.
Thank you. Next time I just won't come. So that's what it's really about is...
They're going to cut the labor and do that kind of stuff no matter what, because that's what their S&E permits.
So otherwise, they're going to go bankrupt for a lot of these, if you look at their income sheets.
So it's really about making customers happy and making the in-store experience good again, right?
And that's kind of what we're trying to do.
The way that we do that, there's kind of two big avenues that we do that.
If you take the pie chart of labor in the store, Maybe about 15% of that is in shelf integrity, which are things like scanning out of stocks.
They've got to do it every morning at 6 a.m.
9 a.m. If anyone is working grocery, they know that.
Home Improvement, all these different retailers.
Right. Same deal, what do I have to go order more of, right?
Knowing that answer takes about four hours a day.
So that's a huge amount of labor. Then the cashier is about 42% of labor.
For Walmart, if you look at their income sheet, about $100 billion a year in SG&A.
So that's about $42 billion a year. And just like picking up items, finding the barcode, scanning the barcode, putting it to the right, and then bagging it, making sure you pay.
I want to say something like, there's got to be a better way.
Yeah. Just wait, there's more. And the remainder is in stocking and customer service and stuff like that.
Customer service is getting slashed. There is no more customer service.
Imagine a world where we can automate. The shelf integrity piece is audit.
It's observation. It's not moving anything, right?
Right, right. No. Why can you not automate that with a really simple, small little camera?
The camera on your iPhone, the bomb costs like $4.
How can we not just supply $4 cameras? You need 500 of them in a store.
It's a 2K or something like that investment to automate that task.
Seems pretty trivial, right? So we look for opportunities like that.
You know, the automated checkout. There's a huge swath of folks that are trying to apply deep learning computer vision. to solve that problem.
Amazon has been working on it and has spent billions of dollars in research of Amazon Go.
And then there's the Amazon Go Me Too's that are coming out.
When we started, we started about 2015. This is three years into the AI craze.
And I was at Stanford AI Lab. And at the time, there was a bunch of people leaving to go start companies or join companies.
In my lab, it includes Andrew Karpathy, who's now the head of Tesla Autopilot, Drive AI, came out of the lab, a bunch of other people.
And I was like, there has to be more than just automating my car and automating my doctor.
What else can we do? And we kind of looked at from a macro level, well, if we believe this is the age of automation, everyone's talking about that.
Where is it happening? Where else should it happen?
And who has the highest sensitivity to a decrease in SG&A?
Well, if you decrease retail's SG&A by half, you triple, 300% increase in EBITDA. that goes from Jim Cramer's like, you know, sell, sell, sell to buy, buy, buy, right?
That's like the best, like, Tell me anything that can happen to an industry that would cause a 300% increase in EBITDA.
That's insane. And I think that that's what's going to happen.
I think that's what Jeff Bezos sees. And that's why he's investing so heavily into Go.
And so what does Focal do at this point?
And I'm not sure what stage you're in as far as the startup development.
But what do you guys do or what do you envision yourselves doing for, you know, your customer, the retail store?
Yeah. So we have our two solutions. We basically try to find when we have this macro strategy of like, okay, this is clearly the space we want to be in.
We believe that let's break down the problem like engineers into sub-problems and solve those.
What's the quickest, easiest solution thing that we can automate inside the four walls?
And it was a shelf integrity at first. So we deployed these small little cameras.
They're really inexpensive. We plop them. throughout the store, specifically in high-moving areas to begin with.
Makes a lot of sense. Take the soda aisle, for example.
Pepsi and Coke, you'd be amazed by such a high percent.
It's only, let's say, 20 SKUs account for maybe 15% of sales.
Right. yeah right it's like it's completely the Pareto principle um 80 for the 20 right and uh And they go out of stock all the time.
And even worse, they're DSD, which means direct store distribution.
So Pepsi and Coke don't even know that they're out.
Okay. And so it's a huge problem, and it's an incredible amount of lost sales, and Pepsi and Coke have been growing, and Dr. Pepper and... that category has been growing at very small amounts.
So you've got your fixed cameras and they're able to just automate the process of detecting They take an image once every half an hour.
Okay. And they give this nice chart that says 8 a.m. in.
9 a.m. in, 10 a.m. in, in, in, in, 4 p.m. out.
So I know exactly when it went out. Right. which means I can tell you that on average, we call it out-of-stock hours, not necessarily out-of-stocks.
So now we're trying to reduce out-of-stock hours.
Every single hour that you don't have a product on the shelf is lost sales.
Less products on the shelf, less sales. completely correlated.
So what we do now is we deploy these cameras and just go through some numbers. take the Pepsi Coke, Dr. Pepper example, we calculate that they have about let's say 5%
On average, on-shelf availability. On the holidays, it's insane.
They get down to like... Maybe 5% out-of-stocks, 95% on-shelf availability.
They get down to maybe 20% out-of-stock hours and 80%.
On shelf availability, which accounts for maybe about $30,000 a month in lost sales.
Wow. per store per month. So you have $30,000 you're leaving on the table where people are coming in, they're looking, oh yeah, they don't have Coke today.
All right, so let me go buy. Dr. Pepper, let me go buy something else.
So not only are you not getting the sale, you're funding your competitors.
You're giving them a gift card to go buy your competitor.
There's nothing cocaine's more than losing the sales of Pepsi and vice versa.
And you can solve this with rough numbers.
Let's say a $20 camera. You just deploy and it takes an image once in a while and sends you the data. all that, that recoups about, let's say we can make a small little change to the planogram, increase effacing, maybe change the route. where now we're recouping, let's say, just one-third of that.
That's $10,000 per grocery store. You know, there's...
30,000 grocery stores in America or maybe 50 now at this point.
I think it's like, so it's like, was it $300 million, $500 million difference?
Sounds easy when you lay it out like that.
I mean, and that's, you know, I think per month.
I mean, I'm doing the math in my head here, but it's insane. numbers in an environment where there's very low growth.
And so the second bucket to your approach?
The second bucket is this cashier piece, which is a little bit the holy grail that everyone's been going after.
You've seen Amazon Go and things like that.
We started before Go, and we were looking at the checkout.
And I remember, you know, meeting with... We're at, you know, Stanford...
AI lab and we left to go scout some of the local retailers and see what their appetite would be to apply deep learning computer vision to automate their checkout.
And it went over to Drager's and Bianchini's and small little retailers around here.
And I remember. One of the guys at Drager's, Roger Marin, basically told me, he's like, Francois, I don't know how else to say this to you.
You're never touching my POS. Why? It's the lifeblood, right?
So like, I mean, they make an edit to the POS once a year, maybe twice a year.
You change that thing and it goes down for a day.
You're stuck. That's a lot of money gone.
And you lost those customers too. Right.
Right. If you've ever been in a store where the POS is down, it is a nightmare.
Right. And so... It's a huge issue. Amazon Go came around.
We went after the shelf cam solution, this out-of-stock piece.
And now we're in 11 major retailers in the world.
We have deployments in Australia, in the UK. a bunch of stores, a bunch of chains in Canada, all throughout the United States, obviously.
We have 30 stores just in Manhattan. It's like that is clearly a signal that there's clear need there.
And what's really cool about it is that Yeah, the Pepsi, Coke, General Mills of the world, they make 50% margin on the products, maybe 40%.
And the Walmarts of the world make 20%. So the CPGs are twice as motivated to make it happen.
So there's a marriage there that's really good.
The cashier piece is only retailer-led. There's no way the CPG is going to pay for that and make you a better cashier.
So now that we're post-Amazon Go world, where Amazon's talking about adding 3,000 stores, and that's not a bluff.
They will do that. Mm-hmm. you know, what should happen.
It doesn't make sense for Walmart and Kroger and all those people that have Amazon Go.
Like, that solution is... complete overkill, very, very expensive.
It's unbelievable that it works. It's unbelievable.
There is a lot of, you know, Wizard of Oz-ing behind the scenes.
You guarantee that, right? Sure. Because we've done it where we took the three different sushis, put them in different parts of the stores. what they call the cup and the ball, you know, like try to really pull it.
And, and that one, typically you get your receipt in five minutes.
That one, we got our receipt in about 45 minutes.
Interesting. Right. So, You can see there's a little bit of, no deep learning model takes 45 minutes to run.
I know that. Otherwise they're not on a GPU, right?
There's my NVIDIA plan. So what is the solution?
Walmart doesn't have AWS. Kroger doesn't have AWS.
And even if they did, I still don't think the economics really work on that.
In a 1,000 square foot store like Amazon Go, the equivalent is actually smaller than a CVS.
And how many cashiers are at a CVS? Yeah.
So now you're automating one cashier. That was a $50,000 a year cashier.
Right, right. You can claim that more people are going to go now and things like that, but I've never really had a long line at CVS.
That's not the issue. The issue is the big format stores.
Amazon bought Whole Foods. That was a big splash as well, which was great for our business.
And they just launched last week, they're going to be rolling out their own.
In Whole Foods, they can only do whole products.
That's a laundry list of qualifiers to make it onto a Whole Foods shelf.
They want to sell the other stuff, and that's what Amazon's going to do in the other stores.
And that's going to be directly ripping apart some of Walmart sales and things like that.
They're going to try to get into that $520 billion that Walmart makes every year.
And so how are they going to automate? I don't believe that Amazon is going to put Amazon Go in those format stores.
I do not believe that. I had done the economics.
We've tried to build that in-house. And we said that the cost is enormous. astronomical, just in kilowatt hours, the amount of watts you burn running that many GPUs.
It would be great for NVIDIA if that were true.
But until you guys come out with a chip that's 1-100 the cost and 1-100 the power draw... which I don't think is physically possible.
Otherwise, Jensen would have done it already.
He's the most qualified guy on the planet to do it, and he hasn't.
So it leads me to believe it's not physically possible. means it ain't gonna happen.
And so what does happen? Well, there's lighter touch solutions where I can automate half of the checkout, right?
How do I do that? Well, we have a small little camera we place on top of the conveyor belt.
We call it Focal Scan. Which, barcode scan, I have to grab this product.
I have to scan this thing. It takes on average, anyone in retail knows, it takes on average three seconds per product.
Per item, okay. Right, per item. And each lane is around between 22 and 30 transactions per lane per hour.
Well, if all your products are already on the belt like that and I have a camera and I can identify all of them, there's no more scanning.
We just scanned it in 0.1 seconds. And we use an NVIDIA card, so it scans in 0.1 seconds.
I just scanned the entire basket. It cuts down the checkout speed.
That's huge, yeah. Check out time by 60%.
And that's just computer vision object detection.
You're just using deep learning computer vision to detect amongst 100,000 different SKUs.
Yeah. in 0.1 seconds with 99.9% precision recall.
And that's what we're building. And that's a perfect problem for us because we love that.
That's nice, tight, contained. That is just build, get more accuracy and something that my team that has... you know, a number of deep learning computer vision engineers out of, you know, name and top tier school.
Yeah. Right. So that They love a structured problem like that.
So it's great. And we have that first solution deployed in Manhattan right now. in one of the fast-moving stores on the planet.
It's astronomical how much volume these guys do.
And right now, it's in passive mode. And we're going to flip it on into active mode pretty soon.
So it's pretty exciting. And then we have it live, obviously, internally at Focal.
So we call it Focal Scan. Right. Very cool.
So you mentioned that Focal was kind of born out of Stanford AI Lab and looking for, what would you say?
It's not health. Not self-driving cars.
There's got to be something else we can automate and do this with.
What's your own background? How did you wind up at Stanford?
And did you go in... to AI and everything else you were doing with this inkling towards retail in mind, or did that just kind of come up?
What's your story? Yeah, so myself, I was a math, mechanical engineering undergrad, and then went to Lockheed Martin doing bliss and missile defense.
I thought it was the coolest job on the planet until I... actually dealt with the bureaucracy and things like that.
And I was in this engineering leadership development program that actually forced me to apply to grad school.
And after my year three, you have to go in through, in your last rotation, you have to actually apply.
And everyone goes to UPenn or Rutgers or Columbia or somewhere like that, because it's nearby and you can still work.
And then I just threw out an application to go to Stanford.
I got in and my boss said, yeah, we can't fund that because you'd have to quit and then come back.
And I was like, okay, well, I'm just going to go anyway.
So I quit Lockheed, went to Stanford. And my background... before that was basically on Kalman filter theory, optimal reduce state estimators, basically taking noisy signals from radars sitting in boats in the water. on DDGs and CGs, and then giving very accurate, from that very noisy input, accurate track data.
So you can go hit it with a missile or otherwise.
And so that played perfectly into a project that Fei-Fei Li was working on.
And she kind of asked me to come in to do research on some of the stuff that she was working on.
So slam stuff and seeing this location map. mapping technologies like that.
And then deep learning hit and it was like, you know, stop every single piece of research that you could be possibly working on. in that lab and then apply deep learning to it was really the message.
And it was right. And it was, you know, I remember when the results came out from AlexNet, everyone was like, you know, they faked the results.
You know, we had a bug, something like that.
It was out of this world. And so very much it was right place, right time.
I left Stanford with a double-E master's and then went to work at Apple doing a whole bunch of clandestine stuff and using deep learning computer vision on satellite imagery data and other types of data like that.
And it was really cool. And I think I loved my time at Apple, but that was not where, if you wanted to get really, really good at deep learning, it wasn't the place.
Right. And so I went back to Stanford, what was going to be for a PhD in computer science, to continue on this research.
And then I ended up leaving just with a master's in computer science.
I still owe one course, technically, actually, to be dean of admissions.
Let's be above board here. I have to take Intro to AI.
It's literally the class I have to take.
Irony, man. Irony will always get you. if you're listening.
And then we took a project course, the founders of Focal, and it was basically take your research and go and apply it to a real business problem.
And we did so, you know, we had this idea, we had a thesis where it's like, okay, take all the industries, and measure the beta to decrease an SG&A to EBITDA.
That's by far retail, by far. I mean, there's no other place that has such a discrepancy in terms of retail ratio of SG&A to revenue.
So that's where we chose retail. And then after a long period of humility, of realizing that we got to...
Changing an old industry from their ways, you can't do it yourself.
You can not. So if it would not have been for Amazon Go, we would not be here today.
But Amazon Go hit. We kind of were early in the wave.
We're sitting in the water. And there's just no, it's just still, it's like crystal glass, you know?
And then like, and then all of a sudden now you just bigger and bigger swells every single six months.
Right. And we were kind of out there earlier.
And so we've been focusing on that. So, you know, you've referenced Amazon Go a couple of times and kind of the difference between, and I don't know if it boils down this way, but what I was hearing was kind of the difference between a smaller environment where a cashierless system would make sense, and then these larger stores where it just doesn't make sense for a variety of reasons, clean cost, et cetera.
Mm-hmm. As you think about where Focal's going, where the industry's going, are we looking at kind of just a –
Matter of choice based on the parameters, if you're going to go, you know, full on cashier list kind of thing, or if it's more of a contained, you know, focal scan, for instance.
Or is the industry ultimately headed in one direction in how they're using computer vision, deep learning, and all this stuff?
Well, I don't think the unit economics even makes sense for a full-scale automation.
Like, unfortunately, even if you wanted to do it, it would be so much cheaper to just use people.
And this is like the Tesla growing pains where Elon famously said, it's amazing how good people are.
People are hard to beat in certain tasks so far.
Robotics isn't there yet. You know, so so in that pie chart of like, you know, we have shelf integrity, we have cashiers and we have stocking Stocking is not getting automated.
I've worked on this problem. I can't get a robot to pick up a thing and not crush it or drop it.
So we're not there. They're not restocking the shelves.
So that part, even Amazon is not going to be able to do automated yet, maybe for the next 10 to 50 years.
The cashier piece, it's the way to automate that 100%.
You have to choose your selectively permeable layer.
So if that's the shelf edge, that means you need cameras on the whole shelf and they need to be very high res. for you to do it, that is very expensive.
Both install and compute. So I don't see that as a real solution for anyone that's above 1,000 square foot. stores.
It doesn't have a giant arm that deploys compute for 50% of the market. take Amazon just completely out of it, right?
So take Kroger and Walmart and ShopRite and Stop and Shop and blah, blah, blah. it's just not on the, it's not in the cards.
Like I can show you, like prove that in math, like that it's just not going to work.
That it'd be cheaper to use people. And that's going to be true until the price per kilowatt hour comes down by 10x, the price per GPU comes down by 10x, the power draw comes down by 10x.
And speaking with folks in the video is not happening.
That's not going to happen. So what else is in the feasible set that's going to be below people in terms of cost?
And that's the solutions that we're trying to build.
The AIFI and Standard Cogs, We've worked on those problems.
They're going to face the same issues, I think, the AFI—
I think it's pretty cool. And I think I met Steve a few times.
We talked on stage at NRF, which is pretty interesting.
And I haven't seen his pitch deck until then.
If they're... model is to be a giant vending machine, and you plop those in airports and stuff like that, and there's no cashier.
And it's the Boingobox model just in the United States.
If that's it, you know, I don't know. You can do the unit economics.
But the nice thing about a vending machine in general... is that there's very low operating costs.
If they can get the cost down where the price per kilowatt hour, the price per GPU, and the price per these units per month.
If you can get that down, then it becomes a pretty interesting business.
But it's only for these very small pods, which are a very small selection of products.
Yep. You know, there's the bodega guys.
They have a new name now. I forget what they're doing, but it's the same kind of concept, just even smaller.
The word bodega isn't in vogue anymore? No.
We're done. My mind's blown. I don't want to hear about it.
Yeah. I forget their new name, but it's the same kind of thing.
It seems like an interesting model if you can get the cost down.
But again, it's not going to supplant the need to go to a grocery store.
No, right. For certain products. There's just...
There's a reason why Walmart and Kroger won over the rest.
Way back when, it's because they said, well, we want more selection.
What's the famous Bezos quote? It's like, what's not going to change?
No one's going to ever say, hey, I don't really like... as much selection.
I don't want lower prices. So that's what you need.
And part of that is huge amounts of selection.
And so that's where you need to have a giant mall place, super center, where you get all your stuff done.
So I see that to be the model ongoing. I don't see that going away.
If folks want to follow the future of all this stuff wherever it's headed and track what Focal Systems is up to, where's the best place, best places to go?
Well, our website, focal.systems, is a good place to start.
It's focal.systems. You can try focal.systems.com.
It actually does come to us anyway. Or email us.
Almost everyone's first name at focal.systems.
So you can ping us and let us know what you think.
Perfect. Francois, good luck. You've clearly got a lot you're thinking about in terms of where this is all headed and this intersection of the customer and the business ops and the tech.
So best of luck to you and your team on it.
And I'm looking forward to my time in line at whether it's safe or whatever being shortened. because of focal scan.
So I'm going to let you get back to work.
All right. Thanks for having me. Thank you.
