Afresh Co-Founder Nathan Fenner On How AI Can Help Grocers Manage Supply Chains - Ep. 208

0:10Hello, and welcome to the NVIDIA AI podcast. I'm your host, Noah Kravitz. Our guest today is Nathan Fetter. Nathan is co-founder and president of Afresh. Afresh is using AI to help grocers in the face of skyrocketing food prices, cutting costly, environmentally unfriendly fresh food waste, and yielding measurable impact in grocery stores nationwide. Before co-founding Afresh, Nathan taught at the Stanford Engineering School and worked extensively in Silicon Valley as a robotics engineer. So I think we'll have a lot to get into here technology-wise and also in terms of the future of technology, food, you know, all that basic stuff.

0:47So let's get right into it. Nathan, thank you so much for joining the NVIDIA AI podcast and welcome. Thanks for having me. Really excited. So if you would, why don't you start us off with telling the audience a little bit about what Afresh does, what the mission is behind the company. So Afresh is really a kind of supply chain and inventory management solution for the grocery industry with a laser focus on the fresh side of the business. And I'll say fresh over and over again, probably throughout this conversation. And when I talk about fresh and grocery, that really refers to the kind of perimeter of the grocery store.

1:24So we're talking about the produce department, the meat, seafood, deli, bakery, food service, dairy, and things, basically anything that can perish and go bad. So that's what we do. We build supply chain inventory management software. And really our mission is to eliminate food waste and by making the supply chain more efficient, really increase access to fresh food, which we believe is, you know, really healthy and can kind of nourish the world. How long has the company been around? We incorporated in 2017. And that was kind of at the conclusion of a year or so work, kind of pre-company work that Matt and I were doing together when we were in business school.

2:06Got it. So six, seven, eight years, somewhere in there kind of sense at first. So we started afresh really on the backs of two critical observations. The first insight we had really specifically about the retail and grocery space was that we came to believe that the future of this massive, massive industry, it's over a trillion dollars in the US alone. We came to believe that the future of this industry really was fresh food. And this was based partially on The data, you know, again, this was 2016, 2017. The growth in this industry was really coming in the fresh departments. Health and consumer trends were getting people to eat more fresh food, less processed food.

2:46But it was deeper than that. It was more strategic than that. With the kind of emergence of the pure play online players like Amazon, fresh becoming this critical strategic differentiator for grocers as they competed with each other, but also with kind of new online players. You can buy a box of Cheerios at really any grocery store. You can buy it online. It comes to your house. It's the exact same no matter where you get it. But the quality of your fresh offering is really what most consumers kind of use to determine where they shop. Right. That makes sense. So that was our first observation.

3:22Like to win going forward, grocers really need to be world class in how they handle their fresh food. So that was observation one. Second observation we had was really contrary to the first. When we were looking at the technological landscape in grocery, and specifically the supply chain and inventory management space, we noticed that first and foremost, and this might be a little techno elitist Silicon Valley startup, we saw that the technology was pretty legacy. A lot of it had been built in the 80s. But beyond just being, frankly, pretty old, the software had really been built for the non-perishable side of the business.

3:58It had been built for things that come in boxes and have barcodes. And when you look under the hood of these solutions, they were solutions that were built for a bunch of retail subverticals. So the same software they were using the grocery industry, they were trying to use to manage the supply chain for car parts, electronics, clothing, just hard goods in general. And when you applied, what we saw is when they tried to apply that same software to the fresh side of the grocery business, it just didn't work very well. It's kind of a death by a thousand cuts things. You know, fundamentally, the tasks are reasonably similar.

4:33But when you get into the weeds, all these kind of esoteric or nuances of the fresh food supply chain start to accumulate and drive problems. So fresh food perishes, obviously. A lot of it doesn't have barcodes. Sometimes things are prepared in the store. Like if you buy a cantaloupe, it gets cut up and put in, you know, a fruit bowl, but also still sold as a cantaloupe. There's upstream supply chain variation. So all these things made these systems that were really built for hard goods really inefficient and fresh. And the consequence of that was a really inefficient supply chain that then drove the problems that you talked about at the beginning.

5:10Really tons and tons of food waste and billions and billions of lost dollars of profits for the grocers. So that's kind of what we saw. The future of the industry was fresh food. The technology wasn't being built for fresh and was causing a lot of inefficiency. Simple idea. Can we apply modern technology in a laser focused way to the most strategic part of the grocery store and in doing so drive differentially kind of great results? I have so many questions about the grocery store supply chain, but I don't want to get us off topic. I'm thinking about going through the self-check and you have a sticker on your red pepper, but then are self-checks actually evil?

5:52even though I was good at them because they're faster, but is that dark design on purpose? You know, all these things going off tangents. What I want to ask you is, so what'd you decide to do about it? Yeah. So, I mean, we decided to start the company afresh to solve this very problem, right? So we came together as a set of co-founders when we were in graduate school at Stanford. Like you mentioned, my background, I'd done my undergraduate and graduate work in engineering and been a lifelong engineer and a lecture in engineering school. but I like to say I went and joined the dark side and got my MBA.

6:25That's where I met my co-founder, Matt. He's our CEO. He was doing the joint program in food and ag as well as the MBA. And then our third co-founder, Vladimir, was finishing his PhD in AI at Stanford. Got it. And so we came together and we looked at this problem, this kind of inefficiency in the supply chain, and we thought to ourselves, obviously, this is 2016, 2017. AI, maybe still really hot in 2016, 2017, not as hot as it is right now. But it just struck us as a perfect problem for applying this emergent technology that where we could just drive genuine, tangible kind of pragmatic impact.

7:06And so we formed our company to apply cutting edge machine learning in a really pragmatic way to solve this very tangible problem. And so did you start with one specific sort of sub-problem within the problem or were you taking kind of a, you know, top-down sort of look at the whole industry approach? How did it, how'd you start working? That's a, I mean, that's a great question. We start, the first problem we sought to solve was, it's, you know, getting back into the little bit of the grocery lingo was the store, store level replenishment challenge. So basically, the job we were trying to improve was how do you figure out how many tomatoes to order into your store each day?

7:46And so the traditional way that grocery has worked for decades previous is a produce manager in store is responsible for figuring out how many tomatoes they pull in from their distribution center into their store. The U.S. grocery has a pull model like that. And the tools at their disposal, again, were really sparse and required a ton of manual intuition. Are you working primarily with grocers of a certain size or kind of across the spectrum of, you know, smaller markets, big, you know, mega? I don't know what the word is in the big box equivalent in grocery lingo, but you know what I mean. Yeah, totally.

8:24Our grocers run the gamut from about 25 stores to, you know, close to 2 ,500 stores. Right. We tend to work with grocers who own their own distribution center or have a certain amount of scale that enables us to offset the cost of the kind of data integration, upfront data integration to make the kind of ROI for both sides make sense. Right. Makes sense. Yeah. I think as we continue to mature as a company and we'll be able to figure out how to onboard grocers in a more scalable way, we can open up parts of the market, those kind of mom and pop shops or independents and things like that, who may not have the scale to be able to leverage or basically the scale to kind of offset the integration costs.

9:07So anyway, I guess to close my previous thought, that was the first problem we tackled was how do we optimize that decision? How do we build a tool that enables grocers to do that far more efficiently? And yeah, that was our starting point. And I think a kind of a fun stat. So the first department we started in was produce. The first tool we built was store level replenishment. And I think as of right now, about 13 % of all produce that's sold in grocery stores in the U.S. was ordered through our tool. Wow. In about six years or seven years, something like that. Yeah. And frankly, almost all that commercial growth has been in the last three years.

9:48It was product development followed by commercial growth. So it's been a really, really resonant product in the market. That's amazing. Congratulations. Thanks. So Nathan, kind of as an outsider to the food industry, but somebody familiar with buying too much at the store and then watching it go bad on the kitchen counter, you know, the compost bin, I would imagine, obviously, food waste conjures up, you know, images of awful, very direct, like we're wasting food. People need food to live. This is bad. Beyond that, what are some of the other repercussions of food waste that, you know, myself, listeners might not be so aware of?

10:27Yeah, I think like exactly to your point, right? I think when we all have some guilt, when we throw out extra food at the end of a meal, we think about people who are food insecure and the fact that when we're wasting food, it just feels kind of awful from that kind of at a person level. For those in the know, and I certainly wasn't one of those people seven years ago, but I've become acutely aware, food waste beyond that kind of food insecurity angle is actually really one of the most significant contributors to climate change and our broader kind of climate crisis. There's a really, really cool organization called Project Drawdown that has done really extensive research on what are all the macroscopic kind of initiatives we can take as a society to mitigate climate change.

11:20And they have a couple of different methodologies. It's very scientific. It's very quantitative. and, you know, against a list of over a hundred, a hundred of these kinds of macroscopic initiatives, food waste ranks first and fourth in their two, two methodology. Most impactful thing we can do is to reduce food waste, to mitigate climate change. And so it's really one of the key things that brought me into the business. I think I've always had a keen eye to kind of work in the climate space and it's really motivating for a lot of our team. And I think it's, um, it's a key part of our mission. Is it, as I'm asking, I'm guessing the answer is probably like a mix of different factors, but is it overproduction?

12:01Is it sitting on shelves for too long? Is it, you know, sort of the transportation supply chain? So maybe to answer a slightly different question, food waste is a really big driver of climate change for a variety of reasons. If you think about the percentage of the world's resources that go into creating the food that feeds us, it's absolutely massive. And then in practice, about a third, a rough swag, a third of that food winds up in the dumpster. So you're taking this massive, massive pool of resources. Right. That's really inefficient. And on top of that, when food gets thrown away, it eventually basically emits a ton of methane, which is substantially worse than other greenhouse gases.

12:46But more squarely to your point, Food waste occurs at every part of the journey from seed to fork. It's interesting in developing countries, you see a lot of the food waste happening before it gets to kind of retail, before it gets to the store. It has a lot to do with, you know, less developed cold chain. So, in developed countries like the US, it overwhelmingly happens at the retailer and downstream at the consumer. Okay. And so obviously that's where we put our focus is really making that retailer component of the supply chain maximally efficient. And that has positive knock-on effects for the consumer because the more efficient your supply chain is upstream, the more days of shelf life items have when they get to their consumer.

13:37Right. Yep. So I know that you guys just launched a new inventory tool. We're recording kind of at the end of September, 2023 here. And so a couple of weeks ago as we speak. And I want to ask you about that. But are there, I don't know if highlights is the right word, but some things about, you know, the past three years even of that big growth, but just the work you've been doing and particular problems that you were able to solve or, you know, on the tech side or on sort of the practical, you know, impact side, or even just some things about the industry itself that, you know, you found really interesting or satisfying to work on, you know, from that technical kind of the perspective of applying AI.

14:24Yeah, it's a very broad question. It is, it is. I was, well, again, it's a reverse thing because I was thinking about when you were talking about the replenishing at the store level, I was actually thinking about a video I saw recently on Reddit of the pumpkin pie table at a Costco. And it was like a surveillance camera video. And it just showed this massive table just full of pumpkin pies and how like during the day people would come and buy them and they went down, they went down, then they got restocked and they go down and go down and then they go, you know. And so that was making me think, oh, like at this level with fresh, fresh food, do you count?

15:01You can't count the individual tomatoes, I would think, but there's X number per box. Do you count that way? Are you going off of data from the data center? Are you having real-time data to computer vision? You know, I had a million questions, right? So instead of picking one, I went super broad. So it's, yeah. Well, I think you've probably hit at the crux of the challenge, which I think, I'm not sure it's gonna... When I cast a wide net, I tend to get something. Yeah, yeah. I'm not sure I'm gonna answer, like, I think your question was like highlights, but I'm going to take it in, you know, I'm going to take the breadth and use like have a little poetic license with the perfect, the question here.

15:40Um, but I do think in your kind of your pumpkin, how does the pumpkin replenishment, pumpkin pie replenishment story work? You've hit at the crux of things, which really is the biggest challenge that I think really differentiates fresh from traditional hard goods, non-perishables. I talked about all those nuances, things like perishability things are produced in store but i think the way you can cast like you can kind of characterize all those things is that the data is much messier and fresh and so you know to your to your point about self checkouts like when you have self checking out things that don't have our codes on them are they when they have bananas are they scanning them out as a organic banana or as a conventional banana.

16:28And as a result, if they're scanning out organic bananas as conventional bananas, you have a bad signal about what the actual demand signal is for organic bananas. And you have the wrong signal for conventional bananas as well. And that's just one of the many, many ways that really, really, and I don't know if the listeners of your podcast are like really the kind of people who want me to say the data are, but I'm going to go with data a singular for vernacular benefit here, but the data is so, so much more challenging. And so the really like crux of the problem and the big accomplishment that I think Afresh is really trying to tackle is how do we, how do we bring to bear this cutting edge technology that is machine learning, that is AI into a world where the data is really messy.

17:15And so I think that has been the fundamental challenge and the thing that I think has been our biggest kind of technical achievement at the highest level. Right. Can you share any insight into the solution or the achievements? I feel like I might not be doing the best job of asking because I don't have the technical vocabulary, perhaps, but kind of like some examples of like you described the problems really well, but examples of solutions. Yeah. Let me maybe dive into some of the like the subcomponents of the machine learning models we need to make make work. Perfect. So I'll be a little bit reductive here, but I think some of the like critical.

17:54So, again, let's maybe use the example of our store level replenishment system and what our store level replenishment system is doing in a really kind of, again, being reductive. is it's trying to say how many tomatoes should produce manager X at store Y order tomorrow in order to optimize inventory levels to minimize waste, but also maximize sales and maximize in-stock rate. In order to solve that problem, we use kind of machine learning in a variety of different ways. And each of those kind of subcomponents of the problem requires us to solve this problem of making those components of the system effective in the face of unreliable data.

18:41And so the components of that problem, the most obvious one is there's a forecasting problem. So I think when people think about what we do, everyone, even historically in the industry, kind of as a result of the hard good paradigms that people kind of apply or assume we kind of fit into, the first kind of thing you have to do is forecast demand. right and so you need to figure out how can you accurately predict how many customers are going to buy a certain product over the time horizon that that order covers and that's a really really challenging problem in fresh one because because a lot of like fresh things are kind of commodity driven with their prices their price data is much much much noisier and so you have much much much higher variable demand.

19:33Fresh food tends to be much more promotion driven. You're running these promotions on a weekly basis, whereas the forecasting challenge for Clorox bleach or something like that is really, really smooth throughout the year. There's a very jaggedy demand that's driven by changing prices, promotions, cannibalization, substitution effects, all these things, and then unreliable demand signals caused by things like missed scans at the register. And so we have to come up with really, really sophisticated machine learning models that are able to work in that paradigm. That's one challenge that we've had to tackle.

20:15Then there's a really critical part of the problem is understanding inventory. So in a hard goods world, you calculate inventory in the following way. today's inventory equals yesterday's inventory plus what you shipped into the store minus what you sold out of the store that does not work in fresh all those data inputs are really really problematic uh you have things in additionally going bad yeah some retailers try to scan out what's gone bad that doesn't work at all so you have this mess and right that this is good this This is called perpetual inventory. And perpetual inventory in fresh has been a broken paradigm for a really long time.

20:57Okay. This is why when you order, I don't know if you're an Instacart user, for example. Okay. Yeah, I have. Sure. If you order something on Instacart, why you're plagued with substitutions, right? Because grocers have a particularly bad understanding of what they have in inventory in their fresh departments. And then if you think about what you're trying to figure out what the ideal amount to order is, you need to know how much demand is there going to be. And then you also need to know how much do you currently have to meet that demand. So that's been another really interesting challenge that we've had to tackle, which is building models that are actually, as opposed to using that arithmetic approach that I described to calculate inventory, we're actually modeling what our inventory position is.

21:39So using the understanding things like perishability, understanding things like how unreliable the data is to build a probabilistic representation of what grocers have in inventory and understand also kind of the remaining shelf life of that inventory. So another really, really cool kind of application of call it machine learning, call it modeling, call it AI that we're doing to really crack this unique, the unique challenge that is fresh, that all emanates again from that data problem. The third big area I would just talk about, again is basically the actual output decision of these systems. So forecasting how many tomatoes are going to sell, that's not a decision.

22:23The decision is how many cases of tomatoes are you going to order from your distribution center. And that task of taking that demand forecast, taking that inventory position and turning it into a decision, that's where we apply a lot of the more recent kind of decision-making style algorithms to really make sure that we're optimizing those decisions in the face of all this uncertainty that's caused by the challenges in the data. And so this is where we're drawing on fields like controls or things akin to reinforcement learning and some of the decision-making algorithms that you see in things like self-driving cars.

23:04Again, because what the output of our system is, it's not a forecast. The output of our system is a decision. It's a decision, right. So this is where we're pulling from a lot of that emergent algorithmic, those emergent kind of algorithmic approaches of these other fields that are really pushing the bounds of AI forward. And we're able to leverage that and enable us to make better decisions in the face of uncertainty. I'm speaking with Nathan Fenner. Nathan is co-founder and president of Afresh, a technology company that's using AI to help grocers and the grocery industries writ large get a better handle on fresh food.

23:43Everything from forecasting, inventory needs, to things in the supply chain, all kinds of stuff related to how grocery stores refresh the tomatoes, all the other foods. It's this whole time, Nathan, we've been talking. There's a couple of grocery stores I go to in the real world regularly. And the one is just in my head. And I'm thinking, like, I'm going to have to go there now and, you know, just to just to check in with reality. Right. But it's fascinating stuff you guys are doing. And we were talking off air that we mentioned earlier that coming from a background in robotics, obviously a lot of overlap with, you know, the sort of engineering, computer science, you know, math, algorithmic level stuff that you're doing, I would imagine.

24:23But from the outside, it seems like a fairly big shift. What's that experience been like for you? Have there been, you know, similarities, parallels? Do you tinker with robots in your spare time? What's the move been like? Yeah, I think your last part there that you're like, it's a pretty big shift is spot on. Okay. I would really say that my background as an engineer, when I went back to Stanford to go to business school after I'd been out kind of as a, as a engineering industry, it empowered me to just actually just take, I was technical enough that I was able to take as many machine, upper level machine learning courses as I could.

25:06Like they say across the street is what we say at the Sanford business school, across the street at the engineering school. And so it really, I think it just gave me enough kind of false confidence that I could masquerade as a computer scientist and a machine learning grad student enough to take enough classes to kind of understand things conceptually to then be able to form, I think, high level kind of theses around the business and then hire way, way, way, way, way smarter and better people to actually go implement those things so like i said pulling velatomir onto our team uh and then the kind of amazing set of talent that we've been able to amass i think there's a through line from my engineering background to that but i think it's not because i have some technical it's not a technical through line it's more of a really there's a there's a soft skill is a weird word but you know something to be able to surround yourself with good people is a skill in and of itself for sure other engineers just kind of, they see you, like when you're an engineer and even if you're a different type of engineer, we kind of, we have a knowing nod.

26:18Yeah. I think that's most, most of it to be perfectly honest. So to kind of end on a, on a forward looking note here, a couple of weeks before our taping this in, in late September here, Afresh launched an inventory management solution on addition to the platform. Maybe if you want to take a second, kind of go over what that is, but maybe in the context of, we didn't talk too much. I think it's sort of understood and it was less important to our conversation, but we didn't talk too much about what the end user sees and kind of what, you know, what the solution is there. So maybe how the new tool fits into the Afresh offering.

26:58Yeah. So I think the notion of an end user, it's probably worth kind of immediately clarifying that which is the the vision for a fresh has never been about points a point solution so i talked earlier today about our store level replenishment tool but that was really our first thing we built and we really really expanded our platform uh since then and the vision for what we're doing really is about a fully integrated platform that helps you optimize decisions at each node in your kind of fresh operations. And so different, like, obviously as a result, different end users have different experiences.

27:39The release of our inventory management solution is kind of a continue, like continuing on that journey to optimize in a joint fashion, additional nodes in the supply chain. So that specific tool is really, really exciting. It helps grocers essentially with their, there's a practice in grocery that's really common where at the end of a certain amount of time you take up an inventory of what you have in store and so our this is our solution is kind of an intelligent way of doing that saves users a lot of time but more importantly it generates much much much more accurate information about what the inventory in the store is and that serves as an input into our ordering solution, but it also creates this really invaluable source of data, which is a more accurate understanding of what's available in store.