The climate crisis is one of the most important and complex challenges of our age, and solving it will require collaboration, innovation, and commitment. According to Project Drawdown (a non-profit organization that functions as a top resource for climate solutions), one of the key drivers of climate change that we can meaningfully address as a society, is food waste.

In today’s episode, we learn about Afresh, a company that is leading the way in providing food waste solutions to grocers across America by creating optimized food orders through pioneering AI and machine learning solutions. You’ll hear from Afresh Co-Founder, Nathan Fenner, as we discuss the founding mission behind the company and how they are leveraging AI in a way that is fundamentally different from other established legacy companies in their field. We discuss the challenges of working with perishable products, how it results in noisy data, and why it’s so important for Afresh technology to not only provide predictions but also make decisions in the face of uncertainty.

Today’s conversation unpacks a particularly exciting area of AI and demonstrates how advancements in the field are paving the way for impactful climate solutions. Be sure to tune in to learn about the real-world impact of AI innovation in an area where we need it most urgently!

“We're hyper-focused on building supply chain software to optimize all the perishable supply chains in retail. The big outcome of optimizing that supply chain is that we dramatically reduce food waste. Food waste is one of the biggest macroscopic contributors to climate change.” — Nathan Fenner

“Good machine learning is key to writing an optimal order that maximizes profit, but also minimizes waste.” — Nathan Fenner

“All the technology that had been built for the grocery industry, and that was being used in supply chain and inventory management, had all been built for the non-fresh side of the business. It had all been built for things that come in boxes that have barcodes.” — Nathan Fenner

“We leverage AI in a fundamentally different way. We definitely do forecasting, but the critical thing we're doing is really decision-making under uncertainty. The output from our models is actually a decision as opposed to simply a forecast.” — Nathan Fenner

“Leveraging this more frontier area of machine learning has allowed us to make really good decisions in a really uncertain environment.” — Nathan Fenner

“If we can build a technology that reduces food waste by 50%, it will become uneconomic for grocers to not use our technology (or a similar technology) that produces that much in cost savings.” — Nathan Fenner

Transcript: [INTRODUCTION]

[00:00:02] HC: Welcome to Impact AI, brought to you by Pixel Scientia Labs. I’m your host, Heather Couture. On this podcast, I interview innovators and entrepreneurs about building a mission-driven, machine learning powered company. If you like what you hear, please subscribe to my newsletter to be notified about new episodes. Plus, follow the latest research in computer vision for people in planetary health. You can sign up at pixelscientia.com/newsletter.

[INTERVIEW]

[00:00:33] HC: Today I’m joined by guest Nathan Fenner, President and Co-Founder of Afresh, to talk about eliminating food waste. Nathan, welcome to the show.

[00:00:41] NF: Thanks for having me, Heather. This is really fun.

[00:00:43] HC: Nathan, could you share a bit about your background and how that led to you to create Afresh?

[00:00:47] NF: Yeah. I can definitely share about my background. I’m not sure it’s a really linear progression into starting Afresh, but my background is as an engineer. I studied engineering in my undergrad and in my master’s program. Then I went on to become a robotics engineer working here in Silicon Valley, and contemporaneously, while I was working as an engineer, I was also a lecturer in the engineering school at Stanford. Reasonably technical background, but tying that all to Afresh, I would just say that – and really my focus was as a hardware engineer.

Trying to tie that technical background to what we do from a software perspective at Afresh, I think the real through line that connects that background, the company, is really that I’ve always really believed in one, working with really cool technology, just as a technologist I find that really exciting. Two, being really passionate about working at companies that did unequivocally have positive social impact, so not a company that’s contribution to society was purely adding value to like our GDP, but rather something that in its own rights made a positive impact for the world.

I had a particular bent towards mitigating climate change and the company I worked at after grad school before starting Afresh was really focused on ocean health, which is a big barometer for climate change. That’s really, I think, the through line to what we’re doing at Afresh today.

[00:02:16] HC: What does Afresh do? Why is it important for sustainability?

[00:02:20] NF: Yeah. So Afresh, at its most simple, is a supply chain software and inventory management software for the grocery industry, focusing specifically on the fresh side of the business. In other words, we’re hyper focused on building supply chain software to optimize all the perishable supply chains in retail. The big outcome of optimizing that supply chain is that we dramatically reduce food waste. Food waste is actually one of the biggest macroscopic contributors to climate change.

I think there’s a really cool study done by an organization called, I think, Drawdown, Project Drawdown, where they looked at all the macroscopic initiatives that we as a society could take, and using a variety of different methodologies, reducing food waste was either the top or one of the top three or four initiatives we could undertake as a society. Our work contributes to that reduction in food waste, which in turn is a really critical mitigation of climate change.

[00:03:30] HC: What role does machine learning play in your technology?

[00:03:33] NF: Machine learning is really, really fundamental to what we do. At a high level, we’re making the fresh food supply chain in retail more efficient, and the way we’re doing that is by optimizing decisions at various nodes in the supply chain. The way we optimize those decisions is powering them with machine learning. So to be very concrete, our first product that we built at Afresh is a store level replenishment solution for fresh departments and grocery stores.

This is a solution, and it’s based on an iPad application. The person in the store in a produce department in a grocery store uses our application to do their inventorying and ordering workflow. Our solution essentially gives them recommendations of how many tomatoes, for example, to bring in into their storage data in order to optimize their inventory levels. That recommendation of how many tomatoes to bring into their store, that’s all powered by machine learning. So really, really good key, machine learning is really key to writing an optimal order that maximizes profit, but also minimizes waste.

[00:04:51] HC: What kind of data are you working with in order to train this? It sounds like maybe the stock of different products in the store, or is there other data you collect along the way?

[00:05:00] NF: Yeah. I think a nice way we think about the data that feeds into our system. I think there’s three main sources. There is data that comes from the retailers, from their central data warehouse. Our retailers send us all their sales data, all their shipments data, their item meta data, basically their data that describes their items, their promotion calendar, their delivery schedules, basically all the sources of data that you could possibly need that they have as internally in order to inform the correct decision. That is that internal bucket of data.

I’ll add that for that data, we get a big historical download of all that data. We get a huge several years of sales history. Then we also set up an ongoing data feed where we’re constantly being updated with the most current version of that data. That’s one source of data. The second source of data is external sources of data. This would be things like holiday calendars, local events, weather, basically exogenous factors that contribute to what the perfect order would need to be in the store. For example, the Super Bowl is coming up in a couple of days here, we have that information available to us to know that avocado sales may increase dramatically this weekend. That’s an example of an external data source we use, we leverage.

Then the third data source is data that we collect via our front end. In conjunction with our smart machine learning, we have a really intuitive workflow that our team has built. As our end users use our tool, they’re actually collecting data from the store in terms of things like inventory position in the store and the size of their displays and other key pieces of information that will inform what that perfect order is. Then basically those three data sources come together and those all inform our machine learning models that then kick out the ideal recommendation. Is that makes sense?

[00:07:10] HC: It does. Do you need to annotate your data in order to make these predictions or is that taken care of via the different data sources you have?

[00:07:20] NF: It’s not a data annotation challenge like a computer vision problem from a training data perspective. Instead, I think really the challenge for us is when we initially integrate with a customer, we need to ensure that we are correctly interpreting their data and have all the rules and understandings of their data to convert it from their schema into our schema. Once we accurately write those transformations and it’s ingested into our system, this is not the thing where you need to like annotate that this is a car in this picture. The things we’re training on are basically just sales dollars and things like that. There isn’t this big undefined data source. Once those rules, those ETL rules are created and validated, the data is well understood by our system.

[00:08:19] HC: What sorts of challenges do you encounter in working with those three different sources of data?

[00:08:23] NF: I think it’s come back to the fundamental thesis of our company, which I guess I really haven’t gone over, which is that we started the company basically with two critical assumptions and this is going to be a little bit meandering and I’m long-winded, but hopefully this will be a valuable context for the question eventually. When we started the company, we had two foundational insights about the retail grocery industry. One was that we really thought that fresh was the strategic future of this industry for a variety of reasons that I won’t go into on an AI-centric podcast.

The second insight we had is that all the technology that had been built for the grocery industry and that was being used in supply chain and inventory management had all been built for the non-fresh side of the business. It had all been built for things that come in boxes that have barcodes. When you look at a lot of these solutions that are out there, legacy solutions out there, a lot of them had been built as horizontal solutions that served other retail subverticals like clothing, electronics and such.

The fundamental challenge of using those systems in the fresh departments is that those systems expect fresh or perfect data, because you have non-perishable goods, hard goods, things with barcodes and that’s just not true for all the items that happen that exist in the fresh supply chain. For example, in fresh departments in grocery, you have things that don’t have barcodes that are scanned out at the register and the person who’s checking you out needs to enter a code for them. What will often happen is if you buy an organic avocado, the person who’s ringing you up might think it’s a conventional avocado.

The consequence of that is that your sales data, your demand signal is now incorrect and there’s a variety of other complexities here. These items go bad, they perish inherently by being fresh so they’re thrown out so your understanding of your inventory position degrades. Some of these items are cut up and produced in stores. So you order pineapples into the store, but pineapples are also sold as chunks of pineapples, pineapple bowls and things like that.

All of these factors and there’s more upstream supply chain fragmentation as opposed to getting all your cereal boxes from general mills or craft. You’re getting them from 27 different small farms around the country. All these factors basically lead and this is all back to your original question of really, really, really noisy and uncertain data. This is I think at the heart of why we exist as a company which is that in order to succeed in this area, we needed to figure out how to deal with this challenging data. I think that is really at the core of what’s differentiated us from the legacy companies that have come before.

I think your original question is what are the sorts of problems that we have with the data? I guess to answer it now with that extensive prologue. The problems we have with data are that the data are really, really noisy. There’s tons of issues that hide the true demand signal. There’s tons of issues that make it really hard to have a consistent through line for a given item. A strawberry might get scanned out as 27 different items, because they’re not linked to a parent item. Just a variety of these things make it really, really, really, really hard to run good machine learning models. I think with a traditional approach, you basically just have a classic garbage and garbage out situation.

[00:12:14] HC: What do you do about that, then? Do you need to clean your data extensively or do you need to use alternative algorithms that are more robust to these challenges?

[00:12:24] NF: I think it’s all of the above. I will say that when we talk about our approach to machine learning, our approach is really paradigmatically different from a lot of the legacy players in the space. I think what you see with the existing solutions in the market is basically using data science algorithms to do forecasting, to predict demand. When they claim to leverage AI, which may or may not be true, they’re talking about it in the context of improving their forecast. We leverage AI, I think, in a very fundamentally different way. We definitely do forecasting, but the critical thing we’re doing is really decision making under uncertainty.

The output from our models, the ultimate output of our model is actually a decision as opposed to simply a forecast. I think that is like the foundational answer to your question, is that we’re taking a different algorithmic approach that is basically the inherently acknowledging of the fact that the inputs are uncertain. This really emanates from the PhD research that my co-founder Volodymyr did when he was at Stanford in the PhD program there. One of his areas of expertise is really decision making under uncertainty.

Leveraging this more frontier area of machine learning has allowed us to make really good decisions in a really uncertain environment. Again, that’s paradigmatically different from the approaches out there as supposed to just an evolutionarily better forecasting model. That’s not to say we don’t do lots of work on cleaning up the data, lots of work with our front end to prompt users and intelligently engage humans in the loop, but I think at our most core from a machine learning perspective, we’re just taking a entirely different approach to the end output of the system. Did that make sense?

[00:14:24] HC: It does. Yup. Are there any specific technological advancements that made it possible to build your technology now when it wouldn’t have been feasible a few years ago?

[00:14:34] NF: Well, I think this area of machine learning, specifically decision making under uncertainty, I think a lot of the principles have been around for a long time, but I think it’s an area that’s gained, that’s made a lot of headway in recent years. Maybe it’s not, it’s less technological advancements and we’re probably getting a little over my skis here, but more about a new frontier of knowledge and research and approach that has really started to unlock this approach to solving the problem.

We’re leveraging a lot of the techniques that they’re trying to leverage in autonomous cars, right? Basically, which obviously similarly are taking a ton of inputs, all of which have a probability distribution around them and needing to make an optimal decision. Now, obviously, autonomous cars are leveraging a bunch of other inputs that are really required technological advancements, but I think my assumption would be that a lot of the pushing forward of this frontier of decision-making uncertainty has been critically enabling for us to be able to do this. Does that make sense? I think it’s more about a knowledge frontier moving forward as opposed to necessarily a specific technology, it’s not like the advent of GPUs was critically enabling for this or like a certain computer vision technique, but rather just the boundaries of our knowledge as a scientific community.

[00:15:58] HC: Yeah. Like the theory and the algorithms related to decision making under uncertainty. Yeah.

[00:16:03] NF: Exactly, right. Better said than me.

[00:16:06] HC: How do you measure the impact of your technology?

[00:16:08] NF: Yeah. That’s pretty easy. I think what’s really cool about what we do is that it’s really, really measurable, particularly when we operate pilots at our partners. So we go in and install one of our solutions and we have a set of stores, for example, who are using our solution and a set of stores that aren’t. We get to measure really, really concrete financial impact. We look at how much shrink we are avoiding, and shrink is the term they use in retail to measure food waste. We also measure how we are impacting sales.

The really, really cool thing about our technology is that by optimizing inventory levels, we actually will increase the grosser’s sales and decrease their food waste, which is I think a lot of the times in the industry, people thought that was an impossible tension that either you had to stock your shelves fuller to push sales, but if you did that, you’d increase your food waste or you had to run much more lean at the expense of sales. What we’re seeing with our technology that really optimizes inventory is that we’re able to achieve both.

After we measure that impact of sales and food waste production, we then can plug those, especially the food waste production numbers into calculators and methodologies that enable us to understand our climate mitigation impact. There’s an organization called ReFED, which is the national leader in terms of really understanding food waste, and they have a methodology of how to convert food waste into marker, some of the critical sustainability metrics like a greenhouse gas emissions, water saved, etc. so we convert our food waste into those sustainability metrics.

[00:18:01] HC: One of the challenges that’s common across the industry is hiring for machine learning, because these professionals are very much in demand right now. What approaches recruiting and onboarding have been most successful for your team?

[00:18:14] NF: Yeah. It’s nice to be on a podcast where that’s like a well understood problem, because I think one of the challenges we have, a challenge we have is differentiating our genuine machine learning from the marketing claims of legacy companies. I think there’s a congruence there that it’s pretty much impossible to hire machine learning engineers. So it’s really, really suspect that legacy, these legacy companies that don’t have a real, I don’t know, a real carrot for these highly sought after, highly paid engineers would be able to credibly hire the talent they need to be able to build genuine cutting edge technology that they claim otherwise.

Anyway, to that question, I think overwhelmingly, I think we’ve been able to punch above our weight in terms of our hiring. That all comes back to being a genuinely social impact driven company, I think as everyone’s aware, especially in this, the tightening of the tech economy, a lot more and more people are turning towards climate tech right now. As a climate tech company, I think that’s been a real tailwind for us. I realized for people listening to this, that’s probably not either you have a genuine social impact mission, or you don’t, and it’s probably not really useful advice. I think that’s overwhelmingly the biggest tailwind for us.

I would say in particular, the fact that our business model, you make our customers more profitable by reducing food waste. So there’s this perfect alignment of incentives that allows folks on our team to I think fully embrace our mission in a way that they don’t have to be skeptical of it as greenwashing. So I think having that really genuine, really, really tight alignment of business model and social impact adds a credibility to that mission driven nature of the company in a way that really, really resonates with this like the most competitive job, the most competitive persona to hire. I realize it’s not maybe the most useful advice, but I think it’s been incredibly impactful for us.

[00:20:35] HC: It’s something that you make use of in recruiting and make sure in your job postings that candidates are aware that you’re focused on sustainability and climate change and all that.

[00:20:46] NF: Yeah.

[00:20:46] HC: It helps the recruiting process.

[00:20:48] NF: Absolutely. I think when you hire people as a result, if a person doesn’t care about our mission, they’re not going to go work for us. They’re going to – we’re not going to be able to compete with Facebook, with FAANG companies, Apple, Amazon, Netflix, Google. If they don’t care about the mission, if they’re pure mercenary, they’re going to go take a higher salary elsewhere. That’s fine. What that means is that there’s this virtuous cycle here where the people who come to work at Afresh are really, really passionate about the mission. The consequence of that is that when people interview, I think the authenticity of our team’s belief in the mission really comes through in that hiring practice. As a result, I think people really feel it. Then other missionaries join the company as well. So it’s really virtuous in that respect.

[00:21:35] HC: Is there any advice you could offer to other leaders of AI powered startups?

[00:21:39] NF: Then maybe the generic piece of advice I would give. I think one thing we’ve done well is being really business problem oriented. I don’t think we were like, oh, we’re good at AI. Thus, like we have this AI hammer and everything’s a nail. We didn’t look for a problem to solve with AI. We looked for a genuine business problem to solve and then figured out what was the best way to solve it.

Then it happened to be that AI was the best way to solve it. I’m not the first one to wax poetic about this, but I do think there is a tendency to be a little bit technology oriented in the products we build as opposed to solution, product, and business problem oriented. I think there’s probably an excitement right now to figure out ways to leverage AI. I think that’s the wrong way to go about building companies that add really genuine value. I think you have to start with real business problems.

[00:22:37] HC: Finally, where do you see the impact of Afresh in three to five years?

[00:22:40] NF: I’m really, really excited about our potential as a company. I think right now our partners are reducing their food waste by about 25% in the departments in which we’re active. So we’re currently this year expanding across all fresh departments. When I look three to five years out, I really think we can get to a level of food waste deterrence that is really on the order of 50%. So I think retailers working with us can reduce their food waste by 50%. I think that’s to me a really great North Star for us. What I think is really cool about that is that the net margins in the grocery industry are so slim that if we can build a technology that reduces food waste by 50%, it will become uneconomic for grocers to not either use our technology or a similar technology that produces that much in cost savings.

What I think we can do is drive a real movement. I’m not saying it will be us, but I’m saying we’ll force the hand of the industry to be able to really reduce the food waste by about 50% across the supply chain. When we back that out, that equates to tens of billions of pounds of food waste each year and just a dramatic impact on climate change more broadly. I think we really have the opportunity to lead the market and force the hand of the market towards this far more efficient future and state that’s really better for and consumers and better for the planet.

[00:24:27] HC: This has been great. Nathan, your team at Afresh is doing some really interesting work for fresh food and sustainability. I expect that the insights you’ve shared will be valuable to other AI companies. Where can people find out more about you online?

[00:24:39] NF: I guess probably start with our website afresh.com. Then I think we have a bunch of resources on Afresh, a lot of our blog posts where our machine learning engineers go into way more detail and speak way more accurately and eloquently about what we do. Then I’m sure, I don’t think we have much of a social presence. We have some on LinkedIn, but yeah, I would start at afresh.com and highly recommend our engineering blog. The folks on our team do really cool stuff.

[00:25:05] HC: Perfect. Thanks for joining me today.

[00:25:07] NF: Thanks Heather. It was super fun.

[00:25:09] HC: All right everyone. Thanks for listening. I’m Heather Couture and I hope you join me again next time for Impact AI.

[OUTRO]

[00:25:19] HC: Thank you for listening to Impact AI. If you enjoyed this episode, please subscribe and share with a friend. If you’d like to learn more about computer vision applications for people in planetary health, you can sign up for my newsletter at pixelscientia.com/newsletter.

[END]

Resources for Computer Vision Teams:

LinkedIn – Connect with Heather.
Computer Vision Insights Newsletter – A biweekly newsletter to help bring the latest machine learning and computer vision research to applications in people and planetary health.
Computer Vision Strategy Session – Not sure how to advance your computer vision project? Get unstuck with a clear set of next steps. Schedule a 1 hour strategy session now to advance your project.
Foundation Model Assessment – Foundation models are popping up everywhere – do you need one for your proprietary image dataset? Get a clear perspective on whether you can benefit from a domain-specific foundation model.