From kibble to clicks: Chewy’s path to the Edelman Award competition

Joshua Hale:
It sounds simple from a customer perspective, right? You go on the website, place an order, and then that order arrives at your door. Even from the replenishment side, it sounds simple where just buy products at certain locations. However, to answer this question at Truly requires coordinating over a hundred thousand SKUs for over 20 fulfillment centers, buying from thousands of vendors and executing over a half a million customer orders at a daily basis. Where everything is uncertain, customer demand fluctuates, vendor performance is inconsistent, operational constraints changes. And what especially makes it challenging is that all these decisions are interconnected.
Ashley Klimp:
Welcome to Resoundingly Human, the Informs podcast. I’m your host, Ashley Klimp, and in this special mini series, we’re highlighting the finalists for the Franz Edelman Award, often referred to as the Nobel Prize of OR and advanced analytics. Each year, organizations from around the world demonstrate how advanced analytics and OR are solving real problems at massive scale, from supply chains and healthcare to public policy and logistics. Joining me today are members of the team from Chewy, the largest online pet retailer and pharmacy in the United States. Their project tackles a challenge that sounds simple on the surface but becomes incredibly complex at scale and that’s replenishment. Using a science-driven suite of models supported by a custom engineering platform and continuous analytics monitoring, the team transformed how Chewy manages purchase orders, safety stock and inventory placement all while accounting for unreliable data, supply uncertainty, and vendor constraints.
The result, improved inventory replacement, fewer split shipments, shorter shipping distances, and meaningful financial impact in a thin margin industry. So let’s kick things off with introductions. I’d love for you each to introduce yourselves and including your role in Chewy and this Edelman Finalist project.
Joshua Hale:
Yes, I can start. So I’m Joshua Hale. I’m a staff research scientist within Chewy’s supply chain replacement organization and I focus on developing optimization models that support our inventory systems throughout our fulfillment network. And my main contribution to the replenishment ecosystem is that I led and developed a safety stock model, which is the model that deals with uncertainty to determine the inventory buffer to combat against that on certainty.
Erdem Eskigun:
Hello, I’m Merdem. I’m Merdemet Higan. I’m currently leading the team with the replenishment ecosystem at Chewy. I’ve been working for Chewy for four plus years and the main reason why actually I came to Chewy was because of this opportunity of rebuilding this replenishment ecosystem from ground up and you wouldn’t get that kind of opportunity every time, everywhere. So this was a really exciting opportunity for me to join Chewy and be part of this exciting development. So I’m really happy that our work has been also recognized by this prestigious at Umanaward. On a personal note, I have a PhD in industry and engineering from Purdue and have been working in industries since then solving practical problems.
Ashley Klimp:
Well, it’s a pleasure to meet both of you and thank you so much for taking the time to join me on the podcast today. Let’s dive into the challenge itself. When someone orders pet food or treats online, they probably don’t think much about the logistics behind the scenes. What made replenishment such a complex problem for Chewy?
Joshua Hale:
Yeah, so it sounds simple from a customer perspective. You go on the website, you place an order, and then that order arrives at your door. And even from the replenishment side, it sounds simple where you just buy products at certain locations. However, to answer this question at Chewy requires coordinating over a hundred thousand SKUs for over 20 fulfillment centers buying from thousands of vendors and executing over a half a million customer orders at a daily basis where everything is uncertain, customer demand fluctuates, vendor performance is inconsistent, operational constraints changes. And what especially makes it challenging is that all these decisions are interconnected. Where we decide or where we predict demand happens affects inventory placement and inventory placement affects shipping and shipping affects costs. So even a small decision where placing inventory in the wrong location may cause a profitable item to be a loss. So replacement isn’t just about buying product.
It’s about orchestrating an entire supply chain under uncertainty and especially in a pet retail industry where the margins are relatively thin. The way that we answer this replacement problem is very important.
Ashley Klimp:
Your solution involves a suite of analytics models that influence everything from purchase orders to safety stock and inventory placement. Can you give us sort of a high level look at how these models all work together?
Joshua Hale:
Yeah, so we didn’t want to build just one model to answer this complex question. So instead, we wanted to build a coordinated system of decisions and those decisions is comprised of three components. One is allocation. So there’s a disconnect between where our forecasts are generated and the level of decisions that we have to make. So our forecasts are generated at a geographic region level, but we have to make decisions at the fulfillment center level. So we have an allocation model that can translate those regional level forecasts into a fulfillment center level signal. So that’s kind of the first step. And then the second step is we have to deal with the uncertainty because forecasts are never perfect and vendor’s performance aren’t always perfect either. So we have to build a safety stock model that can combat against that uncertainty. And that model is comprised of two components, but the first component is in charge of constructing the distribution around the era that consists of the error in the forecast and also the error in the vendor lead time.
And then the second is we take those distributions and maximize an expected profit to determine the service level of each item within our catalog. And then lastly, we have a PO generation optimization model that consumes both the demand signal at the fulfillment level and also the safety side target to generate POs while respecting vendor constraints like truck minimums, MOQs, and also operational constraints like capacity. So basically our model flows from the allocation to uncertainty to execution where each layers feeds the next one.
Ashley Klimp:
So one thing that stood out in the project description was a system’s ability to correct unreliable data and continuously monitor analytics performance. Why was building that capability into the platform so important?
Joshua Hale:
Yeah, so we didn’t want to build a model that is frozen in time. So we wanted it to evolve with time. So we embed analytic monitoring directly into the system. So this allows us not only tweak certain parameters of our models, but also redefine underlying logic if we need to. So one example is we saw that safety stock was increasing week over week and analytical monitoring provided us the insight that it wasn’t necessarily the vendor delays, but it was inbound congestion that was causing the problem. So this insight allows us to create a smoothing procedure that allows us to dampen that response in order to not artificially inflate safety stock. So that is just one example of how analytical monitoring allows us not to be stuck in time but also evolve as the environment around us changes.
Ashley Klimp:
All right. Now let’s talk about the results. Your solution improved inventory placement, reduced split shipments, and shortened shipping distances. For customers ordering from Chewy, what does that improvement actually look like in practice?
Erdem Eskigun:
Yeah, this is a really good question. Chewy has been known for its outstanding customer experience and customer service and we have internal goals to meet that customer demand with great customer experience. Basically that means we deliver customer orders in less than two days, but at the same time we try to keep Chuvi financially healthy as well. And Chu is also known for its high level subscription customers and we need to make sure that those subscribed orders arrive at customer doorsteps at the right time with the right contents in each box that we deliver to the customers. Majority of the times our orders include multiple SKUs and we need to make sure that every SKU that inside within a customer order are actually put in the same box and we deliver to the customer within one single box. But even with the subscribed orders, customers can add different SKUs on top of their subscribed order skews and that makes it the order multiple SKUs order.
So beforehand though, we don’t know what SKUs are going to be ordered in each customer order and we need to make sure that each SKU that in each customer order actually is available at the nearest fulfillment center from the customer addres. So that brings the buying and placement as a complex problem that how do we ensure that each skew that the customers will like the order are available as an inventory in our nearest fulfillment centers. So that’s a really hard problem to solve and we try to address with our ecosystem to solve such problems. If you don’t solve this problem, well, one might argue that for example, hey, why don’t you just buy every SKU everywhere and then you will make sure that you are going to have that inventory for every customer order skew. But let’s say if we do that, then we have to buy a lot of inventory in the network and tie a lot of capital in our network and not only the capital, but also we don’t have enough space to put all that much of a high inventory levels in our fulfillment centers.
So that’s not really a practical problem. So the real problem that or real solution would look like, how do we actually optimize and first predict accurately what the customers will likely order even in the multi-skis ordering and how do you optimize the buying and the placement and make sure that we have the right skills at the right FCs and to deliver to the customer orders in the least possible shipping costs. So that’s what we address with our replenishment ecosystem and that’s what practically the customers will see better customer service. And also because of the savings that we would actually enable, now we have the ability to add more selection to our catalogs and also make sure that the customers get the best price from Q website
Ashley Klimp:
So projects like this don’t happen in isolation. What did the collaboration look like between data scientists, engineers, and the business teams that rely on these decisions every day?
Erdem Eskigun:
Yeah, so this requires a lot of collaboration and teamwork together. Our product managers work with business teams to get business requirements in place. And then once we have versions of the model, they take it back to the business teams to understand whether those business use cases are fully satisfied or not, or we need to do more iterations around it. And they work with our scientists and engineers to make sure that those business requirements are implemented correctly. Our scientists work together to make sure that not the models we develop are budget efficiently but also work coherently together so that we have an efficient ecosystem in place. Our engineers work with the scientists and analytics teams to make sure that we have the right platform to be able to run our models efficiently, but for production runs, for daily production runs, but also for the experimentation scenario analysis that we need to do for business use case purposes.
Our analytics teams make sure that they work with every team member, make sure that we have the right metrics and the right analytical tools available to monitor the performance of our ecosystem and also how the business are being impacted based on the decisions we take every day. So that in a sense, in short, basically for the last three plus years, we had to work in collaboration as a team very closely, but even after we put this ecosystem on, we kept doing this, this teamwork and collaboration to keep up with the system we developed as well.
Ashley Klimp:
The Edelman Award celebrates analytics that make a real world impact. When people hear the story of your project, what do you hope they take away from it?
Joshua Hale:
So for me, I hope people do not take away that analytics isn’t just about building a better model, but it’s about building a better system. And while we developed this specifically for Truly, many industries face similar challenges. So this system level approach that we took on still remains powerful. So I hope that people see that as more as just this project is more as a truly story, but also a blueprint for how operations research and analytics can be embedded into the daily operating system of an organization.
Erdem Eskigun:
Yeah. Let me add my perspective on this one as well. We hope that others get really excited about the problems that we are trying to solve here or we solve in the industry and they get motivated to join us to the industry scientists or the science teams, I would say. And the biggest advantage of working in industry is we have ton of data and that will enable us, that would enable us to find out the opportunities to solve real practical problems, improve our cost structures, and more important, improve our customer experiences as well. So a lot of opportunities in industry and we hope that this kind of recognition also would highlight that kind of cool problems that we are working on in industry.
Ashley Klimp:
These are very cool problems. Finally, what does it mean to your team to be recognized as a Franz Edelman Ford finalist?
Joshua Hale:
So for me, the Franz Element Award represent real world impact at scale and that’s what motivates our team to be recognized alongside global leaders in operating research is incredible, meaningful. It’s especially meaningful because the project supports millions of pet parents and behind every optimization decision is a real customer and a real animal depending on us. So we’re proud of the work and grateful to represent Truly and the broader operation research community.
Erdem Eskigun:
Yeah, we are really proud of our work and great really to see that our work is being recognized by this prestigious award. We are really happy to be here.
Ashley Klimp:
Congratulations again to you both and to the rest of the team from Chewy on being named a Franz Element Award finalist. The winner will be announced that the upcoming 2026 Informs Analytics Plus Conference where all the finalist teams will present their work during the competition and the winner will be announced at the Edelman Gala. To learn more about this project and the other finalist teams, visit resoundingthehuman.com. And to all our listeners and viewers, be sure to like and subscribe wherever you watch or listen to the Resoundingly Human podcast.