The bridge between supply and demand
How Amazon’s Delivery Experience team acts as a concierge for customers.
On Prime Day earlier this year, shoppers purchased more than 250 million items worldwide from the Amazon store. Each of those orders entailed a series of choices, including deciding when and how they want an item delivered. Making that choice might require just a couple of clicks, but it sets in motion a complex and constantly adapting process, one that is made seamless with help from the Amazon delivery experience team (DEX).
The DEX team’s job: balance the information coming from Amazon fulfillment centers around the globe (which items are stored where and in what quantities, plus the relevant logistics) and the information coming from customers (which items do they want and by when) in a manner that allows those two ends of the spectrum to interact efficiently.
“The DEX team is a bridge that connects supply and demand,” said Siwei Jia, a DEX team principal research scientist. In its simplest form, the DEX team is responsible for helping customers find delivery options that best suit their needs, while simultaneously ensuring those goods arrive as efficiently as possible. Between those two points is a sea of complexity that is both vast and, in some ways, unpredictable.
Jia likens the DEX team to an incredibly knowledgeable wait staff, albeit one working at a 24/7 restaurant with a massive menu that constantly changes over time.
“In one direction is the kitchen, which is our inventory control and logistics, they have all these plates prepared, and DEX talks to the kitchen and we know which items are available and how soon we can deliver them to the table,” he explained. “But there's another direction as well, which is from our customers. They read the menu, and they choose what they like. And what they like can change over time, or vary by season. DEX acts as an advocate for customers, communicating preferences and desired shipping speeds so fulfillment centers can adjust to meet those needs.”
The restaurant simile is evocative, but there isn’t a restaurant in the world that operates at Amazon scale.
“Every day we have huge numbers of packages to be delivered and our customers are highly heterogeneous,” Jia said. “We have customers everywhere, we have customers in dense metropolitan areas, we have customers in remote rural areas.”
And no restaurant on Earth has anything quite like Prime Day. “Prime Day is a great opportunity to showcase Amazon’s customer obsession in all dimensions, and providing earth’s best delivery experience is one of them,” Jia said. “At every point in our customers’ shopping journey, DEX science tries to optimize delivery information and options to help our customers quickly get what they want.”
At every point in our customers’ shopping journey, DEX science tries to optimize delivery information and options to help our customers quickly get what they want.
Each time one of those customers makes a decision about shipping speed, that choice acts as a multiplier that involves layers of additional variables, including how to actually deliver the items.
“If you think about one product, first we need to decide whether that product is able to go to a certain place. Then we need to try to figure out how it gets there — and there might be 20 different routes,” explained Tian Chen, a DEX senior applied science manager. “So, we need to call all this inventory information and the routing information to make sure this one product can arrive at a certain destination at a certain speed.”
The DEX teams relies extensively on machine learning to manage that process. DEX team scientists have created machine-learning models that do everything from helping make recommendations that reduce the number of boxes a customer receives by consolidating shipments, to optimizing package delivery speed. The models range from deep neural networks that leverage hundreds of features to make predictions, to causal regressions that conduct what-if analyses, to dynamic stochastic optimization such as contextual bandits.
“We rely, for example, on machine learning to do predictive delivery speeds,” Chen explained. “In particular when serving speed as a feature for ranking models, where there is strict latency requirement. For any delivery speed higher than two days, there are so many variables that making an estimate can be difficult, so we built a model to predict those speeds.”
For any delivery speed higher than two days, there are so many variables that making an estimate can be difficult, so we built a model to predict those speeds.
To further complicate matters, the DEX team also must account for the unknown. That includes shifting preferences, fads and trends, and weather.
“In our machine-learning models, from the scientific point of view, we need to address this highly heterogeneous situation and be flexible enough to respond to all these outside shocks,” Jia said. “At the same time, our system must be robust enough to handle all the demands.”
As an example, during the holiday seasons in the past two years, DEX implemented machine learning solutions to assist customers in discovering items that would be delivered by Christmas, helping to proactively improve the customer shopping experience.
Jia said contending with both known and exogenous events requires a comprehensive approach.
“We tackle this from three perspectives,” he explained. “One is scalability. Our platform needs to have the right flow from data to machine learning to implementation to address different situations. The second one is a decision-making mechanism, which is largely driven by machine learning, but includes human audits. The third is exploration-exploit methodology: For some things we try, history doesn't necessarily tell us anything. It might be totally new. So sometimes we just need to experiment to determine whether something works or not. If it does, we double down. If not, we quickly rewind.”
That customer obsession leads to applied science that has measurable real-world impact.
“Whenever we implement a solution, we immediately see how customers respond, and this gives us a great sense of fulfillment,” Chen said. “One of the examples is at the mobile checkout page. Previously, customers saw some default delivery options, and all the other delivery options were folded into another page. We saw a clear gap in how customers selected options on mobile versus desktop. So, we adjusted to make the other options more apparent on mobile. Addressing those pain points is what we do.”
Jia said the scale and variety of challenges make DEX an exciting place to work for scientists with a passion for solving challenging optimization problems.
“I’ve been working in DEX for two years and I feel it's getting more and more exciting,” he said. “DEX is a concierge service for our customers. We’re going to help a customer through the entire journey. Once they select a product on the store, we make sure packages arrive to their doorstep on time, and in the most efficient way possible. We’ve only begun to scratch the surface of what’s possible in getting customers the product they want in the shortest time frame, and most effective means possible. The opportunities for business and customer impact are almost endless.”