Nordstrom CTO on AI-powered analytics and consumer insights

Customer service is Nordstrom’s bedrock, and that wasn’t going to change just because the retail giant was engaging with customers online and not in the store, Edmond Mesrobian, Nordstrom’s chief technology officer and chief information officer, said during VentureBeat’s Transform 2021 virtual conference event today. In conversation with Hari Sivaraman, head of AI content strategy at VentureBeat, Mesrobian shared how Nordstrom revamped its data infrastructure in ways designed to enhance customer experience.

When customers are in the store environment, it is easy to have a compelling personalized experience, thanks to the attentions of a stylist or a salesperson. It is more of a challenge to connect digitally when the online presence tends to be more of a cold and distant experience. This difference forced Nordstrom to take steps to find out how to become closer with customers.

“We had to figure out how to get analytical data to make predictions that help us make additions, so that we can actually offer the most compelling experience to our customers, whether online or in-store, and that’s the journey we embarked on,” Mesrobian said.

Building NAP

Nordstrom Analytical Platform (NAP​), the analytics platform that Nordstrom employs, is a real-time, event streaming–centric analytical platform that provides insights on everything from customer services to credit. Reflecting on the start of this project, Mesrobian said that the most important thing is to remember that presenting the information is not about reporting data, but about collecting events that can be translated into actions.

Nordstrom starts by getting the business events created, organized, and streamed in real time. Then the multi-layered analytical models translate the events into predictions that ultimately lead to customer benefits and services. NAP employs open source technology and cloud computing, stitching existing components together to create an analytical platform that drives machine learning in a robust way.

“It’s off-the-shelf commodity components with a lot of special sources, which essentially translate our business events into a semantically rich object, something that could then be actioned through a model,” Mesrobian said.

Before adding the sources, however, Mesrobian needed to make sure that the data was clean and readable to use. That is, the business events had to be completely and accurately represented to start with. As a result, when those events arrive at a point of being transformed into high-level objects, they are accurate and searchable.

“We want to transfer the responsibility from the data engineers in the past … to the application owners to make sure that their business events were pure, curated, and correct to begin with, and [they are] also aligned with their business goals,” Mesrobian explained. The shift in responsibility, he said, is key to the platform architecture process.

With more than 100 AI models leveraged daily at Nordstrom, an end user will indirectly see the benefit from the timely delivery of information. While users have control over their privacy and permission settings, Nordstrom collects user preference information and employs it to provide better selection, better dynamic looks, better style boards, better choices, and so on.

Through AI, the company presented a more robust way of driving discovery and personalization. Nordstrom also launched a fashion map effort to take a natural language-based approach with deep learning so that the model gets to know customers better through conversation, as opposed to taxonomy-driven keyword searches.

In terms of the various AI technologies that Nordstrom uses in building NAP, Mesrobian listed examples such as linear classifier algorithms for computer vision. Those algorithms generate attribution vectors about the product based on pictures, and those vectors are used to drive dynamic loops. The company also uses computer vision algorithms that take shots of customer closets and give them insights about what’s there. Tree-based models are also used to detect abnormal signals and fraud. “Name it, and we are probably using it,” he said.

Lessons from Nordstrom

When building an analytical platform, Mesrobian pointed out that it is important to keep in mind that data is not the crucial aspect, since the end result is to present predictions and leverage them in real-time events. Building from that broad recognition, the three takeaways from the platform include:

1. Make sure that it’s analytics so that the platform actually makes predictions rather than presenting raw information.
2. Get the business aligned so that the model speaks to business events.
3. Make sure that the systems are wired from the store with privacy and security in mind, so that the analytics take into account both the business event and the privacy impact of customers.

When talking about data privacy, Mesrobian said that developers need to be clear about customer proposition and about what information is being collected. “We are very mindful of those regulations to make sure that [the customer] is always aware of what we are collecting, and that you have access to that information,” he said. In the end, the data needs to create accessible value for the customer, he concluded.