Transform 2022: How enterprises crawl, walk, then run into their AI/ML deployments

SAN FRANCISCO – Enterprises don’t stand up AI/ML deployments overnight, and when the decision is made to do so, it involves most of the C-level leadership of the company and a lot of recruiting for qualified data analysts and scientists. It also involves an evolution that can be likened to a person learning to crawl, walk and then run.

None of this is easy or simple, but it’s becoming necessary in this third decade of the 21st century. Companies are learning to crawl, walk and run when it comes to utilizing their data in order to give them deeper insight into their protected business data, all the extraneous data that’s in storage coffers but not accounted for, and all their historic data. Don’t forget all the social networking and outside data (customer opinions, product reviews, etc.) that float around in the gigantic universe that is the internet and affect a company directly or indirectly.

At VentureBeat’s Transform 2022 conference here at the Palace Hotel, a panel consisting of Fiona Tan, CTO of Wayfair; Rajat Shroff, VP of product, DoorDash; Kevin Zielnicki, principal data scientist, Stitch Fix; and moderator Sharon Goldman, senior editor and writer, VentureBeat, discussed how their automated AI/ML processes are providing scale and speed to market. Their paths all eventually took them from proof-of-concept to production in sustainable ways. 

DoorDash’s approach

“At DoorDash, one of our values is that we dream big but start small,” Shroff said. “We apply this to our AI efforts as well. We’ll start by using manual means to do unscalable things to learn and understand how to find product market fit. Once we see the signal, that’s when we start inventing algorithms and start scaling this up. 

“For example, when we did our analytics, we found that only about 8% of our business was delivering pizza. Some of us thought it was maybe half our business. We realized we needed to be much more accurate in our assessments, so we got the team together and said ‘We’ve got to get to 99% precision.’ After a few months of manually annotating data gathering, the team found a small sample (identifying a market, a category). Once they got some signal, they expanded the whole project. Once they got to a level of precision they liked, that’s when they handed it over to the ML team. And they started building (AI models).”

After a few months of building the team and the deployment, DoorDash went from 60% accuracy in analyzing its business to its goal of 99%, Shroff said. 

How Wayfair is using AI/ML

“We started our (AI) project by looking at the accessibility and quality of data available for the problems we were trying to solve,” Tan said, “so we wanted to make sure we had the ingredients to apply to our AI/ML project. The second consideration we wanted to know was ‘How much tolerance do we have for faulty predictions?’ So the first place we decided to go with our project was in areas within Wayfair that could tolerate faulty predictions. 

“For example, we want to use our AI deployments in (Wayfair) marketing and advertising bidding. The worst thing that could happen there is that you pay too much for an ad, right? It was an area where I thought we could learn and lean in and get a quick turnaround on results. It’s a little bit harder using analytics to determine the quality of an item in our catalog; we wanted more humans doing that.”

Stitch Fix specializes in personalization

Stitch Fix specializes in matching its customers with items of clothing and accessories, so its recommendation engine makes a lot of use of AI and ML, Zielnicki said. “This is very important to get right when you’re sending people a box of things that you think that they’ll like when you try them on at home,” he said.

Stitch Fix has integrated AI and ML into every facet of its business, Zielnicki said. 

“The problems can be as diverse as deciding which warehouse to ship out of, the ‘pick paths’ within those warehouses, choosing which stylist to match with which client, assembling items out of sets of items, and so on,” Zielnicki said. “When we started 10 years ago, we had very little data about our items, our clients. We started with some simple popularity-based systems, then moved to some standard statistical models – things like multilevel regression that work well with relatively small amounts of data. As we gathered more data about our clients and got more of a history built up, we evolved into doing collaborative filtering approaches, matrix factorization, and most recently a sequence-based model that is based on the sequence of interactions a client has with us across their journey.

“This all adds up to a more personalized experience for our clients.”
VentureBeat Transform 2022 continues virtually through July 28.