Shopping online is nothing if not efficient. With a few clicks, just about anything can be found and comparison shopped for a low price.
Jet.com is taking things a step further. The New Jersey startup uses AI to optimize the cost of entire shopping carts, rather than simply tallying up the prices of the individual items inside.
With the online shopping market forecast to double in 2024 from an estimated $390 billion in 2016, according to Forrester E-Commerce, there’s billions of dollars at stake.
Acquired by mega-retailer Walmart last year, Jet reduces costs for baskets of items by assessing a variety of merchants, distributors and shipping costs. Its goal: find the right combination of the items so that the total order cost, including shipping and any commissions, is as low as possible.
The bigger the shopping cart, the larger the savings it can generate. However, optimizing the contents of a cart can easily spin into an almost incalculable number of combinations too time consuming to compute.
Daniel Egloff, managing director at QuantAlea, which consulted with Jet on new pricing software, shared how they’re tackling this “basket-optimization” problem, during a talk at the GPU Technology Conference earlier this year.
Mind the Gap
Calculating all possible combinations to find the minimum overall price of a basket of less than a dozen items would take years using traditional algorithms and a CPU-based “brute force approach,” Egloff said.
Switching to a GPU sped things up on traditional algorithms, but not enough. And the slightest delay between choosing and buying a product can drive consumers away.
“We want the time to be seconds, not days, or years,” he said.
Jet solved the problem using a clever combination of machine learning, new algorithms and GPUs in the Microsoft Azure cloud. Moving processing to the cloud allows Jet to scale its compute resources dynamically and cost-effectively, Egloff said.
The team designed their algorithm to find a merchant combination that fulfills a shopping cart at the lowest price, and to be fast enough to provide real-time results from millions of calculations.
Computing the optimal order fulfillment using these methods leads to significant savings on orders with multiple items, Egloff said. This adds up fast for consumers, and can boost sales for retailers.
The backend of the Jet platform is developed in the F# programming language. The GPU algorithms are coded in F# and compiled to CUDA with QuantAlea’s Alea GPU, which simplifies and speeds GPU programming. The deployment is done on Azure N-series virtual machines powered by Tesla GPU accelerators.
AI Podcast: Looking to Add Some Hot Dogs to Your Cart?
Drop in some buns and mustard while you’re at it. We’ve got the inside story on the latest gag on the show that everyone is buzzing about this year.
Learn more on the AI Podcast, where developer Tim Anglade shares his experience developing “Not Hotdog,” an app that spawned as a joke on HBO’s Silicon Valley TV show, but is now available on iOS and Android stores.