Everli Cart Assistant

Reading Time: 6 minutes

Your AI-based grocery helper

Have you ever dreamt going in a supermarket and finding all your favorite products ready for you on a single shelf?
Are you the kind of person who’d like to get the grocery done as fast as possible, and move onto more interesting stuffs?
Or would you like finding all the things you usually buy ready for you, and spend more time exploring offers and new products?

Well, here in Everli we have designed an AI-powered feature that meets all these needs at once.
It is named Everli Cart Assistant and it works like magic 🤩.

The origins of Cart Assistant

How did this new feature come to light?
Everybody at Everli is obsessed with just one thing; we want to deliver customers peace of mind with a trouble-free grocery experience that they love and trust.
This has been our mantra and main driver behind the creation of Cart Assistant; we wanted to delight customers with a feature capable of:

  • Making their grocery process faster and smoother
  • Letting them feel Everli knows what they like and what they don’t
  • Simplifying the grocery process and leave them more time to explore the platform

  • The first signals we could build something like this came up in April-May 2022, thanks to the research work done by our Data Science and Product team around customer grocery behaviors.
    From that point on, the wheel started spinning, and it could not get stopped 💪.

    Everli Cart Assistant timeline
    We came up with a first prototype of cart assistant engine during our 2022 Everli Firebreaks (click here if you you want to know more about Firebreaks) and from that it took us another couple of months to get to to the final solution.
    Proper AB testing of the feature took place between April and May 2023 and it was a huge success, leading to a significant uplift in conversion rate for test users 😎; finally, the 15th of May, we released the feature to the full customer base.

    Cart Assistant in action

    That’s enough with the history lesson, how does Cart Assistant effectively work 🤓?
    By intelligently analyzing customers’ grocery history, it identifies their preferred items, and collects them into a “warm up” section, ready to be used by customers. From this page, customers can rapidly start a new grocery by adding their favorite products to the cart and, once they are done, either move to the checkout or continue buying by browsing other products on the store.

    Cart Assistant in action
    To facilitate the access to the new Cart Assistant page, a new, ad hoc widget has been created within the Store Home Page, and strategically positioned to capture users’ attention.
    Cart Assistant widget: is there anything promising more “peace of mind” than “grocery in a flash” ⚡?

    Cart Assistant core engine

    In the previous paragraph, the expression customers’ favorite products has come out, but what does this actually mean?
    In the context of Everli Cart Assistant, it represents:

  • products a customer will quite likely buy when she starts a new grocery
  • a set of additional items Cart Assistant thinks would fit customer needs

  • The first set is created by analyzing the overlap across contiguos groceries done by a customer in a given store, and by retaining items with high frequency of occurrence.
    The second set is obtained by looking at the overlap across groceries done by a segment of similar customers (see next paragraph for more details), in the same store, and by keeping items with high relevance for such segment.
    The final recommendation for customers is the union of the two sets:
    the bulk of the recommendation being retained by the first term in the equation, while the second one works more like a “fill up” term.

    How can Cart Assistant effectively identify these two sets of products? This is possible thanks to the smart recommendation engine our R&D team has developed in the past months, and around which the whole feature then evolved. The engine, powered by Machine Learning algorithms, has been devised to identify relevant patterns from users’ grocery history and use them to come up with specific, customer tailored suggestions.
    One nice feature of such engine is its ability to intercept behavioral changes as time goes by, so that its recommendations are always up to date and on spot 😍.

    All this sounds great, but what happens when Cart Assistant has to deal with a new customer, who has not done enough groceries to provide a history to learn from?
    Don’t worry, our R&D team found out a solution also for this type of edge cases 😉.
    Even if Cart Assistant doesn’t know what a new customer would like to find in his cart, it can still infer some of his preferences by looking at customers similar to him who have already matured some grocery history. For example, it could look at customers:

  • with similar age
  • living in nearby areas
  • using the same type of device
  • and, by inspecting their grocery histories, detecting common patterns from which a rose of potentially preferred items for new customers can be built.
    Cart Assistant and new customers: learning tailored suggestions by looking at similar users

    DynamoDB: a new approach to improve site speed

    An important goal releasing this feature was the ability not to impact site speed and loading performances. This is a key goal for us, as we are really obsessed with providing customers with the smoothest possible user experience.

    For Cart Assistant, we were able to explore a new way to serve Machine Learning models in our marketplace, using DynamoDB as storage. DynamoDB is a noSQL database available as AWS cloud service allowing low latency access.
    How has it been used?

    Cart Assistant flow via DynamoDB

  • Every day a training job of the Cart Assistant algorithm takes place
  • Sets of customer-store level recommendations are built
  • The latter are then stored into a DynamoDB table that can be called on the backend of our marketplace

  • Why is this better than providing a standard, real time endpoint? Well, compared to the former, the DynamoDB solution is 10 times faster in terms of retrieval time 😎.

    Cart Assistant: to infinity and beyond

    Is this the end of the story? Of course not 😄!
    The first version of Everli Cart Assistant proved to be a great success, but there’s still room for improvement and, why not, the introduction of new features on top of it.
    Are you curious to see where this is going? Then stay tuned for the next upcoming episodes 😁.

    Join the Everli tribe!

    Are you also interested in delivering peace of mind to customers?
    Would you like to work for a company whose goal is to become the most loved European marketplace?
    Do you want to be part of a group of talented people, who are always pushing the boundaries to deliver the coolest features, like Cart Assistant?
    Well, then you should probably have a look at our open positions.

    Ciao! 🚀🚀🚀

    Authors: @Fracappo87 @MariantoniettaDG @mattiau

    Leave a Reply

    Your email address will not be published. Required fields are marked *