Pon Automotive - Building an AI Personalization Framework to Ensure a Smooth User Experience
Names:
Helmer Koppelman
Thomas Helling
Sander Berkers
Country: Netherlands
Organization: Pon Automotive
Pon Automotive excels in (luxury) automotive brands, commercial vehicles, and numerous innovative mobility services. We are a market leader across a large number of markets. We represent the automotive brands Volkswagen, Audi, SEAT, ŠKODA, and CUPRA, and luxury automotive brands, such as Porsche, Bentley, Bugatti, and Lamborghini in the Netherlands (primarily). In addition, we are active in the luxury segment in the US. In the Netherlands, we also represent the brands MAN (trucks, buses) and Volkswagen Commercial Vehicles (light commercial vehicles).
Every day, more than 1.8 million Dutch drivers rely on our passenger vehicles and associated services.
Awards Categories:
- Moonshot Pioneer(s)
- Most Impactful Transformation Story
- Most Extraordinary AI Maker(s)
- Partner Acceleration
Business Challenge:
User experience is everything. At Pon Automotive we strive for a seamless customer journey, from landing on our homepage to receiving your car keys and driving off into the sunset. Personalization of our website is the key starting point of this journey. Are you interested in a sports car? Or in an SUV instead? Our website will recognize your preference and give you a tailored experience.
Personalization as a concept sounds intuitive and straightforward – the implementation was everything but. We needed an adaptive and contextual model that can predict what version of the website a user prefers seeing. User preferences change quickly. Today’s most popular model can be out of fashion tomorrow. Modeling all of this as a set of fixed business rules is an insurmountable task that rapidly grows in complexity with the number of car models and brands that are included.
Instead, we developed an AI model, a deep neural network reinforcement learning model predicting a user’s preferences. The model runs as a containerized app in the cloud and interacts with the website. When a new visitor lands on the homepage and gives their consent for personalization, the AI framework is informed of the visitor’s interaction with the website. Upon reaching the homepage a second time, the website will be personalized based on the stored interactions. To keep the model informed of the latest changes in user preferences, we designed it using reinforcement learning principles. The AI framework adapts and learns about the latest changes in user preferences several times a day.
To enable building, monitoring, and maintaining the AI framework we used at least 10 different tools and platforms. Numerous steps were required to push new updates to production. Implementing modifications was slow. What started off as a simple AI project quickly turned into a complex and intractable product.
Business Solution:
Moving the entire AI personalization framework to Dataiku has been bliss. From CI/CD pipelines and model monitoring to publishing the model as an API endpoint, everything is now managed within Dataiku.
The first step is building the data archives using automated scenarios. Every two hours the latest user interactions are retrieved from the data storage and processed using Dataiku recipes. Then comes the periodic training of the model together with hosting the new version as an API endpoint. Again automated scenarios are used to trigger the model training. The API endpoint, through which the website interacts with the model, is updated automatically after each model training phase.
The migration to Dataiku and the automation functionality that comes with it decreases the time spent doing MLOps and, in turn, enables us to focus on implementing new features. One such feature is a functionality to boost certain types of personalization depending on the active marketing campaigns. For example, when a certain type of car is actively advertised using channels other than our website (e.g., television, social media) we can now choose to boost the personalization for certain types of users and display the campaigned model instead.
Business Area: Analytics
Use Case Stage: Built & Functional
Value Generated:
After clicking on a pair of red boots do you also get suggested 10 different other types of red boots? Did you just buy a chair? Here are 10 suggestions for other chairs you might like! How often have you been frustrated by simple recommender systems or a buggy website? Personalization is a true art, and a smooth user experience is invaluable.
The website often is the first interaction we have with a customer. The true value of a smooth and personal experience is hard to measure. Useful feedback about how visitors experience our website is hard to get by. Moreover, cars are slow-moving products. Impactful changes to our website might only translate into more sales after several weeks to months. Tracking customers between the time they land on our website and when they visit a car dealer is near to impossible.
Nonetheless, there are several direct measurables that we can compare. For example, in terms of the click-through rate and completed car configurations we see an uplift of up to 100% compared to a standard and non-personalized website. And in the number of planned test drives, we are seeing an improvement of 70%!
The self-learning nature of the AI model greatly simplifies the life of marketers. They no longer have to develop and maintain business rules to personalize the website. With the time gained they can focus on optimizing the marketing campaigns.
Value Brought by Dataiku:
Dataiku has greatly aided in making the AI personalization framework tractable. By automizing what can be automated and pulling the entire framework to one platform, we have enhanced the team’s efficiency. As a result, we have been able to shift our focus from managing and maintaining to optimizing and improving.
Together with Dataiku data scientists, we pushed the boundaries of Dataiku. The implementation of our AI model required some customization. It does not fit into the standard templates because it requires multiple types of input together with a feedback loop to make it adaptive. Some functionality, like monitoring the model, is not yet available and will be included in newer versions of Dataiku. Because of our fruitful collaboration we were able to migrate the personalization framework in record time.
Finally, Dataiku opens up exciting new possibilities. For example, we plan to monitor the model together with the other AI products that we created. Having a single platform that hosts all models significantly improves our ability to do proper model governance. Moreover, we plan to extend the framework and add more personalization options to the website. Finally, having the framework on Dataiku enables us to roll out the current personalization options to other websites too. Currently, it is in production for a single brand, and we plan to add other brands in the next period using the same API service.
Value Type:
- Improve customer/employee satisfaction
- Increase revenue
- Save time