The design and supervision of Sephora’s artificial intelligence models are provided by a central team of 12 people integrated into the Digital & IT department of the retail subsidiary of the giant LVMH.
Thanks to these skills and profiles of product owners, data scientists and ML & Data engineers, Sephora develops machine learning solutions aimed at improving the customer journey and optimizing operations.
A Dataiku platform on Google Cloud
A catalog of existing use cases has been designed. The applications are then rolled out gradually and locally to Sephora’s various country subsidiaries, 15 in total. This existing one is also maintained and improved incrementally.
To improve performance in its Data Science operations, the company’s factory has redesigned its organization and technological platform. Machine Learning developments have been carried out since 2017 on Dataiku.
However, the use was previously reserved for data scientists. Now, the platform is shared with the company’s data engineers. This decision was justified by the desire to shorten the production cycle of the algorithms.
By bringing together development and engineering, Sephora believes it has increased collaboration, including with business teams, and has also accelerated deployments. The factory highlights in particular the gains in terms of data exposure and sharing of model results – via dashboards and APIs.
For the design and deployment of its AIs, Sephora uses a data platform hosted on the public cloud. The retailer is a customer of GCP (Google Cloud Platform) services, including BigQuery and Kubernetes clusters.
From the V cycle to agile development on AI
In addition to the technological stack, there are organizational and methodological developments. Since 2022, the Data Science team has switched from V cycles to agile (Scrum). A CI/CD (DevOps) chain has also been set up.
In order to shorten the time it takes to deploy models to the various subsidiaries, Sephora is following an approach based on ‘core models’ combined with a reduced starting perimeter. This is then enriched incrementally via short sprints.
Always with a view to industrialization and acceleration, Sephora implements the principles of MLOps in terms of monitoring and alerting. For each model in production, technical and business metrics are reported. The time spent on maintenance has also been halved.
These various changes allow the company to deploy an existing use case in a new country in a few days, compared to several weeks or months previously. A strategic gain to deploy a catalog of 20 use cases.
The improved time-to-market frees up time for the design of new models, the development cycle of which has also been shortened. Creating a new machine learning use case went from 12 months to 4 months.