St. Petersburg Bank, together with GlowByte, personalizes communications with customers in a mobile application using an RL model

2025: Application Implementation of Model to Personalize Customer Offerings

BSPB, with the support of GlowByte, has introduced a model for personalizing offers for customers into the Bank's mobile application. The solution is based on the Reinforcement learning technology. This approach allows you to form an individual plan for displaying banners to each client: now the Bank will be able to conduct interactive marketing actions that will allow you to create customer-relevant offers. GlowByte announced this on February 13, 2025.

The RL model was designed to automate marketing. The model has strengthened the ensemble of previously used ML models that analyze customers' inclinations towards various products of the Bank. The RL model provides an optimal combination of variety of offers and maximizing customer interest. Unlike classical approaches, Reinforcement learning technology allows the model to constantly learn more from new incoming data and adapt to changes faster. 

Thanks to this solution, the marketing automation system can combine a strategy for investigating the degree of customer interest in various offers with a strategy for maximizing responses: an offer from a pool of offers available to customers will necessarily be shown to a larger or smaller proportion of customers, depending on their response to this offer. Among other things, RL models provide quick hypothesis testing for new offers (the model will reduce their impressions if they are not of interest to customers), and also allow you to effectively solve the problem of personalizing offers for customers without a history in the Bank's mobile application. The system easily adapts to changes in customer behavior, seasonal trends and is able to update on the basis of fresh data as often as possible, even every minute. 

We actively monitor the development of new methods and IT tools and try to constantly improve the quality of work with data, improve communication with customers. We move away from the needs and interests of the Bank's clients. Now, thanks to the implemented project, together with GlowByte, we can easily, quickly and, most importantly, safely test hypotheses for new and existing products, focusing as much as possible on the current needs of our customers. said Konstantin Kiselev, Director of the BSPB Client Base Development Department.

"We are coming from the needs and interests of the bank's customers. The developed tool will determine the effective audience coverage for each offer, add an interesting product to the list of recommended ones, and if the response to the new product is too low, the system itself will reduce its impressions to customers, " explained Maria Dukhnich, head of the targeted marketing department of BSPB.

2024: Implementation of MLOps tools to improve the performance of the modeling platform

Bank St. Petersburg, with the support of integrator GlowByte, has introduced MLOps tools to improve the efficiency of the modeling platform. GlowByte announced this on January 13, 2025. The project made it possible to speed up the cycle of updating models, centralize the processes of preparing data, training models and their deployment, as well as implement an automated system for controlling the versions of model code.

As part of the project, Airflow, Gitlab CI/CD and DVC (Data Version Control) tools were introduced, thanks to which the lifecycle of models is managed. Airflow provides task coordination and automatic start-up of processes, using the Gitlab CI/CD tool, a single pipeline of model output from the development stage to production is built, DVC allows you to save data to S3, as well as versioning large data files, models and other artifacts.

We integrated all stages of development and testing with the GitLab version control system, thanks to which we automated the assembly processes for models machine learning and the process of model production, made the history of all changes in the code base transparent. In addition, Airflow made it possible to automate the launch of batch models, and also made it possible to optimize the management of cluster resources, "explained KubernetesSergey Novoselov, architect of Advanced Analytics GlowByte.

The introduction of MLOps approaches increased interest in using the ML platform, which, in turn, contributed to attracting more business users at St. Petersburg Bank. 

In the bank "St. Petersburg" ML-platform has existed for a long time. In 2024, thanks to joint work with GlowByte, it received an additional impetus for development on the way to a single bank space to solve complex multifactorial problems. The implemented tools will not only increase the efficiency of model developers, but also greatly facilitate the availability of their results to users, said Kirill Svetlov, Managing Director of the Treasury Directorate of PJSC Bank St. Petersburg. "

In the future, it is planned to continue the development of the platform, improving the stability of its operation by better managing resources in the Kubernetes cluster and applying modern approaches in the field of machine learning and MLOps, including the MLFlow tool for tracking experiments with models.