Customer
The Tier-2 private bank in Eastern Europe (over 22 million private clients, the number of corporate clients exceed 1 million).
Challenge
The growth in the number of ML-models used by the bank in various business processes (lending, marketing, customer service, cyber security) and the modes of their use (batch, REST, streaming) required a revision of the processes for working with ML-models to ensure the ability to operate a large number of ML-models and fast time to market of new versions of models.
Project Results
An MLOps platform has been implemented that automates the processes of deploying and executing models, allowing data scientists to use any machine learning libraries and programming languages, and also provides monitoring of models during operation. 30+ models have been transferred to the platform.
Business Value
The implementation time of new versions of ML-models has been reduced from several months to one day. Processes of implementation and monitoring of models are unified. High performance of models on large arrays of client data (30+ mln. records) has been provided.
Tech Stack
JupyterHub, Tensorflow, Scikit-learn, MLFlow, Apache Airflow, Apache Spark, Cloudera CDP, Kubernetes, Jenkins, ArgoCD.
The Tier-2 private bank in Eastern Europe (over 22 million private clients, the number of corporate clients exceed 1 million).
Challenge
The growth in the number of ML-models used by the bank in various business processes (lending, marketing, customer service, cyber security) and the modes of their use (batch, REST, streaming) required a revision of the processes for working with ML-models to ensure the ability to operate a large number of ML-models and fast time to market of new versions of models.
Project Results
An MLOps platform has been implemented that automates the processes of deploying and executing models, allowing data scientists to use any machine learning libraries and programming languages, and also provides monitoring of models during operation. 30+ models have been transferred to the platform.
Business Value
The implementation time of new versions of ML-models has been reduced from several months to one day. Processes of implementation and monitoring of models are unified. High performance of models on large arrays of client data (30+ mln. records) has been provided.
Tech Stack
JupyterHub, Tensorflow, Scikit-learn, MLFlow, Apache Airflow, Apache Spark, Cloudera CDP, Kubernetes, Jenkins, ArgoCD.