Design and Implementation of a Cloud infrastructure for the Deployment and Inference of Machine Learning Models

Abstract

This thesis presents the design and implementation of a scalable infrastructure to support a Machine Learning model used for predicting warehouse stocking and developed with industry-standard technologies such as AWS, Docker, and Apache Airflow.

The infrastructure follows the key DevOps practices, including CI/CD pipelines, Infrastructure as Code (IaC), and automated testing and monitoring.

The project focuses on the following contributions: a Data Extraction, Transformation, and Load (ETL) process to deliver data from SAP to the ML model; the design of a cloud-based architecture that meets high scalability and reliability standards; and the implementation of a CI/CD pipeline to automate the deployment and test of new features and bug fixes.

This cloud architecture is able to suggest new changes that can be made to the Stocking Policy in an automatic way, meaning that the planner will get the latest predictions directly on the data platform, and they will be able to decide how to update the policy. This work highlights the importance of integrating DevOps practices and cloud technologies to create a flexible and efficient ML pipeline in a production environment, that is able to scale and adapt based on business or technical requirement. The thesis explains how new features and bugfixes can be tested and implemented by the developer.