Scalable cloud-driven pipeline for machine learning algorithms deployment

Abstract

As machine learning (ML) continues to permeate industry and research, the need for robust, scalable, and maintainable deployment pipelines has become critical. Despite advances in model development, operationalizing ML models remains a non-trivial challenge due to the complexity of managing infrastructure, ensuring reproducibility, and integrating with heterogeneous data sources and consumers. This thesis proposes a generalized, cloud-oriented deployment framework for ML algorithms, addressing these challenges through modular and extensible architecture. The framework is built upon containerized microservices, enabling environment-agnostic deployment and seamless transition from local development to production. A core component of the system is its pluggable inference interface, which abstracts model-specific logic, allowing arbitrary ML algorithms to be integrated with minimal engineering overhead. The pipeline supports dynamic parameterization of models, automatic resource provisioning, and concurrent execution, promoting scalability and adaptability to varying workloads. A lightweight, web-based interface facilitates user interaction with the system, supporting dataset upload, model parameter configuration, execution of inference tasks, and retrieval of outputs in multiple formats. Workflow orchestration, monitoring, and logging are integrated into the pipeline to ensure operational transparency, fault tolerance, and streamlined debugging. All system components are loosely coupled, promoting maintainability and compatibility with multiple cloud service providers and on-premises environments. The generality of the framework is validated through the integration of the Growing Self-Organizing Map (GSOM) algorithm and deployment on Amazon Web Service (AWS), demonstrating its ability to support computationally intensive models while preserving system responsiveness and reliability. However, architecture is agnostic to specific ML methods or cloud platforms, making it applicable across a wide range of deployment scenarios. This work contributes to a reusable and scalable Machine Learning Operations (MLOps) aligned infrastructure that bridges the gap between ML development and production deployment, supporting continuous delivery, reproducibility, and operational efficiency in modern ML workflows.