MLOps Architecture for Machine Vision: A Practical Architecture for Industry Application
Linna, Matti (2025)
2025
Tietotekniikan DI-ohjelma - Master's Programme in Information Technology
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
This publication is copyrighted. You may download, display and print it for Your own personal use. Commercial use is prohibited.
Hyväksymispäivämäärä
2025-06-09
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202506086926
https://urn.fi/URN:NBN:fi:tuni-202506086926
Tiivistelmä
Despite the growing adoption of machine learning (ML) across industries, many organizations struggle to transition from proof-of-concept (PoC) solutions to production-level systems. This gap often stems from the lack of standardized practices and infrastructure for operationalizing ML workflows. MLOps has emerged as a discipline to address these challenges, but practical guidance remains lacking and fragmented.
To help address this gap, this study applies a Design Science Research (DSR) approach to develop both a reference architecture and a concrete architecture for machine vision MLOps pipelines. The reference architecture is based in the existing literature and validated through the design and implementation of the concrete architecture. The concrete architecture is built for the industry partner’s production-level system and is designed to be easy to adopt, cost-effective, extendable, and secure. With these objectives, it specifically targets organizations at the early stages of their MLOps journey.
The study indicates that the modular structure of the designed reference architecture allows for incremental development while providing guidelines for how the architecture can evolve toward higher maturity levels. This makes it a valuable foundation for practitioners when designing and developing MLOps architectures. A production-ready concrete architecture was built using Azure Machine Learning, MLflow, GitHub, and FastAPI. This combination of cloud-managed and open-source tools makes the architecture easy to adopt and improves reproducibility in model development. The architecture has no upfront costs, and Azure ML provides enterprise-grade security, making it cost-effective and accessible for organizations of any scale. This work contributes to the MLOps field by offering a generalizable reference architecture and a concrete implementation example, helping practitioners start with MLOps practices and move their ML workflows from experimentation to production.
To help address this gap, this study applies a Design Science Research (DSR) approach to develop both a reference architecture and a concrete architecture for machine vision MLOps pipelines. The reference architecture is based in the existing literature and validated through the design and implementation of the concrete architecture. The concrete architecture is built for the industry partner’s production-level system and is designed to be easy to adopt, cost-effective, extendable, and secure. With these objectives, it specifically targets organizations at the early stages of their MLOps journey.
The study indicates that the modular structure of the designed reference architecture allows for incremental development while providing guidelines for how the architecture can evolve toward higher maturity levels. This makes it a valuable foundation for practitioners when designing and developing MLOps architectures. A production-ready concrete architecture was built using Azure Machine Learning, MLflow, GitHub, and FastAPI. This combination of cloud-managed and open-source tools makes the architecture easy to adopt and improves reproducibility in model development. The architecture has no upfront costs, and Azure ML provides enterprise-grade security, making it cost-effective and accessible for organizations of any scale. This work contributes to the MLOps field by offering a generalizable reference architecture and a concrete implementation example, helping practitioners start with MLOps practices and move their ML workflows from experimentation to production.