Many publications elaborate on the so-called AI engineering or MLOps (Machine Learning Operations) processes, also from an architecture point of view. However, it remains a challenge to translate this into a practical approach for designing an MLOps architecture from the very beginning of a project. In this paper, we define an integrated approach to develop an MLOps architecture (including data, models and software) based on Google’s MLOps maturity levels and publications on architectural design decisions for machine learning. We demonstrate this approach on a real-life machine learning project, where we designed a streaming wearable data platform for detecting stress. The approach we took and the lessons learned are valuable for practitioners working on similar projects. Furthermore, the challenges we encountered in our project can serve as future research directions.