Presentation at the Nordic-RSE conference, 9 June 2026, Tromsø, Norway. (Link) Abstract: Invasive marine species threaten marine ecosystems and biodiversity across Europe. Preventing their spread requires early detection and rapid response, motivating monitoring using underwater video. However, processing these data at scale becomes computationally demanding. To address this challenge in the EU-funded Horizon Europe DTO-BioFlow project, we are extending the SUBSIM subsea image analysis platform to support training and inference of computer-vision species detection models across diverse HPC platforms. In this talk, we share our experiences from the research software engineering work behind this effort: refactoring to improve maintainability and stability, and developing CI/CD pipelines to automatically build container images for AMD ROCm, NVIDIA CUDA, and CPU backends. These containers run consistently from laptops through the EDITO platform (the infrastructure of the European Digital Twin Ocean) up to EuroHPC systems such as the LUMI supercomputer. This approach simplifies user setup and enables reproducible workflows across heterogeneous hardware, while supporting both Jupyter-based interactive use and batch execution on HPC systems.