Worked on the Aleph-Alpha-Research/eval-framework repository to address a critical reliability issue in artifact-based model initialization. Focused on backend development using Python, the work involved debugging and resolving premature deletion of downloaded Weights & Biases artifacts during model loading. By refining context management within the WandbFs component, the solution ensured that temporary directories remained valid until the model load process completed, thereby eliminating load-time errors. This targeted bug fix improved error handling and reduced support incidents, resulting in more stable production pipelines. The approach demonstrated strong skills in root-cause analysis, context management, and backend reliability engineering without introducing new features.