Worked on multilingual evaluation and model infrastructure across swiss-ai/lm-evaluation-harness and swiss-ai/Megatron-LM. Delivered a WMT translation task pipeline supporting multiple language pairs, integrating CometKiwi22 metrics and robust data handling for missing or incomplete datasets. Enhanced error handling, logging, and dependency management to improve evaluation reliability and maintainability. Refactored translation task infrastructure for readability and optimized GPU memory usage. In Megatron-LM, implemented checkpoint parameter enhancements and led a brand migration from SwissAI to Apertus, updating model configurations and file paths. Used Python and Markdown extensively, focusing on codebase management, testing, and reproducible workflows for scalable machine learning development.