Worked on stabilizing distributed training workflows and dataset construction across the sbintuitions/flexeval and NVIDIA-NeMo/Megatron-Bridge repositories. Addressed critical bugs in Python-based machine learning data pipelines, focusing on multi-GPU reliability and reproducibility. Improved tensor parallel argument handling in flexeval to prevent duplicate flags, ensuring correct model behavior in distributed environments. Enhanced the Megatron-Bridge pipeline by making diffusion data encoder shuffling deterministic and fixing sequence length propagation in the Energon task encoder. Emphasized robust unit testing with PyTest and mocking, expanding test coverage and centralizing synchronization logic to reduce regression risk and improve maintainability in backend development workflows.