Developed deterministic training resumption for the NVIDIA-NeMo/Megatron-Bridge repository by implementing logic to persist the dataloader’s stream position per data-parallel rank during checkpointing. This backend feature, built using Python and PyTorch, enables exact sample-position recovery after interruptions, ensuring reproducibility and reducing wasted compute in distributed training environments. The solution integrated the Energon dataloader with robust checkpointing and resume flows, improving reliability for long-running training jobs. No major bugs were reported during this period, reflecting careful attention to stability. The work demonstrates proficiency in distributed systems and backend development, with a focus on practical improvements to training workflows and infrastructure.