Developed and integrated the BailingMoeV2.5 model architecture into the inclusionAI/AReaL repository, focusing on production readiness and distributed training reliability. Leveraged Python and deep learning frameworks to enable advanced features such as Lightning Attention, Multi-Latent Attention, Mixture of Experts, and Context Parallelism. Enhanced deployment workflows by supporting HuggingFace checkpoint loading and bridging with Megatron, while improving save/load stability and configuration integrity. Addressed distributed training challenges by refining expert shard handling and ensuring consistent parameter behavior. Collaborated with cross-team contributors to align model architecture improvements, demonstrating expertise in machine learning, model optimization, and parallel computing within a complex engineering environment.