Li Wenhao developed a production-ready DPA3 architecture for the metatensor/metatrain repository, focusing on mixed-precision support and seamless pretrained model loading. Using Python and PyTorch, Li implemented a pathway that ensures pretrained weights are consistently applied and safely serialized, streamlining experimentation and deployment across machine learning pipelines. The work included expanding unit and serialization tests, improving CI reliability, and addressing CUDA device handling for deterministic initialization. Li also enhanced documentation with detailed usage guides and code references, reducing onboarding time. The project emphasized maintainability through trainer deduplication, type annotations, and a shared test structure, reflecting strong software engineering practices.