Developed and integrated the AdaMuon optimizer into the marin-community/marin repository to enhance Muon training with element-wise adaptivity. The work involved subclassing MuonConfig to create AdaMuonConfig, implementing sign-stabilized orthogonalization, element-wise second momentum, and dynamic RMS alignment using Python. Conducted end-to-end experiments on the Qwen3 model, utilizing a single node with eight A100 GPUs and performing a grid search over learning rate, momentum, and epsilon parameters. Documented comprehensive configuration and results in speedrun_results.json to ensure reproducibility. This contribution established a foundation for future optimizer exploration and performance improvements, demonstrating skills in Python programming and machine learning.