Worked on the ecmwf/anemoi-core repository to enhance the stability of autoregressive forecasts by introducing learnable residual connections, including scalar and spectral Ornstein models with per-variable and spectral parameters. Leveraged Python and deep learning techniques to implement low-pass filtering and optional truncation, effectively bounding error growth during extended rollouts. Developed inverse spectral transforms to support comprehensive end-to-end spectral processing. Expanded validation through unit and integration tests, including multi-GPU scenarios, and improved documentation and configuration schemas. The work focused on robust production deployment, reproducibility, and providing clear on-ramps for future experiments in data science and machine learning workflows.