During this period, contributed to the bytedance-iaas/sglang repository by developing a penalty-based speculative decoding feature for the Eagle model. This enhancement introduced frequency and presence penalties into the token generation process, allowing for more controlled and predictable model outputs. The implementation aligned with spec v2 overlap scheduling, ensuring compatibility with evolving project standards. Leveraging Python for both API development and machine learning tasks, the work established a foundation for future penalty parameter tuning and experimentation. Emphasis was placed on robust testing to validate output reliability, ultimately supporting improved user trust and model quality in production environments.