Worked on stabilizing and hardening the sampling pipeline in the Blaizzy/mlx-audio repository, focusing on the Fish and Whisper samplers. Addressed a bug where random seed handling was ineffective due to compile-time state capture, updating Python decorators to use a state-threaded random source. This change ensured that sampling functions consistently utilized the correct random state, resulting in deterministic and reproducible behavior across runs. The work required proficiency in Python, audio processing, and machine learning, and involved aligning new implementations with existing utilities for random state management. These improvements enhanced the reliability and maintainability of the audio sampling infrastructure.