Rylan developed a flexible token generation feature for the menloresearch/torchtune repository, focusing on enabling optional tensor-based softmax sampling within the token generation path. Using Python and leveraging deep learning frameworks such as PyTorch, Rylan implemented a configurable approach that allows for rapid experimentation with different sampling strategies during inference. This enhancement increases adaptability and supports a variety of production workloads by making the sampling process more modular and maintainable. The work was delivered as a focused, traceable change, laying the groundwork for future improvements and demonstrating a strong grasp of AI development and machine learning engineering practices.