Worked on the exo-explore/exo repository to enhance chat reliability by implementing GLM 4.7 stop-token mapping, ensuring chat completions halt precisely at the assistant’s turn boundary. Used Python to map stop tokens to GLM 4 IDs within mlx utilities, aligning end-of-sequence handling with conversational flow and preventing unwanted continuations. Developed automated unit tests and integrated tooling checks to validate the solution, with manual verification performed on an Apple Silicon cluster. Focused on machine learning and natural language processing, the work improved compatibility for OpenAI-compatible clients while requiring only minimal prompt-template adjustments, demonstrating careful attention to both robustness and user experience.