Worked on the NVIDIA/NeMo-Skills repository to enhance model response parsing by implementing token-level probability logging. Developed a feature in Python that extends the parse_openai_response workflow to accept and propagate a top_logprobs argument, enabling detailed capture of log probabilities for each generated token. This addition improved backend observability and facilitated more granular analysis of output generation, supporting model tuning and quality control. Focused on API integration and backend development, the work maintained full traceability to the relevant commit and ensured that the new logging capabilities integrated smoothly into the existing response handling pipeline without introducing regressions or breaking changes.