Developed and delivered a Reasoning Performance Metrics Tracking feature for the NVIDIA-NeMo/Eval repository, enhancing the observability of model reasoning during evaluation. This work introduced new metrics, including unfinished reasoning counts and finished ratios, by refining the logic within the ResponseReasoningInterceptor to ensure accurate data collection. The implementation was supported by comprehensive unit testing to validate correctness and maintain reliability. Documentation was updated in Markdown to clearly describe the new metrics and their impact on evaluation quality. Leveraging backend development and data analysis skills, this contribution enables more data-driven optimization, supports faster iteration cycles, and informs better business decisions for model evaluation.