Worked on the Arize-ai/openinference repository to improve the accuracy of token usage metrics by addressing a bug in the prompt token counting logic. Focused on backend development using Python, the work involved identifying and fixing an issue where cached tokens were being double-counted, which previously led to discrepancies in usage reporting and cost estimation. Updated the logic to exclude cached tokens from the total prompt count and enhanced the test suite to validate the corrected behavior. This change improved the reliability of downstream dashboards and billing insights, demonstrating careful debugging, data analysis, and a disciplined approach to testing and code review.