During March 2026, Gusti Pardo focused on improving data integrity and billing accuracy in the BerriAI/litellm repository by addressing a bug in Gemini usage metadata token accumulation. He implemented logic to ensure image tokens were properly accumulated across both prompt and response modalities, rather than being overwritten, which enhanced the accuracy of usage analytics. Using Python, he incorporated robust error handling and unit testing to validate the new accumulation logic and added normalization to support both camelCase and snake_case key naming conventions. This work resulted in a more stable metadata pipeline and reduced discrepancies in billing and analytics reporting.