Worked on the UKGovernmentBEIS/inspect_ai repository to deliver a new multimodal content handling feature and a Thinking Mode with an enable_thinking option. The approach involved refactoring backend content processing in Python to normalize data by converting lists to strings, which improved reliability for downstream workflows. Integrated explicit support for multimodal models and implemented robust error handling to provide clear feedback when unsupported content types are encountered. Focused on API integration and model integration, the work addressed specific issues to reduce integration risk and enhance versatility for multimodal workflows, resulting in clearer developer guidance and more maintainable backend infrastructure.