Developed the CyberGym Framework within the UKGovernmentBEIS/inspect_evals repository, creating a reusable platform for evaluating AI agents on real-world cybersecurity vulnerability tasks. The work centered on building unified task templates, automated dataset handling, and a sandboxed execution environment to enable standardized risk assessment and reproducible experiments. Leveraging Python for both API and full stack development, the framework introduced evaluation workflows with YAML configuration and direct data usage. Quality was enhanced through comprehensive unit and end-to-end testing, as well as rigorous code linting and typing with ruff and mypy. Documentation and template standardization further improved reproducibility and maintainability across experiments.