Worked on improving test reliability for the keras-team/keras repository by addressing instability in the Torch MPS backend tests. Focused on backend development and testing, the work involved enforcing correct numeric type handling by explicitly casting input data to float32 within dynamic backend tests. This technical approach reduced flaky failures and enhanced continuous integration reliability. Utilizing Python and leveraging knowledge of backend systems, the developer fixed a persistent bug that was causing test failures. The contribution did not introduce new features but demonstrated attention to detail and a methodical approach to debugging, resulting in more stable and predictable test outcomes for the project.