Worked on the oneapi-src/oneDNN repository to deliver a critical robustness fix for recurrent neural network (RNN) handling. Addressed a regression by implementing explicit validation of the prop_kind parameter in both forward and backward RNN descriptor initialization functions, ensuring that incorrect propagation kinds are rejected and silent errors are prevented. Enhanced the reliability of RNN-based model deployment by adding targeted unit tests to cover negative cases, thereby strengthening API correctness across all RNN descriptor creation paths. Utilized C++ for API design and validation logic, applying machine learning domain knowledge and test automation to improve production model reliability.