Worked on the keras-team/keras repository to enhance the robustness of floating-point sign bit detection in the TensorFlow backend. Addressed a potential integer overflow issue by replacing a hardcoded sign bit with the use of tf.int32.min, ensuring more reliable signbit handling in backend computations. This fix improved the stability and maintainability of the codebase, particularly in scenarios involving edge-case integer values. The work was implemented using Python and leveraged both Keras and TensorFlow, demonstrating a focus on backend development and numerical reliability. The contribution centered on bug resolution, reflecting careful attention to detail and code safety in machine learning frameworks.