Contributed backend enhancements to the tensorflow/tensorflow repository, focusing on improving shape flexibility and model deployment workflows. Developed shape inference capabilities for the tfl.Pack operator, enabling the system to infer unknown input and output dimensions during legalization, which supports more dynamic tensor shapes. Added variable bias support to convolution operations, allowing different input and bias types and increasing flexibility in model design. These features were implemented using C++, MLIR, and TensorFlow, targeting the TOSA path to facilitate robust provider workflows and experimentation with dynamic shapes. The work emphasized backend stability and adaptability for deep learning and machine learning applications.