Worked on the tensorflow/tensorflow repository to enhance graph optimization and debugging workflows within TensorFlow. Developed a configurable graph-pruning feature that allows users to preserve specific operations, addressing the challenge of maintaining side-effectful ops during optimization passes. Introduced DebugIdentityOp, which provides identity mapping for non-reference tensors and includes device, tensor name, and debug URL parameters, streamlining the debugging process for developers. Adjusted the MLIR implementation to ensure DebugIdentityOp remains available for debugging by removing the 'Pure' trait, preventing unwanted optimization. Leveraged C++, MLIR, and compiler design expertise to deliver targeted improvements that support stability and developer productivity.