Worked on the Xilinx/onnx-mlir repository to enhance ONNX operator support and conversion reliability. Developed robust type inference for the ONNX RandomNormalLike operator, refactoring shape inference logic to accurately propagate output element types based on input or attribute dtypes. Added comprehensive unit tests in C++ and MLIR to validate type inference across multiple data types, improving correctness in IR generation and downstream optimizations. Later, expanded the ONNX to TOSA conversion path by enabling runtime padding support in PaddingOp and refactoring constant pad value handling, increasing robustness for dynamic tensor operations. Focused on compiler development, type systems, and tensor operation reliability.