Worked on the pytorch/executorch repository, focusing on improving correctness and reliability in core backend components using C++. Addressed two critical bugs over a two-month period, including fixing an uninitialized accumulator in the portable batch normalization kernel and correcting the output data format for quantized convolution in NHWC layout. Employed strong memory management and unit testing practices to ensure robust solutions, adding regression tests to verify output integrity even under adverse conditions. These contributions enhanced the determinism and stability of batch normalization and quantized inference paths, reducing the risk of incorrect results for models deployed across diverse hardware environments.