Over a two-month period, contributed to the tenstorrent/tt-xla and tenstorrent/tt-mlir repositories by enhancing benchmarking frameworks, optimizing model performance, and improving distributed system reliability. Developed new benchmarking features and expanded CI coverage for vLLM and Llama-3.1 models, using Python and C++ to enable consistent performance tracking and regression detection. Implemented decoding optimizations that reduced graph compilation overhead and accelerated token generation. Enhanced normalization efficiency through RMSNorm fusion patterns across multiple model families. In tt-mlir, improved distributed data-parallel cache reliability by introducing user-dimension sharding and enforcing MeshWorkloadFactory integration, leveraging MLIR and deep learning expertise for robust, scalable solutions.