Worked on targeted enhancements to ONNX Runtime, focusing on performance and memory efficiency for quantized models. Addressed a CUDA backend issue in the intel/onnxruntime repository by correcting after_gather_dim indexing for 4-bit weight nibbling, improving model compression and deployment on GPUs. In microsoft/onnxruntime-genai, implemented broader quantization configurability and introduced shared embeddings to optimize memory usage and model size. Added a new option to untie QKV projections, increasing flexibility for quantized model architectures. The work leveraged C++ and Python, applying expertise in CUDA, GPU programming, and model optimization to deliver both a feature and a bug fix within the month.