Worked on the intel/onnxruntime repository to deliver and stabilize the TensorRT Execution Provider (TRT EP) for ONNX Runtime, focusing on CUDA-accelerated inference in multi-stream scenarios. Addressed a critical bug by ensuring proper CUDA context preservation when user-provided streams are present, which improved execution correctness and reduced context-switch issues. Introduced auxiliary streams support, allowing explicit control over additional streams and enabling finer-grained stream management. Collaborated across TRT EP and NV Execution Provider components to align context handling and stream logic. Utilized C++, CUDA, and TensorRT, demonstrating a deep understanding of ONNX Runtime internals and multi-stream workload optimization.