Worked on backend development and performance optimization for deep learning systems, focusing on reliability and throughput improvements. In the HabanaAI/vllm-fork repository, addressed a padding handling bug in the padding-aware sequence processing path by refining index calculations and ensuring correct application of padding during scheduling, which stabilized hidden state updates and reduced edge-case misalignment. Later, contributed to the vllm-project/vllm-gaudi repository by optimizing Chunk Scan operations in PyTorch for the Gaudi backend, simplifying code paths, and improving numeric precision through vectorization and native PyTorch operations. Utilized Python, PyTorch, and algorithm design to enhance maintainability and execution efficiency across both projects.