Developed scalable AI tooling across the microsoft/onnxruntime-extensions, microsoft/onnxruntime-genai, and microsoft/olive-recipes repositories, focusing on advanced model deployment and optimization. Delivered a multi-step sequence pre-tokenizer in C++ and Python, enabling complex tokenization patterns and improved performance. Integrated Qwen3.5-MoE model support with a 256-expert mixture-of-experts architecture, enhancing runtime scalability and architecture dispatch. Built a multilingual multimodal translation model supporting 55 languages, leveraging ONNX Runtime for efficient inference. Contributed an ONNX export recipe for a vision-language MoE model, broadening deployment options. Work emphasized cross-repository collaboration, robust testing, and performance tuning in computer vision and natural language processing domains.