Kai Yao · Selected writing
Technical writing on machine-learning systems, developer tooling, and trustworthy AI. This archive begins with my work on Neural Coder at Intel and will grow with future writing across research and engineering.
Featured archive · Intel, 2022
At Intel, I helped develop Neural Coder, a component of Intel Neural Compressor that automated the insertion and evaluation of deep-learning optimizations in existing Python code. It combined static program analysis—syntax analysis, type inference, and call-graph parsing—with automated benchmarking and brought the workflow from PyTorch scripts to Hugging Face Transformers, Alibaba Cloud PAI-DSW, JupyterLab, and Visual Studio Code.
Neural Coder was demonstrated in Intel CEO Pat Gelsinger’s Intel Innovation 2022 keynote. This work was recognized internally at Intel with the AIA Division Achievement Award for Neural Coder innovation (2022) and the CESG SW AI Division Recognition Award for the Alibaba Cloud collaboration (2023); both are listed on my professional profile.
Writing archive
Model optimization
Showed how Neural Coder automatically enabled Optimum Intel INT8 quantization for Hugging Face Transformers. The evaluated workflow covered 18 Trainer-based PyTorch tasks and reported about 1.7× inference acceleration on 3rd Gen Intel Xeon processors, typically with less than 1% accuracy loss.
Developer tooling
Presented a Visual Studio Code extension that applied static or dynamic INT8 quantization and BF16 optimization, preserved generated code changes as inspectable patches, and benchmarked candidate transformations directly inside the editor.
Industry collaboration
Documented an Intel–Alibaba integration that brought Neural Coder into Alibaba Cloud PAI-DSW and exposed BladeDISC as an optimization backend, enabling one-click optimization and automatic before-and-after benchmarking from JupyterLab.
ML systems
Introduced Neural Coder, a one-click workflow combining deep-learning optimization rules with Python syntax analysis, static type inference, and call-graph parsing to insert quantization code into PyTorch scripts and benchmark applicable optimizations.