About
I’m Zihao Zheng, an AI Infrastructure & Speech-Language Systems Architect.
I build AI systems that operators can inspect, verify, and trust — from cross-framework ML tooling adopted across ecosystems, to patented speech and language architectures serving tens of millions of households.
My work sits at the intersection of deep-learning infrastructure, speech recognition, and natural language understanding, with a focus on systems that remain reliable under real-world noise and scale.
Background
I graduated from South China Normal University in 2015 with a B.Sc. in Information and Computing Science. From my second undergraduate year, I taught myself machine learning through open courseware.
In July 2015 I joined Alibaba. From 2017 I served as a founding core AI architect for Tmall Genie, designing its NLU domain-classification engine and end-to-end ASR architecture — the two foundational AI layers every voice command passes through. During that period Tmall Genie grew to tens of millions of households and became China’s leading smart speaker platform.
In parallel, I created standalone TensorBoard (dmlc/tensorboard) so non-TensorFlow teams could inspect training runs without framework lock-in. That work was adopted by AWS as the basis for mxboard, referenced by PyTorch’s tensorboardX, and publicly acknowledged by Google’s TensorBoard engineering leadership. I later served as an ASF Board-appointed member of the Apache MXNet Project Management Committee.
I am first inventor on two granted patents covering memory-network NLU domain classification and Indicator Loss for context-aware ASR.
Design philosophy
Accuracy alone is not enough. AI deployed to millions of users must be inspectable by the engineers who maintain it, verifiable against the knowledge it encodes, and reliable without treating the model as a black box. Every major contribution I have made expresses that conviction across a different layer of the stack.