Personal site

Zihao Zheng

AI Infrastructure & Speech-Language Systems Architect

Building reliable, explainable, and verifiable AI at production scale

2 Granted patents, first inventor
~10 Years building production AI
40M+ Households on Tmall Genie by 2023

About

More about me →

I am an AI Infrastructure & Speech-Language Systems Architect. Previously, I was a founding core AI architect of Alibaba's Tmall Genie, where I designed the NLU domain-classification engine and end-to-end ASR architecture serving tens of millions of households. Before that, I received my B.Sc. in Information and Computing Science from South China Normal University.

I am the primary author of standalone TensorBoard (dmlc/tensorboard), first inventor on two granted patents in speech and language systems, and a former Apache MXNet PMC Member and Committer. I care about building AI that operators can inspect, verify, and trust at production scale.

Selected work

All projects →

Two granted patents running on Tmall Genie at national scale, and the open-source infrastructure that came before them. Every page carries a working model of the idea you can operate yourself — no video, no slides, just the mechanism.

7 requests · one cascade, two ways to wire it Rules first 5/7 Model first 5/7 the failures move — the total never does

Patent CN112002313B · Interactive

Memory-Network NLU

Rules or the model — which one gets to win? Neither answer routes all seven requests. That impossibility is the patent.

~65% → ~85% accuracy 15.61M units shipped, 2019
Try both orderings
the user’s list 郑梓豪 transcript 打 电 话 给 郑 梓 豪 calls Zheng Zihao 正 子 号 no contact matches

Patent CN113808593B · Interactive

Indicator Loss for Context-Aware ASR

Delete one contact and watch the transcript break. Same sounds, valid words, wrong meaning — fixable without retraining.

−32.9% negative feedback In Alibaba Cloud materials, 2023
Edit the bias list
to see your own MXNet training run mxnet tensorflow no viewer log written, unreadable install the framework you didn’t choose tensorboard charts, with no framework attached dmlc/tensorboard — primary author

Open source · ML infrastructure · Interactive

Standalone TensorBoard

Why should seeing your own training run cost you a second framework? I rebuilt TensorBoard as a standalone logging and rendering tool. Google's own now ships the same way.

Top 3% of ML-visualisation repos tensorboardX: 5.9M downloads/mo
Try the install →

Academic Services

  • Conference Program Committee Member / Reviewer
    • Empirical Methods in Natural Language Processing (EMNLP) Industry Track, 2026
    • ACM Conference on Recommender Systems (RecSys) Research & Practice Notes and Demos Tracks, 2026
  • Workshop Reviewer
    • Workshop on Noisy and User-generated Text (WNUT), @EMNLP 2026
    • Third Workshop on Uncertainty-Aware NLP (UncertaiNLP), @EMNLP 2026
    • Efficient Representation Learning for Multimodal Information Retrieval (EReL@MIR), @ACM MM 2026
    • AI Measurement Science (AIMS), @COLM 2026
    • Social Simulation with LLMs (Social Sim), @COLM 2026

Recent Posts

All posts →