Diwen Zhu

System Engineer → AI Infrastructure → Agent Builder

diwen.zhu@gmail.com

Experience

Alibaba Group - Tongyi Lab - LLM / AI Infrastructure

AI Infrastructure Engineer - Jan.2023 ~ Now

  • Co-designed GPU pooling and request/resource scheduling on 10K-GPU cluster; reduced inference cost by ~90% for mid/long-tail models, ~20% for head models, ~30% overall while meeting SLO. Work published at SOSP 2025.
  • Side Project: Built an AI-native stock selection system for A-share market. In the age of AI, a beautiful backtest curve means nothing. What truly matters are two fundamental questions: What edge am I capturing from the market? and How much of that edge can my approach realistically harvest? AI is simply an accelerator for arriving at that understanding.
  • Side project: Built ReAct + Self-Evaluation agent loop for game-like programming tasks in a no human-in-the-loop setting (conceptually similar to Ralph Loop).

Alibaba Group - DAMO Academy - Distributed Systems

System Engineer - Jan.2020 ~ Jan.2023

  • GraphScope: Designed and implemented the abstraction layer between compute and storage for this large-scale distributed graph system; project ranks #1 on LDBC benchmark.
  • Vineyard: Co-founder of this CNCF Sandbox distributed in-memory storage project; led architecture and core module development.

Shanghai Beizhou Investment Management Co., Ltd. - Quantitative Research and Trading

Technical Partner & Co-founder - Feb.2016 ~ Jan.2020

  • Led end-to-end design and development of A-share mid/high-frequency quantitative research and trading systems.

Education

Nanyang Technological University (Singapore) — Ph.D. in Computer Science (Database Systems)

3 first-author papers at SIGMOD (2) and VLDB (1). Graduated 0.5 year early.

Fudan University — B.S. in Computer Science

ACM/ICPC regional gold medal.

Projects

Vineyard

CNCF Sandbox distributed in-memory data sharing; co-founder, led architecture and core modules.

v6d.io

GraphScope

Large-scale distributed graph engine; designed compute–storage abstraction layer; #1 on LDBC.

graphscope.io

ReAct + Self-Evaluation Agent

Side project: agent loop for game-like programming tasks without human-in-the-loop (Ralph Loop–style).

Skills

Distributed systems · GPU scheduling · high-performance computing · LLM inference optimization · ReAct / agentic workflows