Lewen Zheng

AI agent systems / LLM infrastructure / performance engineering

I build agent systems that turn frontier AI into reliable work.

I am a researcher and systems builder working across agent orchestration, shared knowledge infrastructure, operator-generation workflows, and high-performance LLM inference.

Email me View work
A systems map connecting agent systems, inference, and research into reliable AI work.
20+ teams Agent-system adoption across research teams
200+ operators Supported through CANN-Bench operator-generation workflows
1.5x-4.5x Measured inference performance gains on Ascend deployments

Current focus

Agent systems that preserve context, coordinate work, and produce evidence.

MemAgora

Shared team knowledge infrastructure for agent systems, deployed into CANN-Bench operator generation and used to support production of 200+ Ascend operators.

Jarvis

Agent-collaboration capabilities across task orchestration, workflow coordination, and reusable team knowledge, adopted by 20+ research teams.

Operator Generation

Agent-driven workflows for Kirin chip and HarmonyOS integration, moving from prototype to system-level delivery.

Infrastructure

High-performance LLM inference on real hardware.

Large multimodal models

Optimized Qwen3-VL, Qwen2.5-VL, Wan2.1, Llama, and DeepSeek-family inference workloads across multi-card Ascend deployments.

Performance systems

Work spans multi-stream concurrency, hybrid parallelism, compute-communication pipelining, automated profiling, and deployment-risk triage.

Research-to-production

I care about the path from mathematical reasoning and system design to measured product behavior under production constraints.

Research

Selected publications and awards

  1. Efficient Simulation of Polyhedral Expectations with Applications to Finance Mathematics of Operations Research, published online 2025. Winner, INFORMS Section on Finance Best Student Paper Competition, 2022.
  2. Best-Arm Identification with High-Dimensional Features Proceedings of the 2024 Winter Simulation Conference.
  3. Conditional Importance Sampling for Convex Rare-Event Sets Proceedings of the 2023 Winter Simulation Conference. INFORMS-Sim Best Student Paper Award, WSC Ph.D. Colloquium, 2023.
  4. Efficient Simulation for Linear Programming under Uncertainty Proceedings of the 2021 Winter Simulation Conference. Runner-up, Best Theoretical Paper Competition, WSC 2021.

Contact

For research, systems, or agent-product conversations.

I am interested in reliable agent systems, AI infrastructure, evaluation, and products where tool use and memory create real leverage.

zhengrockman@gmail.com