👷‍♂️Experience

Currently， I'm a Senior Engineer at Huawei's XXX CTO Office, working on Large-scale clusters LLM Training and System-Algorithm Co-Design and Agent.

2026-now — Senior LLM engineer building production-grade Agent systems and harnesses. I turn frontier model capability into tools people actually use — heavy hands-on with Claude Code / Codex / Cursor — and I close the loop by feeding real internal tasks back into product and model post-training iterations.

2023-2025 — Senior LLM engineer specializing in large-scale training & RL framework optimization on 10K+ Ascend clusters. I work at the framework level — squeezing throughput out of training/inference pipelines, post-training, and distributed communication — and contribute back to the open-source LLM infra ecosystem.

What I Experienced on:

LLM Agent
- Vibe Coding – Extensive hands-on with Codex/Claude Code; independently built Ascend Bot to streamline daily workflows, which gained department-wide adoption and grew into a 10+ people dedicated team.
- Ascend Optim Agent – Identified critical pain points in Ascend-based optimization; designed and deployed two internal tools:RAG-based retrieval assistant for rapid performance diagnostics & knowledge lookup;Agent-based tuning assistant for automated performance optimization and decision support. Project officially approved at department level, with a dedicated team formed for continuous maintenance and R&D.
- Focus Evaluation – Designed and maintain evaluation frameworks for agent systems, using internal real-world tasks as key feedback signals to continuously improve product capabilities and inform model training iterations.
LLM Training
- Data synthesis & cleaning, post-training strategies (RLHF/DPO) to boost model performance.
- Led continued pre-training and post-training projects for our key clients (China Merchants Bank, PICC).
- Developed internal ascend skill-specific models via distillation (OPD) for Ascend infra scenarios.
LLM Infra
- Deep into the Ascend software stack - MindSpeed-LLM/RL, VeRL-Ascend，vLLM-Ascend - optimizing inference and RL training pipelines. Currently obsessed with pushing RL training efficiency on Ascend 910B，Ascend 910C，Ascend 950 Cluster ;
- Built and optimized 10K+ card clusters, including large scale 910B cluster，910C SuperPod deployments. Tuning HCCL and UB-Mesh configs to squeeze every bit of performance from the network fabric.
IDC Infra
- CLOS → ROFT/MP-FatTree → HPN2.0 → ZCUBE
- HCCL， UB-Mesh，RDMA
- 掉卡现象，快慢卡定位，ECC比特，模组更换

06/2023-Now | LLM ( LLM System-Algorithm and Agent Harness ) Engineer. | HUAWEI
11/2022-03/2023 ｜NLP Research&Engineer**( Ernie** ) Intern.｜BAIDU-NLP
04/2022-08/2022 ｜NLP Research( Randeng-T5 ) Intern.｜IDEA-CCNL

🏫Education

National University of Defense Technology – Changsha, China

M.S. in Software Engineering, School of Computer Science | Sep 2020 – Jun 2023