ABOUT / MOYUAN
墨鵷
在模型、系统与安全之间
做一些真正能跑起来的东西
也写下一些与技术有关
又不完全关于技术的东西
01 / NOW
正在做
- 正在做
- 大模型推理 · AI 安全 · Agent
- 最近在想
- 长上下文到底值不值得那么贵
- 最近在折腾
- 这个网站
02 / NAME
名字里的方向
墨落
风起
羽翼向远
鵷yuān
非梧桐不止
非练实不食
非醴泉不饮取意于鵷雏，择善而栖
03 / METHOD
让现实参与判断
先把问题拆开
再让现实有机会说不
能测量的就测量
能留下证据的就留下证据
模型跑通不是结束
系统真正工作才算
04 / ELSEWHERE
一些旁枝
05 / SELECTED WORK
把想法做出来
PeerAssist
证据驱动 AI 审稿工作台，连接原文阅读、引用核查、人工确认与报告导出。
Unlimited-OCR
OCR 多模态模型的昇腾 NPU 适配与同口径评测。
AIGC 内容安全检测与审计系统
结合视觉伪造证据、敏感词与语义检索的多模态内容安全审计。
06 / JOURNEY
一路走来
北京大学大数据实验室 / Data & AI 安全课题组
AI 安全实习生（远程） 2026.04 - 至今
北京大学大数据实验室 / Data & AI 安全课题组
AI 安全实习生（远程） 2026.04 - 至今
负责 JoySafeter 测评执行链路工程化，开展 JoySafety 框架源码审计。
- 串联源码解包与隔离沙箱、MCP/Skills 动态工具调用、Worker 异步执行、证据归档与结构化报告生成。
- 重构 Run 状态迁移与稳定 ID，修复重复消费并发竞态，实现任务恢复与失败重试幂等性。
- 完成 12 个 MCP 服务、79 个安全工具加载验收（0 失败），覆盖 53 项后端与 21 项前端测试。
- 分析 JoySafety 双向审核、Prompt Injection、多轮会话、红线知识库代答、可配置 DAG 路由与流式异步检测。
深圳市大数据研究院（香港中文大学（深圳））
大模型 AI Infra 实习生 2026.06 - 2026.09
深圳市大数据研究院（香港中文大学（深圳））
大模型 AI Infra 实习生 2026.06 - 2026.09
在昇腾 910C 上优化长上下文推理、分布式部署与解码性能。
- 为 Qwen3-235B-A22B W8A8 落地 HSP score-once 在线压缩。
- LongBench v2 全量 387 条 >32K 样本中，平均 Prompt Token 降低 87.04%，TTFT p50 降低 83.89%，准确率提升 1.03 pp。
- 基于 vLLM-Ascend 与 CANN/ATB 8.5.1，完成 TP8 + EP8、A3 Flash Decoding 优化，持续解码吞吐由 4.71 提升至 27.99 tok/s（约 5.9 倍）。
- 优化 TAOmem embedding、KV 缓存与 lazy decode，使 BEAVER prepare p50 降低 31.26%，加入 batch token cap 等 OOM 防护。
绿盟科技集团股份有限公司广州分公司
安全实习生 2025.03 - 2025.04
绿盟科技集团股份有限公司广州分公司
安全实习生 2025.03 - 2025.04
参与南方电网及高校的安全设备交付、漏洞扫描与基线核查。
- 现场部署 WAF、IDS 等安全设备，进行策略调试与高可用验收。
- 执行客户现场定期漏洞扫描与基线核查，结合业务场景输出风险整改分析报告。
广州立源电子科技有限公司
信创技术助理（产品侧） 2024.06 - 2024.09
广州立源电子科技有限公司
信创技术助理（产品侧） 2024.06 - 2024.09
参与立源云在统信 UOS、麒麟 Kylin 环境中的兼容测试与部署适配。
- 负责国产操作系统环境下的兼容性测试与部署适配，沉淀信创技术方案文档。
07 / SKILLS
技能不是收藏
推理
从长上下文压缩到推理性能，关心代价，也关心答案有没有变。
PyTorch · vLLM-Ascend · W8A8 · HSP
安全
让检测过程留下证据，让安全判断能够复查。
garak · Semgrep · Prompt Injection · Jailbreak
昇腾 NPU
模型适配、精度对齐与分布式推理。
torch_npu · CANN / ATB · Ascend 910C
智能体
把模型、工具与上下文接进可恢复的执行流程。
MCP · Agent Skills · RAG · ChromaDB
基础设施
部署、并行和资源边界，都要在真实机器上检查。
Docker · Linux · TP8 / EP8 · KV Cache
后端
状态、数据与接口，是系统日常工作的部分。
FastAPI · PostgreSQL · Redis · WebSocket
工程工具
把重复的步骤自动化，把可能出错的地方留给测试。
Python · C/C++ · Git · pytest · CI/CD
08 / RECOGNITION
一些认可
- 挑战杯人工智能领域专项赛 · 全国一等奖
- 长城杯信息安全大赛 · 全国三等奖
- 全国大学生数学建模竞赛广东省一等奖
4 项已登记软件著作权
- 基于深度学习的文本生成智能平台 V1.0
2024SR1044803 - 证据盾录音器软件 [证据盾录音器] V1.0
2024SR2079705 - 智能知识库问答系统 [智能问答系统] V1.0
2026SR0546067 - AIGC内容安全检测系统 [AIGC检测系统] V1.0
2026SR0926607
另有 2 项发明专利申请，已受理 / 公开
09 / CONTACT
可以来找我
如果你也在做
LLM Systems · AI Safety · Agent
或者一些奇怪但有意思的东西
可以来找我
ABOUT / MOYUAN
MOYUAN
I build at the intersection of
AI systems, security, and software.
And every now and then,
I write about the world beyond the code.
01 / NOW
Right now
- Working on
- LLM inference, AI safety, and agent systems
- Thinking about
- how much context is actually worth paying for
- Tinkering with
- this website, as usual
02 / NAME
A name, a direction
Ink leaves a trace.
The rest is a matter
of direction.
鵷yuān
非梧桐不止
非练实不食
非醴泉不饮Shaped after the Yuanchu, a bird that chooses where it rests.
03 / METHOD
Give reality a say
Break the problem down.
Give reality a chance to prove you wrong.
Measure what can be measured.
Keep the evidence.
A model running is not the finish line.
The system has to work.
04 / ELSEWHERE
Elsewhere
05 / SELECTED WORK
Ideas, made real
PeerAssist
An evidence-driven AI review workspace for source reading, citation checks, human confirmation, and report exports.
Unlimited-OCR
Ascend NPU adaptation and comparable evaluation for multimodal OCR models.
AIGC Content Safety Detection and Audit System
Multimodal content-safety auditing with visual forgery evidence, keyword checks, and semantic retrieval.
06 / JOURNEY
The journey so far
Peking University Big Data Laboratory / Data & AI Security Group
AI Security Intern (remote) April 2026 - present
Peking University Big Data Laboratory / Data & AI Security Group
AI Security Intern (remote) April 2026 - present
Engineered JoySafeter evaluation workflows and audited the JoySafety framework.
- Connected source unpacking, isolated sandboxes, MCP/Skills tools, asynchronous workers, evidence archives, and structured reports.
- Reworked run-state transitions and stable IDs to resolve duplicate-consumption races and support idempotent recovery and retries.
- Validated 12 MCP services and 79 security tools with no loading failures, covering 53 backend and 21 frontend tests.
- Audited JoySafety moderation, prompt-injection defenses, multi-turn handling, knowledge-base responses, DAG routing, and asynchronous streaming detection.
Shenzhen Research Institute of Big Data / CUHK-Shenzhen
LLM AI Infrastructure Intern June - September 2026
Shenzhen Research Institute of Big Data / CUHK-Shenzhen
LLM AI Infrastructure Intern June - September 2026
Optimized long-context inference, distributed deployment, and decoding on Ascend 910C.
- Implemented HSP score-once online compression for Qwen3-235B-A22B W8A8.
- Across all 387 LongBench v2 samples over 32K, reduced mean prompt tokens by 87.04% and TTFT p50 by 83.89%, improving accuracy by 1.03 percentage points.
- Optimized TP8 + EP8 and A3 Flash Decoding with vLLM-Ascend and CANN/ATB 8.5.1, increasing sustained decoding from 4.71 to 27.99 tokens/s (about 5.9 times).
- Optimized TAOmem embeddings, KV caching, and lazy decoding, reducing BEAVER prepare p50 by 31.26%, with batch-token caps to guard against OOM.
NSFOCUS, Guangzhou
Security Intern March - April 2025
NSFOCUS, Guangzhou
Security Intern March - April 2025
Supported security deployments, vulnerability scans, and baseline checks for China Southern Power Grid and universities.
- Deployed WAF and IDS systems on site, tuned policies, and supported high-availability acceptance.
- Performed periodic vulnerability scans and baseline checks, producing remediation reports grounded in customer business scenarios.
Guangzhou Liyuan Electronics
Technical Product Assistant June - September 2024
Guangzhou Liyuan Electronics
Technical Product Assistant June - September 2024
Tested Liyuan Cloud compatibility and deployment on UOS and Kylin.
- Tested compatibility, adapted deployments, and documented technical solutions for domestic operating systems.
07 / SKILLS
Skills aren't collectibles
Inference
Long-context compression and inference performance, with answer quality kept in view.
PyTorch · vLLM-Ascend · W8A8 · HSP
Safety
Keep evidence from evaluation so safety decisions can be reviewed.
garak · Semgrep · Prompt Injection · Jailbreak
Ascend NPU
Model adaptation, numerical alignment, and distributed inference.
torch_npu · CANN / ATB · Ascend 910C
Agent
Models, tools, and context in workflows that can recover from failure.
MCP · Agent Skills · RAG · ChromaDB
Infra
Deployment, parallelism, and resource limits tested on real hardware.
Docker · Linux · TP8 / EP8 · KV Cache
Backend
The state, data, and interfaces that keep a system working.
FastAPI · PostgreSQL · Redis · WebSocket
Tools
Automate the repeatable parts; test where things can go wrong.
Python · C/C++ · Git · pytest · CI/CD
08 / RECOGNITION
A few milestones
- Challenge CupAI special competition · National first prize
- Great Wall CupInformation security · National third prize
- Mathematical ModelingNational competition · Guangdong first prize
4 registered software copyrights
- Deep-learning text generation platform V1.0
2024SR1044803 - Evidence Shield Recorder V1.0
2024SR2079705 - Intelligent knowledge-base question answering system V1.0
2026SR0546067 - AIGC content safety detection system V1.0
2026SR0926607
Also 2 invention patent applications, accepted or published
09 / CONTACT
Say hi
If you're building around
AI systems, safety, or agents
or just working on something
weird enough to be interesting.
