ABOUT / MOYUAN
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
02 / NAME
墨落
风起
羽翼向远
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
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
05 / SELECTED WORK
面向 Agent 的长上下文编排项目，参与上下文压缩与昇腾 NPU 推理优化。
OCR 多模态模型的昇腾 NPU 适配与同口径评测。
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 / RESEARCH
最近在研究一个问题：当任务最终只需要一个决策时，我们真的还需要让大模型生成 Token 吗？
Hao Chen · arXiv 预印本 ·
Calibrated Standalone Direct-Decision Guardrails with Heterogeneous CPU-NPU Confidence Cascading under Explicit Latency and False-Positive Constraints
面向实时 LLM 安全护栏，探索什么时候需要生成、什么时候可以直接决策。将 CPU 快速门控与 NPU 上的直接决策模型级联，在显式时延与误报约束下研究模型路由、概率校准与安全检测效率。
Hao Chen · arXiv 预印本 ·
Controlled Study of Autoregressive Generation versus Direct Decision Interfaces for Closed Multimodal Forensic Tasks
研究封闭式多模态取证任务中，自回归生成是否真的有必要。在统一 Qwen3-VL-8B 主干下，对比生成式输出、Logit Slice、Binary Head 与对比式直接决策接口，分析准确率、校准能力与推理时延之间的关系。
07 / JOURNEY
AI 安全实习生（远程） 2026.04 - 至今
担任 AI 大模型与 Agent 安全测评平台核心开发，基于 JoySafeter 早期框架进行深度二次开发与核心架构重构，将 Agent 对话与单点工具原型扩展为面向 LLM、Agent、RAG 与 MCP 生态的统一安全测评平台。
大模型 AI Infra 实习生 2026.06 - 2026.09
参与面向 Agent 长上下文的 TAOctx 相关研发，在昇腾 910C 上开展上下文压缩、长上下文推理与分布式解码性能优化。所在团队成果于 HUAWEI CONNECT 2026（华为全联接大会）开放演讲中展示，项目已开源。
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。
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。
08 / SKILLS
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
让检测过程留下证据，让安全判断能够复查。
garak · Semgrep · Prompt Injection · Jailbreak
模型适配、精度对齐与分布式推理。
torch_npu · CANN / ATB · Ascend 910C
把模型、工具与上下文接进可恢复的执行流程。
MCP · Agent Skills · RAG · ChromaDB
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
状态、数据与接口，是系统日常工作的部分。
FastAPI · PostgreSQL · Redis · WebSocket
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
09 / RECOGNITION
2024SR10448032024SR20797052026SR05460672026SR0926607另有 2 项发明专利申请，已受理 / 公开
10 / CONTACT
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
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
02 / NAME
Ink leaves a trace.
The rest is a matter
of direction.
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
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
05 / SELECTED WORK
A context runtime for long-running agents, with contributions to context compression and Ascend NPU inference optimization.
Ascend NPU adaptation and comparable evaluation for multimodal OCR models.
Multimodal content-safety auditing with visual forgery evidence, keyword checks, and semantic retrieval.
06 / RESEARCH
When a task only needs a decision, does a foundation model still need to generate tokens?
Hao Chen · arXiv preprint ·
Calibrated Standalone Direct-Decision Guardrails with Heterogeneous CPU-NPU Confidence Cascading under Explicit Latency and False-Positive Constraints
Explores when real-time LLM guardrails need to generate tokens and when they can decide directly. Combines a fast CPU gatekeeper with an NPU direct-decision model to study routing, probability calibration, and detection efficiency under explicit latency and false-positive constraints.
Hao Chen · arXiv preprint ·
Controlled Study of Autoregressive Generation versus Direct Decision Interfaces for Closed Multimodal Forensic Tasks
Studies whether closed multimodal forensic tasks require autoregressive generation. With a matched Qwen3-VL-8B backbone, compares generation, Logit Slice, Binary Head, and contrastive direct-decision interfaces across accuracy, calibration, and inference latency.
07 / JOURNEY
AI Security Intern (remote) April 2026 - present
Core developer of an AI and agent security evaluation platform. Substantially extended and rearchitected the early JoySafeter framework, evolving its agent-chat and individual-tool prototype into a unified platform for evaluating LLM, agent, RAG, and MCP security.
LLM AI Infrastructure Intern June - September 2026
Contributed to TAOctx research and development for long-context agents, working on context compression, long-context inference, and distributed decoding performance on Ascend 910C. The team's work was presented in an open talk at HUAWEI CONNECT 2026, and the project is open source.
Security Intern March - April 2025
Supported security deployments, vulnerability scans, and baseline checks for China Southern Power Grid and universities.
Technical Product Assistant June - September 2024
Tested Liyuan Cloud compatibility and deployment on UOS and Kylin.
08 / SKILLS
Long-context compression and inference performance, with answer quality kept in view.
PyTorch · vLLM-Ascend · W8A8 · HSP
Keep evidence from evaluation so safety decisions can be reviewed.
garak · Semgrep · Prompt Injection · Jailbreak
Model adaptation, numerical alignment, and distributed inference.
torch_npu · CANN / ATB · Ascend 910C
Models, tools, and context in workflows that can recover from failure.
MCP · Agent Skills · RAG · ChromaDB
Deployment, parallelism, and resource limits tested on real hardware.
Docker · Linux · TP8 / EP8 · KV Cache
The state, data, and interfaces that keep a system working.
FastAPI · PostgreSQL · Redis · WebSocket
Automate the repeatable parts; test where things can go wrong.
Python · C/C++ · Git · pytest · CI/CD
09 / RECOGNITION
2024SR10448032024SR20797052026SR05460672026SR0926607Also 2 invention patent applications, accepted or published
10 / CONTACT
If you're building around
AI systems, safety, or agents
or just working on something
weird enough to be interesting.
