🔬 Research & internships
Peking University Big Data Laboratory / Data & AI Security Group
AI Security Intern, remote · April 2026–present
- Engineered execution workflows for the JoySafeter agent security evaluation platform: source unpacking, isolated sandboxes, MCP and Skills tools, asynchronous workers, evidence archives, and structured reports.
- Reworked run-state transitions and stable IDs to resolve concurrent duplicate consumption and support idempotent recovery and retries.
- Integrated tool schemas, risk labels, and evidence from tools including Semgrep and garak. Validated 12 MCP services and 79 security tools with no loading failures, covering 53 backend and 21 frontend tests.
- Audited JoySafety’s input/output moderation, prompt-injection defenses, multi-turn handling, knowledge-base responses, configurable DAG routing, and asynchronous streaming detection.
Shenzhen Research Institute of Big Data / CUHK-Shenzhen
LLM AI Infrastructure Intern · June–September 2026
- Implemented HSP score-once online compression for long-context inference with Qwen3-235B-A22B W8A8 on Ascend 910C NPUs.
- On all 387 LongBench v2 samples longer than 32K, reduced mean prompt tokens by 87.04% and p50 time to first token by 83.89%, with accuracy improving by 1.03 percentage points.
- Optimized distributed inference using vLLM-Ascend and CANN/ATB 8.5.1, including TP8 + EP8 and A3 Flash Decoding. Increased sustained decoding throughput from 4.71 to 27.99 tokens/s, approximately 5.9×.
- Optimized embeddings, KV caching, and lazy decoding in TAOmem, reducing BEAVER prepare p50 by 31.26%, and added batch-token limits to guard against OOM failures.
NSFOCUS, Guangzhou
Security Intern · March–April 2025
Supported security deployments for China Southern Power Grid and universities, including WAF and IDS configuration, policy tuning, high-availability acceptance, vulnerability scans, baseline checks, and remediation reports.
Guangzhou Liyuan Electronics
Technical Product Assistant · June–September 2024
Tested compatibility and deployment of Liyuan Cloud on UOS and Kylin operating systems and documented technical solutions.
🚀 Open-source & applied projects
1. PeerAssist — Evidence-driven AI review workspace
Independent design and full-stack development · July 2026–present · Repository
- Connects PDF range reading, line-level evidence highlighting, citation checks, human confirmation, and immutable report exports.
- A ReviewJob state machine and event timeline support interruption recovery, cancellation, and idempotent retries.
- Built with FastAPI, React, PostgreSQL, Keycloak, MinIO, and Celery workers.
2. Unlimited-OCR — Ascend NPU adaptation and evaluation
Independent development · July 2026 · Repository
- Adapts OCR multimodal models and CUDA operations to torch_npu, addressing AiCPU fallbacks and numerical alignment.
- Achieved 49/76 (64.47%) on the OCRBench classic STR subset under the same evaluation protocol as the official GPU implementation, with 0.73 seconds per image.
3. AIGC Content Safety Detection and Audit System
Independent full-stack development · February–August 2026 · Repository
- Awarded national third prize in the Great Wall Cup and recognized as an outstanding graduation project.
- Combines CLIP ViT-L/14, PEFT/LN-tuning, multimodal models, and PaddleOCR to produce visual forgery evidence.
- Uses ChromaDB and a knowledge base of over 70,000 sensitive terms and policy boundaries for keyword and semantic moderation, with heatmaps and structured audit recommendations.
4. NullHunter-Agent — MCP multi-agent security platform
Independent design and development · May 2026
- A Tencent security hackathon entry for CTF and competition labs.
- Uses a ChallengeState state machine and deterministic ToolRouter to coordinate reconnaissance, JavaScript analysis, exploitation, forum research, and reporting agents, from initial analysis to flag submission and review.
5. Real-time multimodal digital-human assistant
Independent development · January 2026–present · Repository
- Connects ASR, intent parsing, LLM responses, TTS, and lip synchronization.
- FastAPI and WebSocket provide full-duplex interaction; input filtering and jailbreak defenses protect the generation workflow, while Redis supports per-IP rate limits.
6. RAG-based anti-fraud question answering
Project Lead · February 2023–September 2024
- Built a multi-agent planning, retrieval, and risk-analysis workflow for tracing complex fraud scenarios.
- Fine-tuned open models with LoRA for domain-specific recognition of scam language and early warnings.
🛠 Technical stack
| Area | Technologies & tools |
|---|---|
| LLM systems | PyTorch, torch_npu, vLLM-Ascend, CANN/ATB, Ascend 910C, W8A8 quantization, TP8/EP8, Flash Decoding, HSP compression. |
| AI safety | garak, prompt injection and jailbreak defenses, Semgrep, moderation, evidence-based audits, and red/blue-team evaluation. |
| Agent and backend engineering | MCP, Agent Skills, FastAPI, WebSocket, PostgreSQL, Redis, ChromaDB, Docker, Linux, Caddy. |
| Cybersecurity | web logic vulnerabilities, authorization and concurrency issues, API testing, Burp Suite, traffic analysis, and binary exploitation debugging. |
| Development | Python, C/C++, Shell, Git, uv, pytest, CI/CD, Cursor, Claude Code, and Copilot. |
🏆 Honors & qualifications
- National first prize, 19th Challenge Cup special competition in AI.
- First prize (silver award), 2024 Greater Bay Area AI for Science competition.
- National third prize, 19th National College Student Information Security Competition / 3rd Great Wall Cup.
- National second prize, ISCC 2024 individual challenge.
- Guangdong first prize, National College Student Mathematical Modeling Competition.
- International first prize, Huashu Cup mathematical modeling competition.
- Finalist award, 2nd National Advanced Computing Technology Innovation Competition.
- First prize, university network attack-and-defense invitational.
Two invention patent applications accepted or published; four registered software copyrights.
- Deep-learning text generation platform V1.0
基于深度学习的文本生成智能平台 V1.0
Registration: 2024SR1044803; certificate: 13448676; registered: . - Evidence Shield Recorder V1.0
证据盾录音器软件 [证据盾录音器] V1.0
Registration: 2024SR2079705; certificate: 14483578; registered: . - Intelligent knowledge-base question answering system V1.0
智能知识库问答系统 [智能问答系统] V1.0
Registration: 2026SR0546067; certificate: 17760348; registered: . - AIGC content safety detection system V1.0
AIGC内容安全检测系统 [AIGC检测系统] V1.0
Registration: 2026SR0926607; certificate: 18140888; registered: .
Qualifications include Information Security Engineer, NISP Level 2, National Computer Rank Examination Levels 3/4 in information security, and a CNVD-confirmed medium-severity vulnerability. University honors include outstanding student and student leadership awards, and Qianxin innovation and merit scholarships.
Contact
WeChat publication: SecAI前沿, author of the “Growth for Technical Professionals in the AI Era” series.
🔬 科研与实习经历
🏛 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
- JoySafeter 智能体安全测评平台工程化:
- 面向大模型安全评测工作流,负责测评任务执行链路,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自动生成。
- 重构 Run 状态迁移与稳定 ID 机制,修复重复消费引发的并发竞态,实现任务恢复与失败重试幂等性。
- 统一接入 Semgrep、garak 等工具的 Schema、风险标签与调用证据,完成 12 个 MCP 服务、79 个安全工具加载验收(0 失败),并覆盖 53 项后端与 21 项前端测试。
- JoySafety 框架源码审计与分析:
- 深入分析企业级大模型安全中间层源码,梳理输入/输出双向审核、Prompt Injection、多轮会话与红线知识库代答链路;解析可配置 DAG 路由与流式异步检测机制,沉淀完整的技术架构与工程文档。
🏢 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
- 长上下文推理优化(gde-rag):
- 针对
Qwen3-235B-A22B W8A8在昇腾 910C NPU 上的长上下文推理进行端到端优化,落地 HSP score-once 在线压缩方案。 - 在 LongBench v2 全量 387 条 >32K 长文本样本评测中,平均 Prompt Token 降低 87.04%,首 Token 延迟(TTFT)p50 降低 83.89%,准确率(Accuracy)逆势提升 1.03 pp。
- 针对
- NPU 推理部署与分布式性能优化:
- 基于
vLLM-Ascend与CANN/ATB 8.5.1完成分布式并行优化(TP8 + EP8、A3 Flash Decoding),将持续解码吞吐由 4.71 提升至 27.99 tok/s(提升约 5.9 倍)。 - 针对 TAOmem 链路优化 embedding、KV 缓存与 lazy decode,使 BEAVER prepare p50 降低 31.26%,并构筑 batch token cap 等硬件 OOM 防线。
- 基于
🛡 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
- 企业安全交付:参与南方电网及高校的企业级防护体系建设,负责 WAF、IDS 等核心安全设备的现场部署、策略调试与高可用验收。
- 漏洞扫描与基线核查:前往客户现场执行定期的漏扫与基线核查,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
- 负责“立源云”解决方案在国产操作系统(统信 UOS / 麒麟 Kylin)环境下的兼容性测试与部署适配,沉淀信创技术方案文档。
🚀 代表性开源与实战项目
1. PeerAssist · 证据驱动 AI 审稿工作台
独立设计与全栈开发 | 2026.07 - 至今 | GitHub 仓库
- 核心定位:针对学术审稿中幻觉严重、断章取义的痛点,构建具备强证据约束的 AI 审稿工作流。
- 架构亮点:
- 打通 PDF Range 原文阅读、行级证据高亮标注、事实引用核查、人工确认与不可变评审报告导出。
- 以 ReviewJob 状态机与事件时间线实现审稿任务的中断恢复、动态取消与幂等重试。
- 技术栈:
FastAPI+React+PostgreSQL+Keycloak+MinIO+Celery/Worker异步分布式架构。
2. Unlimited-OCR · 昇腾 NPU 适配与评测
独立开发 | 2026.07 | GitHub 仓库
- 核心定位:将前沿 OCR 多模态模型迁移至国产华为昇腾算力生态。
- 技术突破:
- 将模型中的 CUDA 算子与推理逻辑精准映射至
torch_npu,排查解决 AiCPU 算子回退与 NPU 精度对齐难题。 - 在 OCRBench 经典 STR 子集上准确率达到 49/76(64.47%),达到与 GPU 官方实现同等口径效果,单张推理耗时仅 0.73s。
- 将模型中的 CUDA 算子与推理逻辑精准映射至
3. AIGC 内容安全检测与审计系统
“长城杯”全国大学生信息安全竞赛三等奖 | 优秀毕业设计 | 独立全栈开发 | 2026.02 - 2026.08 | GitHub 仓库
- 核心定位:面向生成式 AI 伪造图文难识别、检测黑盒不透明的问题,构建可解释的多模态内容安全审计系统。
- 技术突破:
- 基于
CLIP ViT-L/14、PEFT / LN-tuning、多模态大模型(MLLM)与 PaddleOCR 构建视觉伪造证据链。 - 基于
ChromaDB构建 70,000+ 条敏感词与红线知识库,实现关键词与语义向量混合检索审核,输出带证据热力图与结构化建议的安全审计报告。
- 基于
4. NullHunter-Agent · 智能渗透测试 MCP 多智能体平台
独立设计与开发 | 2026.05 | 腾讯智能渗透黑客松参赛作品
- 核心定位:面向 CTF 与安全攻防比赛靶场的多智能体自主渗透闭环平台。
- 架构设计:基于
ChallengeState状态机与确定性ToolRouter,将任务自动解耦并协同分发给 ReconAgent、JSReverseAgent、ExploitAgent、ForumAgent 与 ReportAgent,实现从入口侦察、JS 逆向、漏洞验证到 Flag 自动提交与复盘报告全自动化。
5. 实时多模态数字人智能体交互系统
独立开发 | 2026.01 - 至今 | GitHub 仓库
- 低延迟 AI 数字人语音交互系统,打通 ASR 语音识别、意图解析、LLM 对话生成、TTS 语音合成与唇形同步驱动。
- 基于
FastAPI与WebSocket实现全双工低延迟通信;在 LLM 生成链路嵌入输入敏感词过滤与越狱防护约束;引入 Redis 实现 IP 级高并发限流保护。
6. 智谷谣擎 · 基于 RAG 的反诈风控问答系统
项目负责人 | 2023.02 - 2024.09
- 主导引入 Multi-Agent 协同架构,构建“规划-检索-风控”自动化流水线,解决传统单轮 RAG 难以应对复杂诈骗逻辑溯源的难题。
- 采用 LoRA 对开源大模型进行垂直领域微调,显著提升诈骗话术捕捉与预警能力。
🛠 技能图谱 · Tech Stack
| 领域 | 核心技术与工具 |
|---|---|
| 大模型系统与 AI Infra | PyTorch, torch_npu, vLLM-Ascend, CANN / ATB, 昇腾 910C, W8A8 量化, TP8/EP8 并行, Flash Decoding, HSP 长上下文压缩 |
| AI 安全与内容风控 | garak, Prompt Injection, Jailbreak 越狱防御, Semgrep 源码沙箱, 敏感词审核, 证据化审计报告, 红蓝对抗评测 |
| Agent 与全栈后端 | MCP (Model Context Protocol), Agent Skills, FastAPI, WebSocket, PostgreSQL, Redis, ChromaDB, Docker, Linux, Caddy |
| 网络安全与攻防 | Web 业务逻辑漏洞挖掘 (越权/并发等), API 接口渗透, Burp Suite, 流量分析, 二进制溢出调试 (PWN) |
| 开发工程素养 | Python, C/C++, Shell, Git, uv, pytest, CI/CD, Cursor / Claude Code / Copilot 深度开发流 |
🏆 荣誉证书与知识产权
🏅 核心竞赛荣誉
- 🏆 第十九届“挑战杯”全国大学生课外学术科技作品竞赛“揭榜挂帅”专项赛(人工智能领域) —— 全国一等奖
- 🥈 2024 年“大湾区杯” AI for Science 科技竞赛 —— 一等奖(银奖)
- 🥉 第十九届全国大学生信息安全竞赛暨第三届“长城杯”信息安全大赛 —— 全国三等奖
- 🥈 第 21 届信息安全与对抗技术竞赛(ISCC2024)个人挑战赛 —— 全国二等奖
- 🥇 全国大学生数学建模竞赛 —— 广东省一等奖
- 🥇 “华数杯”国际大学生数学建模竞赛 —— 国际一等奖
- 🥉 第二届全国先进计算技术创新大赛总决赛 —— 决赛获奖
- 🥇 校第一届“网络攻防”邀请赛 —— 一等奖
📜 知识产权与学术成果
- 📑 国家发明专利申请:2 项(已受理/公开)
💻 计算机软件著作权:4 项(均已登记)
基于深度学习的文本生成智能平台 V1.0:登记号 2024SR1044803;证书号 13448676;登记日期 2024-07-23。
- 证据盾录音器软件 [证据盾录音器] V1.0:登记号 2024SR2079705;证书号 14483578;登记日期 2024-12-13。
- 智能知识库问答系统 [智能问答系统] V1.0:登记号 2026SR0546067;证书号 17760348;登记日期 2026-04-10。
- AIGC内容安全检测系统 [AIGC检测系统] V1.0:登记号 2026SR0926607;证书号 18140888;登记日期 2026-09-10。
🎖 专业资格证书
- 网络安全:软考中级(信息安全工程师)· 国家信息安全水平考试(NISP 二级)· 全国计算机等级考试三级/四级(信息安全技术)· CNVD 国家信息安全漏洞平台认证中危漏洞
- 综合荣誉:校园“十佳三好学生” · 奇安信实践科创奖学金 · 奇安信优秀学生奖学金 · 优秀学生干部
📬 联系方式 · Contact
- 📧 Email:2572045628@qq.com
- 🐙 GitHub:@moyuan10086
- 🌐 Personal Website:moyuan10086.github.io
- 📝 WeChat Official Account:SecAI前沿(公众号《AI时代技术人成长系列》专栏作者)
