CAPABILITY GRAPH / 能力图谱CAPABILITY GRAPH

模型、Agent、落地，
设计与安全贯穿。Models, agents, delivery —
design and security throughout.

  1. 模型训练Model trainingTRAINING

    数据决定上限：数据、后训练、AI for Science 与 AI Infra。Data sets the ceiling: data, post-training, AI for Science and AI infra.

  2. AgentAgentAGENT

    Agent 的架构、设计与开发，让它可靠地完成多步任务。Agent architecture, design and development, so it completes multi-step tasks reliably.

  3. 落地DeliveryFDE

    驻进业务，把 AI 从需求做到上线：政务工作台、律所系统。Embedded in the business, taking AI from requirements to launch: a government workbench, a law-firm system.

  4. 设计DesignDESIGN

    3D 设计与 3D 游戏交互，把技术做成能被感受的体验。3D design and 3D game interaction, turning technology into something you can feel.

  5. 安全SecurityBASE

    网络空间安全科班，也是其他方向的底线。Trained in cyberspace security; the baseline under every other direction.

方块是方向，圆点是技术，菱形通向能证明它的作品或文章。一根线跨到别的方向，说明这项技术在两边都用上了。Squares are directions, dots are technologies, diamonds open the work or writing that proves them. A line crossing into another direction means the skill serves both.

能力索引Capability index

01 / 数据与后训练Data & Post-Training DATA & POST-TRAINING

把数据质量、模型训练与结果验证连成闭环，让改进可以在固定评估集上比较。Close the loop from data quality to model training and evaluation, enabling measurable improvements.

技术：Skills: PyTorch、Transformers、TRL / PEFT、SFT、LoRA / QLoRA、DPO / ORPO、Rubric、A/B 评测PyTorch, Transformers, TRL / PEFT, SFT, LoRA / QLoRA, DPO / ORPO, Rubric, A/B evaluation

  • 数据标注、质检与事实、推理错误归因Data annotation, quality checks, and attribution of factual and reasoning errors
  • 偏好数据构造、指令微调与偏好对齐Preference-data construction, instruction fine-tuning and preference alignment
  • 固定评估集回归、消融对比与 BenchmarkRegression on fixed eval sets, ablations and benchmarks

有金融长文本、金法医等标注实践，在供应商侧参与腾讯混元 HY4 评测与质量管理；开展 Qwen、LLaMA 后训练实验。Annotation practice on long-form finance text and across finance, law and medicine; vendor-side evaluation and quality management for Tencent Hunyuan HY4; post-training experiments on Qwen and LLaMA.

数据研究Data research ↗后训练研究Post-training research ↗

02 / AI Infra 与科学 AIAI Infra & AI for Science AI INFRA / AI FOR SCIENCE

关注模型与 Agent 的运行基础，也把数据评估和算法实验用于科学问题。Focus on model & agent infrastructure, applying data evaluation and algorithmic experiments to scientific problems.

技术：Skills: E2B、Docker、WebSocket、Prompt Caching、KV Cache、GitHub Actions、AI for ScienceE2B, Docker, WebSocket, Prompt Caching, KV Cache, GitHub Actions, AI for Science

  • 控制面与执行面拆分、模型调用与缓存优化Splitting the control plane from the execution plane; optimising model calls and caching
  • 化学合成数据质检、化学与学科资料寻源Quality checks on chemical-synthesis data; sourcing chemistry and subject-matter references
  • 循环 Transformer、参数共享与提前退出实验Experiments with looped Transformers, parameter sharing and early exit

参与 Eazo 执行架构从 Worker 向沙箱迁移。科学方向的实践集中在化学数据质量、学科证据与推理算法实验。Took part in moving Eazo's execution architecture from Workers to sandboxes. On the science side, the practice centres on chemistry data quality, subject-matter evidence and reasoning-algorithm experiments.

AI Infra 研究AI infra research ↗AI for ScienceAI for Science ↗

03 / Agent 架构与开发Agent Engineering AGENT ENGINEERING

设计 Agent 的架构：执行环境、工具、上下文与多 Agent 协作，再把它开发成能可靠完成多步任务的产品。Design agent architectures—execution environments, tools, context, and multi-agent orchestration—into reliable products.

技术：Skills: Harness、MCP、Skills、多 Agent 编排、RAG、Memory、Session / Agent LoopHarness, MCP, Skills, Multi-agent orchestration, RAG, Memory, Session / Agent Loop

  • 插件生命周期、工具接入与能力按需加载Plugin lifecycle, tool integration and on-demand capability loading
  • 上下文管理、会话轨迹与跨任务记忆Context management, session trajectories and cross-task memory
  • 多 Agent 协作、执行沙箱与权限控制Multi-agent collaboration, execution sandboxes and permission control

参与 DeepSeek Harness 交互终端、Evolver 桌面接入与 OpenClaw 扩展；自建 Agent Memory，开发本地 AI 工作台。Contributed to the DeepSeek Harness interactive terminal, the Evolver desktop integration and OpenClaw extensions; built my own agent memory and a local AI workbench.

Agent Memory 范式Agent memory paradigms ↗Claude OrbitClaude Orbit ↗

04 / FDE · AI 应用落地Forward Deployed Engineering FORWARD DEPLOYED

驻进业务现场，从需求梳理、方案设计到开发、部署和交付，让 AI 能力变成业务里每天在用的系统。Embed into operations—from requirements to deployment—turning AI capabilities into daily-use systems.

技术：Skills: 需求梳理、方案设计、私有化部署、TypeScript、Python、Rust、C / C++、React、Next.js、Node.js / Bun、Fastify、PostgreSQL、SQLiteRequirements discovery, Solution design, On-prem deployment, TypeScript, Python, Rust, C / C++, React, Next.js, Node.js / Bun, Fastify, PostgreSQL, SQLite

  • 进入客户业务，梳理流程与需求，确定 AI 能介入的环节Embedding in the client's business, mapping workflows and requirements, and finding where AI can step in
  • 多租户数据隔离、角色权限与知识库接入Multi-tenant data isolation, role permissions and knowledge-base integration
  • 前后端开发、部署联调，到上线后的真实使用与迭代Front-end and back-end development, deployment and integration, through to real use and iteration after launch

参与政务多 Agent 工作台从方案设计到部署交付；独立完成律所案件管理系统的需求、开发与上线，用 MCP 接入案件数据库、文书解析、企业微信归档与周报推送。Worked on a government multi-agent workbench from solution design to deployment; independently delivered a law-firm case-management system from requirements to launch, using MCP to connect the case database, document parsing, WeCom archiving and weekly-report delivery.

opendevopendev ↗开发者工具Developer tools ↗

05 / 3D 设计与游戏交互Design & 3D DESIGN & 3D

做 3D 设计与 3D 游戏的交互设计，也构建界面与动效，把技术做成人能感受到的体验。Design 3D experiences and game interactions, building interfaces and motion into tangible experiences.

技术：Skills: 设计系统、UI / UX、Blender、Three.js、3D 建模与渲染、3D 游戏、高斯泼溅、After Effects、Lottie、AIGCDesign systems, UI / UX, Blender, Three.js, 3D modelling & rendering, 3D games, Gaussian splatting, After Effects, Lottie, AIGC

  • 设计系统、界面规则与交互流程设计Design systems, interface rules and interaction-flow design
  • 建模、材质、实时渲染与三维游戏交互Modelling, materials, real-time rendering and 3D game interaction
  • AE / Lottie 动效与 AIGC 视觉生成AE / Lottie motion and AIGC visual generation

实践见 ART LOOKS BACK、种子与花园和克莱因瓶的漩涡，覆盖三维展览、实体装置与文学叙事。See ART LOOKS BACK, Seeds & Garden and The Klein Bottle Vortex: a 3D exhibition, a physical installation and a literary narrative.

精选作品Selected work ↗进入我的世界Enter my world ↗

06 / 网络与 AI 安全Network & AI Security SECURITY

网络空间安全科班出身，将安全分析、访问控制与 AI 工具的实际执行放在一起考虑。Cybersecurity background, integrating security analysis, access control, and AI tool execution.

技术：Skills: 网络空间安全、OWASP Top 10、流量分析、漏洞扫描、提示词注入、沙箱、多租户隔离Cybersecurity, OWASP Top 10, Traffic analysis, Vulnerability scanning, Prompt injection, Sandboxing, Multi-tenant isolation

  • 授权范围内的安全测试、漏洞分析与安全运维Authorised security testing, vulnerability analysis and security operations
  • 提示词注入检测与防御、工具权限与只读沙箱Prompt-injection detection and defence, tool permissions and read-only sandboxes
  • 租户数据隔离、角色权限与联网边界Tenant data isolation, role permissions and network boundaries

网络空间安全本科在读，有国家级护网与重保经历，持续开展 AI 安全研究与工程实践。Undergraduate in cyberspace security, with experience in national-level cyber-defence exercises and critical-event protection; ongoing AI security research and engineering.

AI 安全项目AI security project ↗提示注入研究Prompt-injection research ↗