What I’m doing
从当前正在寻找的方向开始，向下回溯这段学习、项目实践和写作历程。
Starting with what I’m looking for now, this page traces the study, projects, and writing that led here.
寻找下一段工作Looking for the next opportunity
当前一边继续维护这个知识花园，整理大模型、Agent 工具链、系统架构和项目复盘；一边寻找新的工作机会。目标方向包括大模型算法应用、AI 应用开发、全栈开发和 AI Agent 开发，也愿意参与把模型能力接入真实业务、完善工程基础设施并交付可用产品的团队。
I am maintaining this garden, studying LLMs, agent tools, and system architecture, while looking for my next role. I am interested in LLM applications, AI application development, full-stack development, and AI agent engineering, especially work that turns model capabilities into reliable products.
mini-agent 与 AI/架构实践mini-agent, AI, and architecture practice
5 月到 8 月，主线转向 mini-agent 的持续开发和复盘：从 Akka/Actor、事件总线和双实例入口，到多租户身份隔离、意图识别、会话刷新、SQLite 数据组织、Prompt Injection 防护、Bubblewrap/ECS 沙箱和公网安全整改；同时补上 Android 客户端、SSE 流式输出、推送与提醒、生产部署和故障排查。并行记录了 Pi、Claude、千问、Loop Engineering、RAG、LLM 基础设施、网络 API、可观测性和其他 AI/架构主题。
From May through August, I focused on building and reviewing mini-agent: Akka and Actor patterns, event bus and dual-instance entry points, multi-tenant isolation, intent routing, session refresh, SQLite organization, prompt-injection defenses, Bubblewrap/ECS sandboxes, and public-service security. I also worked through the Android client, SSE streaming, notifications, production deployment, and incident debugging, alongside notes on Pi, Claude, Qwen, Loop Engineering, RAG, LLM infrastructure, APIs, and observability.
AI-Interview 与 RL-Job-SchedulerAI-Interview and RL-Job-Scheduler
4 月把前一阶段的学习落到两个项目上。AI-Interview 这条线涉及 FastAPI 服务、知识库与 RAG 检索、Agent 编排、SSE 流式交互、数据库设计和面试流程；RL-Job-Scheduler 这条线则围绕强化学习调度、Master-Worker、Netty EventLoop 隔离、连接复用和故障恢复展开。期间还把 JWT、Redis、ThreadPool、Top-K、TypeScript、Promptfoo 等支撑技术拆成了独立博客。
In April, I applied the earlier study to two projects. AI-Interview covered a FastAPI service, knowledge bases and RAG retrieval, agent orchestration, SSE streaming, database design, and interview flows. RL-Job-Scheduler focused on reinforcement-learning scheduling, Master-Worker coordination, Netty EventLoop isolation, connection reuse, and recovery. I also turned supporting topics such as JWT, Redis, thread pools, Top-K, TypeScript, and Promptfoo into focused posts.
从基础开始学习Building the foundations
这一阶段先把后端工程的底座重新梳理了一遍：Java、Netty、Redis、Tomcat、线程池、数据库、网络协议、配置中心和 DevOps。随后把学习延伸到 Transformer、LLM Fine-Tuning、强化学习与 PPO/RLHF、向量检索、RAG、Prompt 评测和 Agent 编排，开始把模型能力放回真实的软件系统里理解。
I began by rebuilding my backend foundations across Java, Netty, Redis, Tomcat, thread pools, databases, network protocols, configuration centers, and DevOps. I then moved into Transformers, LLM fine-tuning, reinforcement learning and PPO/RLHF, vector retrieval, RAG, prompt evaluation, and agent orchestration to understand model capabilities in real systems.