我专注于把 LLM 应用与 Agent 系统做成可靠的产品：路由、权限与副作用边界可审计；RAG 回答可追溯、可评测；后端行为用测试与指标证明。我习惯亲手走完「设计 → 实现 → 测试 → 上线」的完整链路，并把关键决策沉淀为仓库里的自动化测试、CI 检查与部署资产——让代码自己讲述它的可靠性。

I focus on turning LLM applications and agent systems into reliable products: auditable routing, permissions, and side-effect boundaries; RAG answers that are traceable and evaluable; backend behavior proven by tests and metrics. I walk the full path — design → implementation → testing → shipping — and encode every key decision into automated tests, CI checks, and deployment assets, so the code tells its own reliability story.

工作中我熟练运用 Vibe Coding——以 AI 编码助手为第一生产力，用自然语言驱动「设计 → 实现 → 测试 → 上线」的全流程，并深度使用 Codex、Claude Code、Cursor、Kimi、DeepSeek 等多模型组合，按任务特性灵活切换。我始终带着产品化意识做工程：不满足于「能跑」，而是追求可维护、可测试、可交付、真正被用户用起来。

In practice I'm fluent in Vibe Coding — AI coding assistants as the primary driver across design → implementation → testing → shipping — working hands-on with a multi-model stack (Codex, Claude Code, Cursor, Kimi, DeepSeek) chosen per task. I bring a product mindset to engineering: code that merely runs isn't the bar; it must be maintainable, testable, shippable, and usable by real people.

• 价值 · Value：面向企业知识问答、数据分析和受控业务操作的多 Agent 工作流系统——回答有据可查、操作有审批、执行有审计。已在内部业务场景中落地并持续运行，核心链路经过真实数据与流程验证。 A multi-agent workflow system for enterprise knowledge Q&A, data analysis, and controlled operations — answers with sources, operations with approvals, execution with audit trails. Deployed and running in internal production environments, with core paths validated against real data and workflows.
• 难点 · Challenges：多 Agent 协作的落地难题——任务流转的显式编排、知识与数据双通道检索、高风险操作的审批与权限、跨环节失败的一致性补偿、故障时的优雅降级；生产对抗性审查修复 12 个漏洞（含 3 个 P0 级认证/凭证类）。 Turning multi-agent orchestration into production — explicit workflow state machines, two-stage retrieval for knowledge and data, approvals and permissions for high-risk actions, consistent compensation when a step fails mid-chain, and graceful degradation; 12 vulnerabilities fixed in adversarial production review (3 P0 auth/credential class).
• 落地 · Shipped：101 项自动化测试、3 道 CI 安全门禁（gitleaks + pip-audit + Trivy）、工单系统真实接入（幂等键 + Saga 补偿 + 审计回写，stub 沙箱验收 7/7）、模型分层（高频 qwen3.5:4b + 推理 DeepSeek，成本降低 40%+）、三级降级链（LLM/检索/Checkpointer）、k6 阶梯压测与备份恢复演练。 101 automated tests, 3 CI security gates (gitleaks + pip-audit + Trivy), real ticket system pilot (idempotency keys + Saga compensation + audit write-back, stub sandbox 7/7), model tiering (high-freq qwen3.5:4b + reasoning DeepSeek, 40%+ cost reduction), three-stage degradation chain (LLM/retrieval/checkpointer), k6 staircase load tests and backup-restore drills.

PythonFastAPILangGraphMilvusPostgreSQLRedisOpenTelemetry

• 价值 · Value：面向气象业务文档的本地优先 RAG 应用——混合检索（BM25 + 向量 + RRF 融合）、证据不足时明确拒答、全链路阶段级可观测。零云依赖，本地运行。 A local-first RAG application for meteorological operational documents — hybrid retrieval (BM25 + vector + RRF fusion), explicit refusal on insufficient evidence, stage-level observability across the full pipeline. Zero cloud dependencies, runs entirely locally.
• 难点 · Challenges：中文专业文档检索的工程难题——CJK 分词避免单字过匹配的 Bigram 策略、BM25 与向量语义两路融合的 RRF 调参、证据评分门控防止大模型幻觉、扫描件 PDF 的按需 OCR（仅处理无文本层页面）、流式生成中断后的优雅降级、全链路阶段计时与状态追踪。 Real engineering challenges in Chinese domain document retrieval — CJK bigram tokenization to prevent single-character overmatching, RRF fusion of BM25 lexical and vector semantic paths, score-threshold gating to prevent LLM hallucination, on-demand OCR for scan PDFs (only pages with zero text layers), graceful fallback when streaming generation breaks, and full-pipeline stage timing with status tracking.
• 落地 · Shipped：49 项自动化测试（含检索评测脚本，支持 hit_rate / MRR / refusal_accuracy / 阶段延迟）、ruff + pytest CI、12 条评测问题（10 条可验证 + 2 条拒答测试）、Docker Compose 一键部署（API + Milvus + etcd + MinIO）、零依赖前端控制台（纯 HTML/CSS/JS，无构建步骤）。 49 automated tests (including evaluation CLI with hit_rate, MRR, refusal_accuracy, and per-stage latency), ruff + pytest CI, 12 evaluation questions (10 grounded + 2 refusal test cases), Docker Compose one-command deployment (API + Milvus + etcd + MinIO), zero-dependency frontend console (pure HTML/CSS/JS, no build step).

PythonFastAPIMilvusOllamaBM25RRFSQLiteSSE Streaming

• 价值 · Value：基于高德候选数据生成结构化行程，预算估算、天气、地图、品牌分享一站式输出。 Structured itineraries from Amap candidate data — budget, weather, maps, and branded shareable output in one flow.
• 难点 · Challenges：真实出行的工程问题——多数据源并行采集、AI 输出格式漂移的自动校验与修复、Kubernetes 部署就绪（多架构镜像 + lifecycle 探针 + Prometheus 指标）、结构化 JSON 日志与请求链追踪、前后端契约 CI 强制同步。 Real-world travel engineering — parallel fetching from multiple data sources, auto-validating and repairing AI output drift, Kubernetes readiness (multi-arch images + lifecycle probes + Prometheus metrics), structured JSON logging with request chain tracing, and CI-enforced frontend-backend contract sync.
• 落地 · Shipped：后端 83 项测试（行覆盖率 90%）+ 前端 Vitest、ruff + mypy strict + pytest + Vitest CI、多模型 LLM 适配（配置切换 OpenAI 兼容端点，默认 DeepSeek）、Docker Compose 一键启动。 83 backend tests (90% line coverage) + frontend Vitest, ruff + mypy strict + pytest + Vitest CI, multi-model LLM adapter (config-switchable OpenAI-compatible endpoints, default DeepSeek), one-command Docker Compose startup.

PythonFastAPIVue 3TypeScriptPydanticRedisKubernetesStructured Logging

• 价值 · Value：专注限量抢购正确性的工程实验——MySQL 始终是唯一可信的账本。从同步下单、Redis 抢购闸门、RocketMQ 消息化，到事务性 Outbox 可靠发布，每个版本都验证同一件事：抢得再猛，账也不会错。 A database-first flash-sale lab — MySQL is always the single source of truth. From synchronous ordering and Redis admission gates to RocketMQ messaging and transactional Outbox publication, each release proves one thing: under the heaviest rush, the books never break.
• 难点 · Challenges：抢购背后的真实并发难题——四种扣库存策略同台对比（含一个故意留的错误对照组）、10 条账目铁律、V4 事务性 Outbox（同一 MySQL 事务内持久化不可变信封，进程重启或 Broker 故障后恢复发布）、围栏对账（仅依据已提交的 MySQL 事实重建 Redis）、毒消息隔离与延迟过期消息加速结单。 The real concurrency puzzles behind flash sales — four inventory strategies compared head-to-head (including a deliberately unsafe control), ten hard bookkeeping rules, V4 transactional Outbox (immutable envelope persisted in the same MySQL transaction, recovered after process restart or Broker failure), fenced reconciliation (rebuilding Redis from committed MySQL facts only), poison-message isolation and delayed-expiry messages to accelerate order finalization.
• 落地 · Shipped：124 项自动化测试与 5 道正确性门禁——单元级状态机检查、Testcontainers 迁移与约束、确定性竞态与幂等、超额需求不变量、k6 HTTP 特征化测试；如实标注实验室定位，不做生产容量承诺。 124 automated tests with 5 correctness gates — unit-level state machine checks, Testcontainers migration and constraints, deterministic race and idempotency tests, over-demand invariant suite, k6 HTTP characterization tests; honestly scoped as a laboratory, with no production-capacity claims.

Java 21Spring BootMySQL/InnoDBRedis LuaRocketMQTransactional OutboxTestcontainersk6

• auto-coding：面向 AI 编码助手的风险感知交付 skill——按风险分级规划执行深度、三层自检 + 逃逸门检测、结构级契约检查、自适应工具链、中断后的安全恢复、CI 与版本发布。已重构为 14 份按需读取的 references，通过 Codex 插件分发。 A risk-aware delivery skill for AI coding assistants — risk-graded planning, three-layer self-check with escape hatch detection, structural contract checks, adaptive toolchain, safe recovery after interruptions, CI, and versioned releases. Refactored into 14 on-demand references, distributed via Codex plugin.
• EquiRebuild：Hermes 生态扩展——可配置的代码审查与加固 skill（5 维度对抗性审查），57 项测试（行覆盖率 95%），仅依赖标准库，新增实例级 timeout 追踪与高熵字符串检测。 A Hermes ecosystem extension — configurable code review and hardening skill (5-dimension adversarial review), 57 tests (95% line coverage), standard-library-only dependency, with instance-level timeout tracking and high-entropy string detection.
• daily-information：个人 AI 每日信息摘要管线——多源信号采集（GitHub / RSS / X）、去重与多维评分（相关性、新颖性、可信度）、LLM 策划与结构化输出验证、中文 Markdown 日报归档。四阶段管道设计（采集 → 候选 → 事件 → 发布），参考实现已跑通 Twitter 监控 + AI 摘要 + 飞书推送。 A personal AI daily digest pipeline — multi-source signal collection (GitHub / RSS / X), deduplication with multi-dimensional scoring (relevance, novelty, credibility), LLM curation with structured output validation, and Chinese Markdown daily archives. Four-stage pipeline design (collect → candidate → event → publish), with a reference implementation running Twitter monitoring + AI summarization + Feishu push.

• ✅ 让 Agent 的路由、权限和副作用边界可检查，而不是依赖模型临场发挥。 Making agent routing, permissions, and side-effect boundaries auditable — not left to model improvisation.
• 📚 让 RAG 回答可追溯来源、可评测质量，证据不足时明确拒答或降级——不幻想、不编造。 Making RAG answers traceable and evaluable — refusing or degrading clearly when evidence is insufficient. No hallucination, no fabrication.
• 🛡️ 让后端系统通过类型、测试、审计、指标与故障路径证明可靠性。 Making backend systems prove reliability through types, tests, audit trails, metrics, and well-defined failure paths.
• 🔍 让 AI 管线每一阶段可观测——检索、融合、重排、生成，每个环节的延迟和状态都可见、可复盘。 Making AI pipelines observable at every stage — retrieval, fusion, rerank, generation, each with visible latency and status, ready for post-mortem.

正在寻找 AI 应用工程 / Python 后端开发 相关机会，也欢迎交流 Agent、RAG、桌面端 AI 与云原生工程实践。

Open to AI Application Engineering / Python Backend Development roles — always happy to chat about agents, RAG, desktop AI, and cloud-native engineering practice.

Keep building. Keep verifying. Let what's shipped speak.
持续构建，持续验证，让交付说话。