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.

近期我持续探索 DeepSeek Harness 这一受到关注的开源 AI coding harness 项目，并将仓库理解、证据驱动分析与有界变更生命周期适配为 dsh-Repo-Atlas · RepoAtlas 插件，重点关注公开工具注册、bundle 加载、兼容性验证与默认只读的安全边界。

Recently I've been following the fast-moving open-source DeepSeek Harness project and adapting repository understanding, evidence-backed analysis, and bounded change lifecycles into the dsh-Repo-Atlas · RepoAtlas plugin, with a focus on public tool registration, bundle loading, compatibility verification, and read-only-by-default safety boundaries.

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：近期持续探索 DeepSeek Harness 这一受到关注的开源 AI coding harness 项目，并完成 dsh-Repo-Atlas 适配插件——面向陌生代码库提供安全优先、证据驱动的分析与有界变更生命周期；默认只读，受控动作显式审批。 Recently exploring the fast-moving open-source DeepSeek Harness ecosystem and building dsh-Repo-Atlas as an adapted plugin — safety-first, evidence-backed repository analysis and a bounded change lifecycle for unfamiliar codebases; read-only by default, with explicit approval for controlled actions.
• 难点 · Challenges：将代码库接手、证据与推断区分、Goal/approval policy gate、ReAct 分析、补丁审阅与验证串成可审计的生命周期，同时保持与 Harness 核心解耦，不扩大 Shell、网络、依赖安装或远程写入权限。 Turning repository onboarding, evidence-versus-inference boundaries, Goal/approval policy gates, ReAct analysis, patch review, and verification into an auditable lifecycle while keeping the plugin decoupled from Harness core and avoiding broader Shell, network, dependency-installation, or remote-write permissions.
• 落地 · Shipped：完成公开 ctx.tools.register 接入、dsh-repo-atlas/harness bundle 加载与 pinned Harness revision 的真实兼容性 smoke，采用 source-first 方式发布 0.1.1。 Delivered public ctx.tools.register integration, dsh-repo-atlas/harness bundle loading, and real compatibility smoke coverage against a pinned Harness revision; released 0.1.1 as a source-first plugin bundle.

TypeScript Node.js DeepSeek Harness Evidence-backed Analysis Sandbox-aware Lifecycle

• 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.

Python FastAPI LangGraph Milvus PostgreSQL Redis OpenTelemetry

• 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).

Python FastAPI Milvus Ollama BM25 RRF SQLite SSE 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.

Python FastAPI Vue 3 TypeScript Pydantic Redis Kubernetes Structured Logging

• 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.
• dsh-Repo-Atlas · RepoAtlas：DeepSeek Harness 插件设计落地——围绕陌生代码库接手、证据驱动分析、公开工具注册、bundle 加载与 pinned revision 兼容性验证，推进默认只读、可审计的变更提案流程。 dsh-Repo-Atlas · DeepSeek Harness plugin design and delivery — repository onboarding, evidence-backed analysis, public tool registration, bundle loading, and pinned-revision compatibility verification, shaping a read-only-by-default and auditable change-proposal lifecycle.

• ✅ 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。