从业务本质到AI智能体编排 From Business Essence to AI Agent Orchestration

针对跨国税制差异与高新技术企业税务场景，1人独立设计发票合规风控 Agent。利用网页爬虫抓取最新案例，构建规则引擎进行批量预警，实现从“被动合规”到“主动风控”。 1-man independent design & deployment of a tax risk Agent addressing cross-border tax variations. Implemented web scrapers for policy retrieval and rule engines for batch risk grading.

• 发票异常识别准确率达 >95%Abnormal invoice identification accuracy >95%
• 财务部门工具采纳率 100%Finance department adoption rate 100%
• 4周内完成从需求、开发到公网部署全流程4-week delivery from requirements to edge deploy

整合福建省历年招生数据，融合专业咨询师决策逻辑与130+院校行业映射模型，搭建 AI 填报咨询助手。支持国内多模型混合路由与高速全球边缘访问。 Synthesized regional admissions datasets, mapped 130+ college networks, and built a conversational admissions consultant Agent using multi-model routers and global edge networks.

• 前端全球边缘加速访问延迟仅 0.3-0.6sEdge caching delivers global latency of 0.3-0.6s
• 双模型路由及基于 Redis 的高并发控制Multi-model fallback routing and rate limits

针对纺织服装制造外贸员工口语培训痛点，开发基于 LLM 工作流的多 Agent 协同训练系统。包含实时互动对话、口语精准评分、生词本抓取与 Anki 卡片自动导出。 Designed a multi-agent oral training pipeline to lower corporate employee language upskilling costs. Incorporates automated score grading and Anki flashcard generation.

• 闭环：对话 → PracticeView 评估 → Anki 导出Closed-loop: dialogue, performance scoring, Anki sync
• 多 Agent 协作工作流，多Provider 灵活切换Multi-agent orchestrations with flexible APIs

从0到1研发“智能跑鞋”、“智能心电监测背心”和“智能足球护腿板”三大穿戴式硬件，并主导配套 APP 的产品规划与数据服务闭环。构建了“硬件 + APP + 数据算法”的业务交付模型。 Developed and scaled smart running shoes, ECG compressed shirts, and smart shin guards. Managed the full hardware + APP + biosensing algorithms lifecycle.

• 智能跑鞋累计出货 200万双 (合作多家国内头部运动服饰品牌)Smart shoes shipped over 2,000,000 pairs (with top national brands)
• 心电背心出货 20,000件 (与军事医科院校合作)Smart ECG vests shipped 20,000 units (with military medical university)

规划实施外部供应链协同平台，连通供应商接单、生产进度、结算流程。实现基于 ERP 的精细化调度与快速反应，打通了“异常预警-智能决策”的前身逻辑。 Implemented the bulk SCM collaboration portal integrating information flow, funding routes, and ordering pipelines. Achieved real-time ERP order tracking and delay warnings.

• 电商供应链反应周期从 80天压缩至15天 (-81%)E-commerce SCM cycle down from 80 days to 15 days (-81%)
• 供应商接单周期由 7-10天缩短为2天 (-70%+)Vendor confirmation down from 7-10 days to 2 days

针对大规模个性化服装定制，自研 3D 前台设计、中台备料与后端产线采集系统。对自有印花及针织工厂进行数字化升级，奠定初期智能流程路由与异常分派逻辑。 Built a custom 3D design editor, inventory compiler, and workshop IoT data collection system to allow one-click ordering and live progress tracking.

• 接单到出货全流程透明，服装产能提升 +50%Full pipeline visibility, textile productivity increased +50%
• 打通“设计-备料-接单-排产-预警”Agent协同雏形Constructed SCM routing workflow (Design-Material-Order-Alerts)

基于 Ollama 框架在本地服务器部署 Google Gemma 4 12B、DeepSeek-V4 等开源大模型，实现企业敏感数据"零出境"的 AI 推理能力。解决制造业与供应链场景中客户订单数据、财务发票等核心数据无法上传云端 API 的合规痛点，数据全程不出企业内网。 Deployed Google Gemma 4 12B and DeepSeek-V4 open-source models on local servers via Ollama, achieving zero data leakage for sensitive enterprise AI inference. Addresses compliance concerns around customer orders, financial invoices, and supply chain data that cannot be sent to cloud APIs.

• 企业数据 零出境，满足高新科技企业数据合规要求Enterprise data stays 100% on-premise, full regulatory compliance
• 推理延迟 <100ms，高频调用成本降至云端 API 的 1/10Inference latency <100ms, cost reduced to 1/10 of cloud APIs
• 支持与云端模型混合路由，敏感数据走本地、复杂推理走云端Hybrid routing: sensitive data runs locally, complex reasoning via cloud

针对跨境电商亚马逊 FBA 头程物流场景，基于领星 ERP 系统的出货单数据，利用 DeepSeek V4 PRO + Minimax M3 双模型驱动，自动解析每一票出货单的 SKU、数量、重量、目的仓等字段，智能匹配物流商报价与税率规则，一键生成符合海关申报格式的头程运输发票。将原本财务人员每票手工制作 15-20 分钟的发票工作压缩至 AI 秒级自动生成，人工仅需终审确认。 Built an AI agent for Amazon FBA first-mile logistics invoicing. Parses shipment orders from LingStar ERP (SKU, quantity, weight, destination warehouse), auto-matches freight quotes and tax rules via DeepSeek V4 PRO + Minimax M3 dual-model, and generates customs-compliant shipping invoices in seconds — replacing 15-20 min manual work per invoice.

• 对接领星 ERP 出货单，自动提取 SKU/数量/重量/目的仓等核心字段Integrates LingStar ERP shipments, auto-extracts SKU/qty/weight/warehouse fields
• 每票发票生成时间从 15-20分钟 → 秒级，效率提升 100xInvoice generation: 15-20 min → seconds, efficiency boost 100x
• DeepSeek V4 PRO 负责结构化解析，Minimax M3 负责多语种发票内容生成DeepSeek V4 PRO for structured parsing, Minimax M3 for multilingual invoice content
• 输出格式符合海关申报 + 物流商对账要求，人工仅需终审Output format meets customs declaration + logistics reconciliation standards

利用 Dify 开源平台搭建企业级多 Agent 协同工作流：将供应链各环节拆解为"调研 Agent → 分析 Agent → 审核 Agent → 执行 Agent"四阶段链路。每个 Agent 使用独立 LLM 策略（ReAct / Function Calling），通过可视化编排实现跨部门业务流程自动化，并集成 Langfuse 实现全链路可观测性。 Leveraged Dify open-source platform to build enterprise multi-agent collaborative workflows: decomposed SCM operations into "Research Agent → Analysis Agent → Review Agent → Execution Agent" pipelines. Each agent uses independent LLM strategies (ReAct / Function Calling), with Langfuse observability integration for full-chain traceability.

• 4 阶段 Agent Pipeline，可视化编排替代硬编码逻辑4-stage Agent Pipeline with visual orchestration replacing hard-coded logic
• 集成 Langfuse 全链路追踪，Agent 推理过程 100% 可审计Langfuse full-chain tracing, Agent reasoning 100% auditable
• 支持 Human-in-the-Loop 人工审批节点，满足企业合规需求Human-in-the-Loop approval nodes for enterprise compliance

将 Claude Code (Anthropic 出品的 CLI 智能体) 作为日常核心开发工具，用于自研项目的全栈代码生成、调试与重构。通过配置 Custom Skills 指令集和 CLAUDE.md 项目规范，将 Claude Code 训练为"理解我业务上下文"的专属编程 Agent，实现从需求描述到可运行代码的端到端交付。 Adopted Claude Code (Anthropic's CLI agent) as the core daily development tool for full-stack code generation, debugging, and refactoring. Configured Custom Skills instructions and CLAUDE.md project specifications to train Claude Code as a business-context-aware programming agent, enabling end-to-end delivery from requirements to running code.

• 个人编码效率提升 3-5x，可独立交付完整前后端项目Personal coding efficiency boosted 3-5x, full-stack delivery capability
• 自定义 Skills 指令集 + CLAUDE.md 规范，Agent 理解业务上下文Custom Skills + CLAUDE.md specs, Agent understands business context
• 本网站即由 Claude Code 辅助从 0 到 1 完成开发This very website was built from scratch with Claude Code assistance

自研一套"本地模型优先、云端模型兜底"的智能路由系统。敏感数据（如发票、客户订单、员工信息）强制走本地 Gemma/DeepSeek 推理；复杂多轮推理、创意写作等高难度任务自动升级至 Claude/GPT 云端 API。路由决策基于任务复杂度评分 + 数据敏感度分级，实现成本与能力的最优平衡。 Built an intelligent "local-first, cloud-fallback" LLM routing system. Sensitive data (invoices, customer orders, employee records) is forced through local Gemma/DeepSeek inference; complex multi-turn reasoning auto-escalates to Claude/GPT cloud APIs. Routing decisions are based on task complexity scoring + data sensitivity classification for optimal cost-capability balance.

• 敏感数据 100% 本地推理，合规零风险Sensitive data 100% local inference, zero compliance risk
• 云端 API 调用量削减 60-80%，Token 费用大幅下降Cloud API calls reduced 60-80%, massive token cost savings
• 任务复杂度自动评分 + 数据敏感度三级分类引擎Auto task-complexity scoring + 3-tier data sensitivity classifier

基于字节跳动「扣子 Coze」平台，搭建面向跨境电商运营的多智能体集群。包含竞品监控 Agent（自动追踪亚马逊竞品价格与评论变化）、选品推荐 Agent（基于销售数据与市场趋势的智能选品）、客服话术生成 Agent（多语种买家消息自动回复草稿），三个智能体通过 Coze 工作流串联协作，形成"监控→分析→执行"的自动化业务闭环。 Built a multi-agent cluster on ByteDance's Coze platform for cross-border e-commerce operations. Includes Competitor Monitor Agent (auto-tracks Amazon competitor pricing & reviews), Product Selection Agent (intelligent product recommendation based on sales data), and Customer Service Agent (multilingual auto-reply drafts). Three agents are chained via Coze workflows into a "Monitor → Analyze → Execute" automated loop.

• 3 个专业智能体协同：竞品监控 / 选品推荐 / 客服话术3 specialized agents: competitor monitor / product selection / customer service
• 竞品价格变动 实时告警，人工巡检时间缩减 90%Competitor price alerts in real-time, manual monitoring time cut 90%
• 多语种客服草稿生成，覆盖英语 / 日语 / 东南亚小语种Multilingual CS drafts: English, Japanese, Southeast Asian languages