10+ 年 Java、TypeScript、AWS、数据平台和企业系统经验。现在聚焦 LLM 工作流、评测治理、审批安全、开发效率和可部署的 AI 产品。 10+ years across Java, TypeScript, AWS, data platforms, and enterprise systems. I now focus on LLM workflows, evaluation, governance, approval safety, developer productivity, and deployable AI products.
精选案例 Selected Work
把选品评分、供应商询价、内容草稿、人工审批、审计和本地私有部署串成安全的运营决策中心。 Connects product scoring, supplier quote priorities, content drafts, human approvals, audit evidence, and private local deployment.
把团队上下文变成 provider-neutral FastAPI + Angular 工程工作台：Memory Hub intake、Codebase Index、Requirement Case、PR Review、Context Trail、Audit/Eval、human ratings、DLP foundation、connector lifecycle 和 Pilot evidence 都保留来源、评测与审计边界。 Turns team context into a provider-neutral FastAPI + Angular engineering workbench: Memory Hub intake, Codebase Index, Requirement Cases, PR Review, Context Trail, Audit/Eval, human ratings, DLP foundation, connector lifecycle, and Pilot evidence with source, evaluation, and audit boundaries.
把发布风险扫描、仓库上下文、测试建议、公开 proof artifacts、DynamoDB 证据和人工审批门禁组织成可审计的工程 AI 工作流。 Organizes release-risk scans, repository context, test recommendations, public proof artifacts, DynamoDB evidence, and human approval gates into an auditable AI workflow.
面向 Qwen Cloud Track 4 的工程自动化代理：CI failure、coverage gate、incident triage、staged remediation、deterministic fallback、GitHub Actions proof 和人工确认门禁。 A Qwen Cloud Track 4 engineering agent for CI failures, coverage gates, and incident triage with staged remediation, deterministic fallback, GitHub Actions proof, and human checkpoints.
读取 JaCoCo 报告，定位低覆盖 Maven 类，生成 JUnit/Mockito 测试，运行校验并进入修复循环。 Reads JaCoCo reports, targets low-coverage Maven classes, generates tests, validates, and repairs failures.
把研究、生成、质量评分、人工审核、发布证据和部署配置做成可追踪的内容生产系统。 A traceable content workflow for research, generation, scoring, review, release evidence, and deployment.
能力证据 Capability Proof
Golden dataset、groundedness、引用正确性、成本/延迟追踪、运行历史和 CI 质量门禁。 Golden datasets, groundedness, citation correctness, cost/latency tracking, run history, and CI quality gates.
GitHubSchema 检索、SQL 安全校验、成本控制、本地 mock 执行和可追踪结果。 Schema retrieval, SQL safety checks, cost controls, local mock execution, and traceable results.
GitHubProvider-neutral FastAPI + Angular 工作台，覆盖 source intake lifecycle、Memory Claim Review、Codebase Index、Requirement Case、PR Review、Context Trail、Audit/Eval、human ratings、SQLite audit trail、DLP foundation、connector lifecycle 和 Pilot evidence。 Provider-neutral FastAPI + Angular workbench covering source intake lifecycle, Memory Claim Review, Codebase Index, Requirement Cases, PR Review, Context Trail, Audit/Eval, human ratings, SQLite audit trails, DLP foundation, connector lifecycle, and Pilot evidence.
GitHub发布风险评分、测试推荐、证据收集、提交材料、UiPath Test Manager proof、DynamoDB proof 和人工审批门禁；公开 demo 使用安全模拟状态，不调用客户系统。 Release-risk scoring, test recommendations, evidence collection, submission materials, UiPath Test Manager proof, DynamoDB proof, and human approval gates; the public demo uses safe simulated state and does not call customer systems.
GitHub Proof Index技术写作 Writing
A passing test is not an AI quality strategy
Why an AI-generated green build or coverage increase is only useful when it comes with evidence of behavior, review value, and quality movement.
A plugin is an integration contract
Why portable skills and MCP configuration matter only when they become versioned, governed workflow contracts.
Your AI rollout needs tiers, not one default assistant
Why this week's clearest signal was role-based AI access, budget tiers, tool allowlists, and review depth becoming policy surfaces.
The second build is where AI products get real
Why this week's strongest signal was that production AI work is shifting into telemetry, policy, evals, and operating structure.
One reviewable path beats a very smart agent
Why last week's tooling signal pushes me toward narrower, inspectable AI workflows instead of broader autonomy.
The repository is becoming the AI rollout unit
Why enterprise AI adoption is shifting from license counts toward repo-level readiness, controls, and evidence.
Longer agent runs need shorter feedback loops
Why stronger AI agents increase the value of tighter eval, approval, and review loops instead of removing them.
The AI workbench is becoming the product
Why this week's bigger signal was domain-specific AI systems packaging models, context, artifacts, and compute into one operating surface.
The next AI bottleneck is the team thread
Why the stronger signal this week was agents moving into shared queues, tickets, and channels rather than staying inside private chats.
AI agents need commit boundaries
Why the deeper signal this week was the move toward staged effects, failure lineage, and cost accountability.
Your agent needs operating documents
Why machine-readable instructions, capability catalogs, eval traces, and usage telemetry are becoming part of the real AI system.
A model name is not a stable architecture
Why upstream changes in model access, permissions, and spend controls now shape system behavior as much as model quality.
职业定位 Profile
Java, Spring Boot, TypeScript, Angular, React, Node, APIs, CI/CD, enterprise release workflows.
AWS S3, Lambda, ECS/Fargate, CloudWatch, DynamoDB, SQS/SNS, API Gateway, RDS, Athena-style workflows.
RAG, agents, CI remediation, test generation, evaluation, SQL generation, content automation, review gates, audit trails.
欢迎用中文或英文交流：企业 AI 落地、RAG / Agent / LLMOps、AWS 数据智能、开发者效率和业务系统 AI 化。 Happy to discuss enterprise AI delivery, RAG / agents / LLMOps, AWS data intelligence, developer productivity, and AI-enabled business systems.