memo-hub — Shared Memory for AI Agents
Independent Project · 2026.07 — Present
A shared knowledge and long-term memory service for AI assistants, helping multi-project teams retain reusable context, decisions, and lessons across local and NAS deployments.
- Established project isolation, candidate-to-formal memory, and knowledge versioning so retained information stays traceable and maintainable
- Connected MCP, management APIs, and CLI tools so AI assistants can retrieve authorized context and retain useful lessons
- Added cross-instance migration and privacy-first usage insights without retaining task content
PythonFastAPIMCPSQLiteChromaDocker
Enterprise Knowledge Assistant & Document Platform
Independent Project · 2026.06 — Present
An end-to-end internal knowledge assistant: converts PDF and PPTX documents into structured knowledge, then delivers cited answers through web, Feishu, and WeCom.
- Delivered multi-knowledge-base isolation, citation previews, user feedback, and quality dashboards for both answer trust and operational visibility
- Built a Windows-ready document conversion and sync tool that lowers the barrier for business teams to onboard and maintain knowledge
- Passed 16 of 17 internal evaluation scenarios (94.1%) with an average response time of about 3.4 seconds
FastAPINext.jsRAGSQLiteChromaFeishuWeCom
AI-Assisted Software Delivery Workflow
Independent Project · 2026.07 — Present
A structured workflow for AI-assisted software delivery, spanning requirements, design, testing, review, and release while keeping quality and accountability clear.
- Connected requirements, design approval, testing, review, and delivery evidence into reusable workflows that reduce rework and missed handoffs
- Made authority, version state, and acceptance criteria explicit at each step for more controlled and traceable AI-assisted development
- Created six reusable workflows covering the full path from idea intake to delivery
GitHubCodexSDDTDDGitHub Actions
Trivy Triage — Vulnerability Diagnosis & Remediation Platform
PEAK3 · 2026.05 — Present
Vulnerability triage and remediation platform spanning security, SRE, DevOps, R&D, and delivery teams. Automates the full Trivy scan → triage → Jira ticket → GitLab MR → rebuild → verify → release pipeline. Covers the full chain: base images (DevOps), middleware, common components (backend), business backend, and business frontend — with risk remediation for both backend and frontend. Auto-classification by component ownership, batch approval, merged ticketing, and FCFS job scheduling with cancellation, retry, and recovery.
- Per-batch vulnerability analysis + project scan in ~10 minutes, freeing security engineers from manual triage
- Zero-touch Jira ticket + GitLab MR generation
- Full-chain multi-dimensional vulnerability coverage forming a traceable, measurable governance loop
Node.jsExpressReact 18MySQLClaude CodeJira APIGitLab APINacos
Graphene — Insurance Core System Frontend
PEAK3 · 2020.12 — Present
Core product frontend serving insurance clients across Japan, SEA, and EUR markets, covering 15+ business domains. Led 5-year architecture transformation from a Web + Node hybrid monolith (~1 hour build) to a pnpm workspace + Qiankun micro-frontend architecture (10–15 min build) with 19 independently developed and deployed sub-packages, supporting 30+ parallel developers at peak. Core modules: Product Factory (AntV G6/X6 graph editing + formula engine), Rules Engine (Monaco Editor + AntV X6 decision trees), Claims Workflow (AntV X6 canvas + Formily v2 dynamic forms + Inversify DI). Full i18n across 5 languages with CLI-managed translation keys.
- Migrated all 19 sub-packages to Rspack: single-module build 1–2 min, full build 10–15 min (down from ~1 hour)
- Built reusable flow canvas component based on AntV G6/X6, adopted across claims, product structure, and BI DAG scenarios
- Micro-frontend architecture enables independent iteration and canary releases per business line, reducing cross-team coupling
- Visual graph editing replaced traditional form-based configuration, multiplying delivery efficiency for complex insurance products
React 18TypeScriptQiankunRspackAnt Design 5.xAntV G6/X6Formily v2Redux ToolkitSWRJotaiInversify DIpnpm workspaceTailwind CSSJestPlaywright
Holmes — Intelligent Diagnostic Tool
PEAK3 · 2025.12 — 2026.03
One-click intelligent analysis covering all 7 Graphene business domains (Marketing, Policy, Endorsement, Underwriting, Claims, Finance, Portal) end-to-end across frontend and backend. Submit a Trace ID or CSV and get root cause analysis with code-aware fix suggestions within 5 minutes. Multi-turn conversational follow-up with one-click knowledge base (Dify) accumulation for similar case matching. Chose Claude Agent SDK — analytical tasks need no local writes, making SDK the lightweight and purpose-fit choice.
- 200+ daily analysis tasks across all 7 Graphene domains, full-stack coverage
- ~5 minutes from submission to conclusion — platformized expert-level log diagnosis into self-service
- Empowered non-technical roles (BA, PM, pre-sales) to directly access root cause analysis, reducing cross-team communication overhead
PythonClaude Agent SDKSolidJSTypeScript
Code Review Agent — Automated Review Platform
PEAK3 · 2025.06 — 2025.09
Automated code review agent for R&D teams — AI review triggered within seconds of MR submission. Uses a dual-model attack/defense strategy: Claude (primary review via Claude Code CLI) + Gemini (independent cross-validation via Gemini CLI) for more thorough coverage. Chose CLI over SDK for this use case: must checkout MRs to disk, analyze diffs, and post results back to GitLab — CLI natively supports Tool Call and MCP for these local-operation-heavy workflows. Incremental and full review strategies based on MR ID vs commit ID diff detection, with per-repo customizable review rules (prompt-based, configured by repo owners via web UI to supplement global rules). Multi-channel delivery (Feishu / web dashboard) forming a measurable quality improvement loop.
- Drove adoption from pilot to company-wide mandate across all R&D teams, with influence extending to ZhongAn Group
- 12,000+ MRs reviewed across 418 repos, 8,000+ high-risk defects prevented (multi-tenancy, thread safety, DB risks, etc.)
- Reduced manual code review time by ~75%
- Won annual innovation award; became core of company's code quality assurance system
NestJSTypeScriptReactMongoDBGemini CLIClaude CodeDify
