AI-Native Application Engineer / Full-Stack Engineer · Shanghai, China

14 years of engineering experience — from frontend infrastructure to platform architecture, now focused on building AI-Native applications that actually ship to production.

🤖  Building with LLM / Agent / RAG / Tool Calling / MCP
🔍  Obsessed with AI observability, evaluation, and knowing when NOT to use an LLM
🛠  Deep in AI Coding — Claude Code / Codex, spec-driven development
✍️  Writing about opinionated technical decisions @ TigerBuilder

I care about the boundary between what the model understands, what rules constrain, and what the business system executes. Most of my recent work lives at that boundary:

• Agent engineering — intent routing, tool calling, multi-turn dialogue, fallback strategies, human-in-the-loop gates on irreversible actions
• RAG systems — hybrid retrieval, reranking, citation traceability, hallucination suppression
• AI observability — OpenTelemetry-based Session → Trace → Observation models, so you can answer "was that the data, the retrieval, the context, or the model?"
• Evaluation — golden sets for ambiguous phrasing, multi-intent, and edge cases; badcase reflow from production telemetry
• Platform engineering — turning messy, repeated work into reusable capabilities

知源 · RAGForge — Self-hostable, traceable RAG Q&A platform

An out-of-the-box knowledge base Q&A foundation. Bring your own model and corpus, get a debuggable, traceable, continuously improvable RAG app.

• Configurable Q&A apps — bind model + knowledge base + prompt + retrieval params per scenario. Immutable versioned configs with a single production pointer, so strategy never silently drifts.
• Knowledge base & chunking — ingestion, parsing, chunk store, search/inspect/bulk delete, citation provenance. Pluggable parsers and chunk templates.
• Retrieval playground — vector recall, keyword recall, reranking, citation generation, streaming answers — with a test bench to compare configs.
• Full-link replay — Session → Trace → Observation over OpenTelemetry, covering rewrite / intent / recall / rerank / generation. Trace any answer back to its execution path.
• Built end-to-end with AI Coding — competitive analysis → prototype → arch design → implementation plan → code → tests → docs.

TypeScript NestJS React + Ant Design PostgreSQL ClickHouse OpenTelemetry Docker

AI LLM Prompt Engineering Tool Calling Agent Loop RAG MCP Vercel AI SDK LangChain

AI Coding Claude Code Codex Spec-Driven Development

Languages TypeScript JavaScript Python Java C#

Frontend React Vue 2&3 Taro UniApp Mini-Programs

Backend & Infra Node.js NestJS FastAPI PostgreSQL ClickHouse Redis Kafka Elasticsearch Docker

Engineering Webpack Vite Rollup Monorepo CI/CD APM AST