AI Engineer · Building Production-Grade RAG, Agents & Data Systems
RAG · AI Agents · Tool Calling · Data Pipelines · Streaming
"AI that doesn't just demo — it deploys, measures, and ships."
I design and build end-to-end AI systems that turn raw documents into grounded, cited answers — and user requests into tools that actually get work done. Not toy demos: production-grade, deployed, evaluated, and measured.
| Pillar | What I Build | Why It Matters |
|---|---|---|
| RAG | Enterprise knowledge bases, document Q&A with citations | Every company must answer questions from its own data |
| AI Agents | Agents that call tools, plan, and finish tasks | The shift from chatbots to agents that do work |
| Data Platform | ETL, real-time streaming, data quality | AI is only as good as the data behind it |
flowchart LR
U[User Query] --> P[Pre-Processing]
P --> V[(Vector DB)]
P --> B[(BM25 Index)]
V --> H[Hybrid Fusion 0.6/0.4]
B --> H
H --> R[Cross-Encoder Rerank Top-20 to Top-5]
R --> G[Guardrail Pre + Post]
G --> L[LLM]
L --> C[Grounded Answer + Citations]
G -.reject.-> E[Safe Fallback]
flowchart LR
A[User] --> O[Orchestrator]
O --> R1[Research Agent]
O --> W[Writer Agent]
O --> C[Critic Agent]
C -->|feedback| R1
W --> O2[Tool Calling: calc / search / KB / DB]
O2 --> T[Structured Trace + Pydantic Validation]
Hybrid retrieval + reranking + guardrails + eval + multi-tenant. Production-grade RAG backend across 4 real business scenarios (support, wiki, product docs, ticket resolution). 2,700+ lines, 23 files, Docker-deployable.
Why it's hard (and how I solved it):
- Hybrid retrieval 0.6/0.4 — pure semantic search misses exact terms; I fused vector + BM25.
- Rerank Top-20 → Top-5 — Cross-Encoder to cut noise before the LLM.
- Hallucination guardrails — pre-check (is context relevant?) + post-check (is answer grounded?) to kill hallucinations.
- Eval harness — Context Precision / Faithfulness / Answer Relevancy; every param change re-validated.
- Multi-tenant isolation — namespace + row-level; the one bug that leaked data across tenants.
ReAct loop + tool calling + memory + error handling. An agent that calls tools (calculator, web search, KB, DB) with a structured trace and self-correction via Pydantic validation.
Orchestrator + specialist agents (Research / Writer / Critic). A content pipeline that shows agent collaboration, not just single loops.
Extract → Transform → Validate → Load. Clean-null, normalize, add-timestamp, quality-gate with abort-on-fail, idempotent loading, full lineage.
Real-time window aggregations over WebSocket. Events in → process → query via API → live dashboard.
Production function-calling patterns (weather, order, calendar, email) with clean schema + multi-turn.
| Layer | Tools |
|---|---|
| AI / LLM | OpenAI · LangChain · RAG · ReAct Agents · Function Calling · Guardrail · Eval |
| Backend | FastAPI · Python · SQLAlchemy · JWT Auth |
| Data | ETL · PostgreSQL · SQLite · Redis · Celery · Real-time Streaming |
| Vector | Chroma (dev) · Pinecone (prod) · Hybrid + Cross-Encoder Rerank |
| Infra | Docker · docker-compose · Structured Logging · Metrics |
- Hallucination in RAG → built pre/post guardrails + eval harness; grounded answers with citations.
- Data leakage across tenants → namespace + row-level isolation; fixed the multi-tenant bug.
- Agent tool-call failures → Pydantic validation + self-correction on malformed args.
- ETL quality failures → abort-on-fail quality gates + idempotent loading + full lineage.
6 production projects · 2,700+ lines in flagship RAG backend · 4 real business scenarios · Hybrid + Rerank + Guardrail + Eval · Docker-deployable · Full-stack (API + frontend)
I'm actively looking for remote AI / full-stack roles (worldwide) and freelance AI projects.
GitHub: @LeonxLJX · Email: liuzhaoxing373@gmail.com Focus: RAG · AI Agents · Data Pipelines · Full-Stack AI