AI Engineer · Building Customer Support AI Agents end-to-end RAG · LangGraph · Text-to-SQL · Production Multi-Agent Systems
Building Customer Support AI Agent in production @ Mono ──────
Sole owner:
├ Architecture · 7-class router · multi-agent orchestration
├ Slot-filling agent (LangGraph subgraph, 16 nodes)
├ Text-to-SQL pipeline w/ Self-Correction (~3,840 LOC, Claude Sonnet)
├ Web labeling console (self-evolving QA loop, 914 LOC)
├ Sales subgraph (Neo4j + Opus 4.7 advisor mode)
├ Mailing system (FastAPI, 30 endpoints, built solo in 3 days)
└ KT/Communis crawling pipeline (4~7h → 2min, 50x+ speedup)
Co-designed with research engineer:
└ Document RAG subgraph — current production code runs on the
infrastructure I built (engine, state, prompts, LLM hub)
Stack: LangGraph · RAG (Dense-X, Recall@5 +2x · MRR +2.78x)
Text-to-SQL · pgvector · ChromaDB · BGE-M3 reranker
Spring Boot · React · PostgreSQL · Anthropic Claude API
- RAG retrieval — Recall@5 +2x, MRR +2.78x, zero hallucinations across 5,636 propositions (Dense-X EMNLP 2024 reproduction)
- KT/Communis crawling — 4~7h → 2min (50x+ speedup, RCS 200x via JWT+RSA reverse engineering)
- Mailing system — 5 flows · 30 endpoints · ~4,510 LOC Python · 3 days end-to-end
- CMS data curation — 46,748 inquiries (9 years) → 2,674 meta patterns · workload 230h → 34h (-87%)
- n8n infrastructure — 12 workflows · 42 days uptime
LangGraph · LangChain · RAG · Text-to-SQL · Anthropic Claude API · OpenAI API · pgvector · ChromaDB · BGE-M3 reranker · Self-Correction · Multi-Agent · Spring Boot · React · PostgreSQL · Neo4j · FastAPI · Python · Java 21 · Docker · Linux