Not training bigger models —
building better plumbing around them: collection, extraction, search, verification, guardrails.
AI pipelines that run without cloud API keys.
Backend engineer with 6 years of experience. Worked across 6 heterogeneous databases (PostgreSQL, Oracle, MariaDB, MSSQL, DB2, Netezza) with native SQL tuning. Built DW/DM, BI/OLAP, and analytics systems for public and enterprise clients. Extensive deployment experience in air-gapped environments.
Currently building local-inference AI pipelines — from data collection to refinement, RAG, and fact-checking — everything runs on-premise without cloud API calls. Tools extracted from this work are published as open source under QuartzUnit.
Each project is independent but connected by data flow. Raw data flows through refinement, mart, and agent layers.
Multi-source raw data to LLM-ready data warehouse. News, blogs, legacy DBs — any source refined. Includes fact-checking at 83.6% accuracy.
Auto-generates data marts on top of Forge DW. Natural language to SQL — BI dashboards in one sentence.
Custom ReAct + RAG (Qdrant + Neo4j) + persistent memory + 2-tier LLM fallback. On-premise, no frameworks.
10 Python packages extracted from the above projects. Collection, extraction, search, monitoring, guardrails — modular tool ecosystem.
All projects run on local inference (vLLM on-premise). No cloud API dependency.
Each tool solves one problem. CLI + async Python API + MCP server — three interfaces as standard.
URL to LLM-ready markdown. HTML, YouTube, PDF, DOCX.
ExtractSchema-based document OCR to structured JSON.
CollectRSS/Atom feed collection. 444 curated feed catalog built-in.
SearchLocal LLM-optimized browser agent. Accessibility tree + token savings.
MonitorURL to screenshot + metadata. Claude Vision optimized.
MonitorWeb page change detection + structured diff generation.
SearchLocal semantic search. Embedding-based grep — search by meaning.
GuardLLM output degeneration detection. 4-signal composite scoring.
GuardAgent infinite loop detection. Sliding window similarity.
GuardDeclarative action policy. 4 built-in templates, zero dep.
10 packages · 959 tests · MIT License · Korean + English docs