Systems running in production / live
A production ops platform orchestrating 6 department AI agents through a Telegram COO interface, blocking content at a medical-advertising-law gate before publish.
A 4-branch operations analytics platform that idempotently reconciles 17 months / 23K rows of revenue ledgers and unifies channel attribution, booking funnels, and no-show risk — with automated weekly WoW reporting.
An operational knowledge base that protects patient data with three-layer PII defense (OS sandbox + hooks + n8n gate) and flags violations in real time across 52 wiki pages and 12 medical-advertising regulations.
A 256-page mobile fortune-telling platform with Toss payments, a coin economy, and idempotent refund logic — grounded in a classical-text RAG corpus with 6 daily push dispatches.
MVP · in progress · sibling builds
A 14-package TypeScript monorepo that runs a crawl→analyze→workorder→recheck loop deterministically without LLM dependence, isolating AI into a replaceable ai-core.
A Layer2 security gate that independently re-scans PII regardless of source declarations — hardened by 3 adversarial red-team rounds (unicode, homoglyph, spread evasion) and never logging PII values.
A medical-tourism marketplace with 6-language i18n, AI smart quotes (Claude vision), a 3-tier hybrid translation pipeline, and a transparent pricing board. (Two iterative builds.)
A premium saju SaaS grounding OpenAI interpretations in classical citations (Jeokcheonsu, Jipyeongjinjeon), with a three-stream revenue model (membership, coins, collectible reports), Toss payments, and web push. A sibling build to 간지사주.
I run Korea's most heavily-regulated vertical — medical and beauty — on code. Eight marketing, operations, and data systems, designed and shipped end to end, by me, actually running in production.
My stance is firm: compliance can't ride on human attention — deterministic gates have to enforce it. Medical-advertising-law violations are blocked before publish, patient PII is quarantined behind layered defenses, and analytics stay reproducible on idempotent pipelines. Regulation isn't a variable I trade against marketing; it's a constraint the system can't break.
Every one of the 8 here was built, deployed, and measured — multi-agent orchestration, 17 months and 23K rows of revenue reconciled, Toss payments, classical-text RAG. Proven by working systems, not slides.