Software architect based in South Korea — building developer tools and MCP servers for AI coding agents.

I spend most of my time on product code, developer tools, and small systems around AI agents, market data, and frontend workflows. Lately that's mostly meant Model Context Protocol servers: three of the four projects below hand an agent structured, verifiable context instead of leaving it to guess.

• Tooling for AI agents — MCP servers and companion skills that turn a messy domain — a design system, a live market feed, a git diff — into structured context an agent can act on. The whole protocol surface: tools, resources, prompts, and the docs the agent itself reads.
• Real-time data systems — long-lived WebSocket connections kept warm across multiple upstreams, ring buffers, cache layers with TTLs, bounded concurrency, REST fallback. Every response carries source and age_ms, so freshness is measured, not assumed.
• Analysis engines — an extensible indicator registry (73 and counting) with a Pine-syntax parser; signal evaluation run as a strictly causal event study, no look-ahead, no repaint; diff analysis across ~25 build ecosystems that emits SARIF for CI.
• Design systems as data — 48 styles and 57 UI-state recipes compiled into role-based tokens (JSON, CSS variables, Tailwind, TypeScript) with WCAG contrast checks, plus the production React site they live on.
• Release engineering — npm and PyPI packages, a GitHub Action, versioned JSON Schemas, CI with lint/type/test gates, READMEs in four languages. A tool counts as finished when someone else can install it in one command and check what it did.

An MCP server that hands the agent scored design contracts — visual directions, UI-state plans, and WCAG-checked design tokens — so it builds from evidence instead of defaulting to hero + three cards. Deterministic, read-only, fully local, no API key, and it ships with a companion skill.

An MCP server for live crypto market data, 73 indicators, and chart-structure recognition — no API keys. Built so agents can ask for market context without scraping around for it.

Reads a git diff and points out the verification that should exist before a patch is merged — no model call, no network, the same answer every time. Emits a portable Proof Pack (Markdown, JSON, SARIF, HTML) that a human, a CI gate, or a model can all inspect.

48 style references, side-by-side comparison, prompt workflows, and a motion lab — the curated catalog that web-stylebook-mcp draws from.

Most of what I build starts from a concrete need — something missing in my own workflow or an agent's. If the need is real, the tool gets built, finished, and documented so someone else can pick it up without me.

I like software that is easy to inspect: clear inputs, clear outputs, tests that mean something, and docs that stay close to the code. I care more about whether a system can be checked and maintained than whether it looks clever at first glance.

TypeScript and Python for most of the work, Rust when a hot path or a native tool needs it, and whatever else the project actually calls for.