About
I am an engineer with a background in both computer science and finance. I use AI as leverage: I design the systems, drive them to done, and move fast by pairing with coding agents across parallel worktrees.
My path has been a series of deliberate switches. I qualified as a CPA and started in financial statement audit at Woori and Bakertilly, then moved into IT audit at EY, reviewing the systems behind Korea's banks and exchanges. Blockchain pulled me in next, so I studied it at KAIST, and went to Texas A&M for a master's in computer science, so I read a system as both an engineer and an auditor.
Research
Bitcoin After Block Rewards
Preprint · arXiv:2606.05503 (cs.CR, cs.GT)
Models miner incentives in a fee-only regime to derive the deviation threshold for honest mining, and shows that combining a base fee, a fee floor, and an adaptive block-size rule raises it.
Projects
Things I started and carried to the end. For each: what it is, why I started it, how the work split between me and AI, and the one thing I took away.
Han, a Korean programming language2026 · 184★
What A general-purpose programming language whose keywords, types, builtins, and error messages are all Korean, compiled to native code through LLVM.
Why Hangul is a designed writing system, so I wanted to see what code looks like when the keywords, types, builtins, and even error messages are all Korean. And I wanted to ship a real LLVM-backed compiler, not a toy.
Me I designed the language: the syntax, the type system, the Korean keyword and error-message decisions, and the overall architecture. I set the direction and made the calls on what to build.
AI AI agents wrote most of the implementation under that design: the compiler pipeline, the LLVM codegen, the test and benchmark harness, the LSP, and the tooling.
Learned How a language is built end to end: lexer, parser, AST, type checking, and the difference between interpreting and compiling to LLVM IR.
EveryMorning2026 · live
What A free service that emails me the three most impactful new STEM papers every morning by 7 AM, at stemem.info.
Why I wanted to keep up with new STEM research without drowning in arXiv, so I built a service that finds the most impactful papers each day, summarizes them, and emails me the top three by 7 AM. It runs for free at stemem.info.
Me I owned the product: what gets surfaced, the constraints (free tier only, send nothing rather than filler), and every operational piece, from the domain and hosting to the database, email, and secrets.
AI AI orchestrators built the implementation under my direction: the pipeline, the scoring, the web app, the unsubscribe flow. Their plans are committed into the repo as an audit trail.
Learned How to stand up a real service on zero-cost infrastructure: the Semantic Scholar API, Vercel, Supabase, and querying the database.
destiny2026 · 65★
What A Claude Code plugin for a daily fortune reading: the saju and I Ching numbers are computed locally, and only the interpretation is generated.
Why A fun Claude Code plugin for a daily fortune reading, built on one principle: the numbers are computed, only the interpretation is generative. The saju and I Ching pillars are calculated locally, and the model just turns fixed numbers into prose.
Me I set the framing and the product taste (deterministic computation with generative interpretation, English by default, prose over data rows), triaged the Hacker News feedback into the next features, and published it.
AI AI wrote essentially all of the code and docs under that direction: the saju and I Ching engine, the skill spec, the README.
Learned How to target user demand and get fast feedback, even from a toy project.
deeper2026 · shipped
What A depth-first interview agent that drills a single claim down to its bedrock first principle instead of branching out.
Why I wanted an interview agent that drills one claim down to its bedrock instead of fanning out, the inverse of the usual breadth-first brainstorming tools.
Me I set the concept and architecture: inverting a breadth-keeping interviewer into a depth-keeper, synthesizing four existing patterns, and designing the self-improving loop.
AI AI agents built the engine under that design: the depth-first search and judge logic, the Claude Code dynamic workflow, the tests, and the docs.
Learned How to build a self-improving agent loop: parallel subagents, a skeptical reviewer, Claude Code dynamic workflows, and calling tools effectively.
readhn2026 · 4★
What An AI-native Hacker News MCP server that surfaces practitioner knowledge through eigenvector-based trust ranking.
Why The real engineering knowledge on Hacker News is buried in practitioner comments. I wanted a tool that surfaces it through the user's lens and explains why each result matters.
Me I set the direction: rebuild around three primitives, make trust the central signal, pick the seed experts, and ship it to PyPI and the MCP registry.
AI An AI agent harness built essentially the whole codebase in one session, around 7,000 lines with full test coverage, under that direction.
Learned How to find real signal in noise: the Hacker News API, and eigenvector-based trust ranking across a network.
codekit2026 · on npm
What An AI-native toolkit that pairs an evidence-based coding constitution with a fully local security scanner for AI-written code.
Why AI agents write better, safer code when you give them a quality bar and a security floor. codekit bundles an evidence-based coding constitution (ai-native) with a fully local security scanner (codesure).
Me I owned the product strategy, the monorepo structure, the npm releases, and all the docs, including toning down the benchmark claims.
AI AI agents wrote most of the net-new implementation: the scanner engine, the rule sets, and the test and benchmark suites.
Learned How to make a codebase readable and safe for AI agents: quality review, security scanning, and writing for LLM readability.
cleanup2026 · tool
What A Claude Code tool that passively tracks which skills and tools I actually use and recommends removing the ones gone quiet, never deleting on its own.
Why My Claude Code sessions got slow carrying around 60 unused skills and tools. cleanup passively logs what I actually call and nudges me to remove what has gone quiet, but never deletes on its own.
Me I brought the problem and the product judgment: the safety design (recommend, never auto-delete) and the README voice and iteration.
AI AI wrote nearly all of the shell tooling and docs under that direction.
Learned How a skill's initial context loads, and how to keep SKILL.md and CLAUDE.md lean.
future-history2026 · live
What A data-anchored forecast of humanity from 2026 to 2200 across three internally consistent branches, with code-generated era art.
Why A data-anchored forecast of humanity from 2026 to 2200 across three internally consistent branches, with code-generated era art. The thesis: capability is near certain across all branches; the open question is distribution.
Me I set the concept (three branches, six eras, confidence tags), pushed it toward plain language, and made the call to strip the multilingual system back to English.
AI More than a dozen AI agents generated the content and art: each branch arc, each era illustration, and the per-era syntheses.
Learned How iterative agentic communication works, with many agents passing work back and forth to converge.
museum-as-code2026 · live
What The National Museum of Korea rebuilt as code, where every artifact record is real, executable Han source.
Why Rebuild the National Museum of Korea as code, where every artifact record is real source in Han. The cultural archive and the language demo become the same executable artifact.
Me I owned the curation and the strict rule of real Han syntax only, no fake pseudocode, and gated every feature through review and merge.
AI AI agents generated the artifact entries, pages, and QA across parallel worktrees under that mandate.
Learned Putting Han to real use: implementing an entire archive in the language I built.
Currently building
- computermoney: a Bitcoin-native wallet for AI agents, on Taproot (Rust).
Experience
Ernst & Young, Senior Consultant, Tech Risk (Financial Services)
Seoul, South Korea
Led IT-audit engagements for the Korea Exchange and major banks; ran K-SOX advisory and IT-control remodeling for financial institutions and cross-border entities.
Woori & Bakertilly, Staff Accountant, Assurance
Seoul, South Korea
Audited financial statements under K-GAAP and valued startups for fund investments.