Hun Labs

Gihun Lee이기훈

Mechanical engineer (PhD) building tools with AI coding agents to fix the frictions of my own day.

I studied mechanics and numerical analysis, and working as a researcher left me with one habit: no data, no analysis. Since the summer of 2025 I have been working with AI coding agents, and my repositories have reached twenty-six.

I decide what to build and why, and I check what is right. AI coding agents write most of the code and documents.

See the LabsRead the story

How I work

  • Rules in documents, prohibitions in machines — irreversible actions are blocked mechanically before they run.
  • A verification level on every claim — documented, built, device responded, seen, matched for accuracy, written apart.
  • Predictions first — written before the result; wrong judgments get corrections, not erasure.
  • Audits through different eyes — an audit agent with no write access and AIs from other companies review separately.
  • Records outside the conversation — decisions, grounds and the next step live in files, so work survives a new session.
  • Hunting for holes — having safeguards is not the same as their working, so I keep finding and fixing the gaps.

Research

A paper on structured prompting and LLM behavioral consistency — arXiv:2607.24765 (preprint, not yet peer reviewed). The reproduction package was released first, and four patent applications that came out of this research are under examination.

The hobby I put the most into — TTAgentRing

Run several agents and you have to flip through terminal tabs to find the session waiting for your answer. This round touch device on my desk is how I set out to fix that daily annoyance.

The ring is split into one segment per session, showing each name and state, and a session waiting for me changes its segment's color. Tapping a segment takes me to that session's window, and I pick options or approve actions right on its screen. The same question appears on my PC screen and on the device at once; whichever I answer first goes in, and the other side withdraws itself. It shows coding agents from several companies on the same ring, and I use it on both Windows and Linux.

The ways of working I built up across my repositories — mechanical safeguards against irreversible actions, verification levels checked on real hardware, writing predictions first, cross-audits by several models — all came together in this one thing.

Hobby project · in development. The photo is a concept image based on the screen in development.

TTAgentRing concept image — a ring split into four sessions