$ tree -L 1 ~/projects

Things built and shared.

In my own workshop, I make useful tools and apps. With the BC Card AX team, I build platforms that bring AI into everyday work.

01 / familyiOS · Android

app/ieum

ieum

How children and parents see each other

A family personality app where parents and children answer the same 12 questions and compare their perspectives through 16 animal types. It began with a wish to use technology to bring families closer.

platform
iOS · Android
flow
Role → age group → 12 questions → result
types
16 animals
Download links will be added when available.
ieum — Different perspectives within a family
Different perspectives within a family · concept illustration
02 / toolsopen source

gh/sh2orc/iyagi-term

IYAGI Terminal

A terminal for AI coding agents

An open-source terminal for running and coordinating multiple AI coding agents. It preserves Claude Code, Codex and OpenCode sessions so work can continue, with agent resources visible in one place.

lang
Rust · TypeScript
shell
Tauri 2 · xterm.js
license
Apache-2.0
terminal — A workspace for AI agents
A workspace for AI agents · concept illustration
03 / toolsmacOS

gh/sh2orc/iyagi-layout

IYAGI Layout

Split windows into halves, then quarters

Working with multiple agent terminals can make window management a task of its own. This macOS menu-bar app uses ⌥⌘ and arrow keys to split windows into halves, then quarters with another press.

lang
Rust · React
shell
Tauri 2
keys
⌥ ⌘ + ← → ↑ ↓
layout — Shortcuts for arranging your workspace
Shortcuts for arranging your workspace · concept illustration
04 / models & dataHugging Face

hf/sh2orc

Korean LLMs & datasets

Models and data that understand Korea

I adapt diverse LLMs to Korean language, knowledge, culture and usage contexts, across model families and sizes. I also create and release the datasets they need. Personal work is on sh2orc; the BC Card AX team’s MoAI models and datasets are on BCCard.

scope
Diverse LLM families and sizes
focus
Korean language, knowledge and contexts
builds
Korean-adapted models · datasets
Korean adaptation of LLMs and dataset creation · concept illustration

$ ls ~/work/bccard

AI built together.

These are the platforms and models I help design and develop as the BC Card AX team leader. They are team projects, presented separately from my personal work.

inference / BC Card

MoAI inference platform

With the AX team, I designed vLLM, FP8 and EAGLE-3 serving on OpenShift AI. SLM Swarm routes across models to serve MoAI LLMs in Paybooc AI.

SLM Swarm results: up to 70% lower GPU usage · more than 3× faster inference

Presented at Red Hat Summit 2026

agents / BC Card

Open MoAI · MoAI MATE

We built the Open MoAI web portal and MoAI MATE, extending computer use into multi-agent orchestration. MoAI MATE was expanded to 200 developers in September 2026.

39-developer POC: 56.4% productivity gain · 64% fewer bugs

POC results before the broader rollout

open models / BCCard

MoAI models and datasets

The AX team released MoAI-Privacy-Filter for Korean financial PII detection, MoAI-Embedding and EAGLE-3 draft models. I trained moai-gemma4-31b-mxfp4 using our in-house MXFP4 quantization method.

Team releases: 49+ models · 12 datasets

2026 figures · moai-gemma4-31b-mxfp4: 10th overall and 7th on K-MMLU in the K-Leaderboard, August 2026

The BC Card team’s public models and datasets ↗

More experiments on GitHub ↗