WIGTN (와 이게 되네 — "Wow, this actually works!") is a 5-person AI-native crew based in Korea. We build practical, domain-specialized AI tools and ship them as open-source — from fine-tuned VLMs to real-time voice translation systems.
Build with TRAE Seoul (ByteDance) — Grand Prize Winner. Our project Wigent won the Grand Prize at ByteDance's Build with TRAE @Seoul hackathon on the theme of AI Agent development.
Google Gemini Live Agent Challenge · Status: Live
AI-powered museum guide. Point your camera at an artifact and get real-time AI explanations with historical context and restoration visualizations. Built with Gemini Live API and Google ADK.
Build with TRAE Seoul Grand Prize · Status: Released
Multi-agent debate platform. Throw in a topic and watch a PM agent + auto-spawned expert agents discuss, argue, retire, and spawn new specialists in a Slack-style chat UI — then auto-generate a landing page from the conclusions. Built in 3.5 hours by 3 members using parallel Claude Code development.
Status: Active
Extract subtitles from YouTube captions or via WhisperX STT, auto-translate English to Korean with GPT-4o-mini, and synchronize in real-time. Built entirely with WIGTN Coding plugin — full microservices architecture (Web, API, AI) through AI-assisted development.
Status: Active · Published on npm
"Drag your UI components — the source code rewrites itself."
Point WIGSS at your running dev server, visually drag and rearrange UI components in the browser, and watch the source code rewrite itself with an always-on AI agent.
Status: Active
Mobile app for recording and managing overseas travel expenses. Receipt OCR powered by our own WigtnOCR model with real-time currency conversion and offline support. Monorepo + Microservices Architecture (Expo mobile, NestJS backend, Next.js admin).
Status: Active · v2.0.0
From idea to deploy, zero friction. A single, unified Claude Code plugin with 12 agents, 3 skills, and 17 design styles — all working together with team-based parallel execution for 3-5x speedup.
ACL 2026 System Demonstrations (submitted) · Status: Live
Real-time bidirectional voice translation over actual PSTN phone lines. No apps needed on the recipient's end — just call. Dual-Session Echo Gating architecture. 0 echo-loop incidents across 148 production calls. 430+ passing tests. Production-deployed on Google Cloud Run.
EMNLP 2026 Industry Track (targeting) · Status: Released
Pseudo-Label Distillation for Structure-Preserving Document Parsing. Distills Qwen3-VL-30B into a 2B student via quality-filtered pseudo-labeling and LoRA fine-tuning. Proves: better parsing leads to better chunks leads to better retrieval on Korean government documents.
| Benchmark | Score | Rank |
|---|---|---|
| OmniDocBench Table TEDS | 0.649 | #1 |
| KoGovDoc-Bench Retrieval Hit@1 | 0.739 | #1 (among 6 parsers) |
| Name | Role | GitHub |
|---|---|---|
| Hyeongseob Kim | Founder & Crew Lead | @Hyeongseob91 |
| Jinmo Kim | Solution Architect | @moriroKim |
| Hyeonsang Kim | AI Product Engineer | @HyeonsangKim |
| Sangwoo Son | AI Engineer | @wigtn |
| Hyunwoo Cho | AI Product Engineer | @starz-woo |
© 2025-2026 WIGTN. All rights reserved. Individual projects are licensed as indicated in their respective repositories.
This project is built and maintained by WIGTN Crew —
an AI-native open-source research crew based in Korea.
We build practical, domain-specialized AI tools. Fast prototyping, strong engineering, shipping real things.
| 🌐 Website | wigtn.com |
| 🐙 GitHub | github.com/wigtn |
| 🤗 HuggingFace | huggingface.co/Wigtn |
| 📦 NPM | npmjs.com/org/wigtn |
| Project | Description | Status |
|---|---|---|
| WigtnOCR | VLM-based Korean government document parser | Research |
| WIGVO | Real-time PSTN voice translation (Korean↔English) | Research |
| Claude Code Plugin | Multi-agent parallel execution ecosystem | Open Source |
Results speak louder than résumés.