An AI company in Korea: we build AI into enterprise systems, ship open-source tools, and publish research.
WIGTN works in three teams. Solutions takes AI transformation (AX) for enterprises from consulting to the systems that run it. Product ships the open-source tools we build and use ourselves. Research publishes peer-reviewed work on LLMs, their evaluation, and speech. The repositories below are the code behind that work.
| Project | What it does | Built with |
|---|---|---|
| WigtnOCR | Document parser distilled from a 30B into a 2B vision-language model. The teacher writes Markdown for 4,501 pages, the judge keeps the pages that score 3 or higher, and the student is LoRA fine-tuned on 2,667 of them. WigtnOCR-2B matches or beats its teacher on 3 of 5 OmniDocBench metrics, with Table TEDS 0.649 against 0.523 | Qwen3-VL-30B-A3B-Thinking (teacher), Qwen3.5-122B-A10B (judge), Qwen3-VL-2B-Instruct (student), ms-swift |
| Project | What it does | Venue |
|---|---|---|
| WIGVO | Real-time Korean–English speech translation over ordinary phone calls, built on the OpenAI Realtime API and Twilio Media Streams. 555 ms median caller-to-callee latency at USD 0.28/min, 430+ tests | ACL 2026 System Demonstrations |
| Project | What it does | Venue |
|---|---|---|
| RCPS | Chooses document parsers by retrieval on a fixed question–answer probe, and separates parser-side from chunker-side loss. It evaluates WigtnOCR-2B as its production parser | EMNLP 2026 Industry Track |
| MDI | The minimum detectable improvement of an LLM judge. All 97,752 judge calls are logged, and every number in the paper resolves to a file | NeurIPS 2026 TAE Workshop |
| Project | What it does | Built with |
|---|---|---|
| WIGTN Plugins | Claude Code plugin with 11 agents, from requirements to a verified commit | Claude Code |
| WIGTN Plugins for Codex | The same requirements, verification and release workflow for Codex | Codex |
| WIGSS | Drag a component on your running dev server and an AI agent rewrites the source. npx wigss |
Next.js, WebSocket, OpenAI API |
Other public projects are in our repositories.
Everything we learn, we share.