# WIGTN

> An independent AI team that builds web products and AI systems for clients,
> and publishes what it learns in the open: peer-reviewed work, released
> models, and reports that say what they do not show. Korea-based. Services:
> web development (Web Agency) and AI transformation (AX Agency).

Every figure below comes from a published paper, a public model card, or a
technical report at https://tech.wigtn.com/. Work that has
not been published is not described here, including work under review.

## Projects

- [WIGVO](https://aclanthology.org/2026.acl-demo.33/): real-time bidirectional
  translation over ordinary phone calls, with no app on the recipient's side.
  Accepted to ACL 2026 System Demonstrations.
- [WigtnOCR](https://huggingface.co/Wigtn/Qwen3-VL-2B-WigtnOCR): a 2B LoRA
  adapter that parses Korean government documents into structured Markdown.
- [WIGSS](https://npmjs.com/package/wigss): browser-side visual editing that
  rewrites the source file behind the component.
- [WIGTN plugins](https://github.com/wigtn/wigtn-plugins): a Claude Code plugin
  ecosystem for multi-agent software workflows.

## Research details

### WIGVO, real-time speech translation over PSTN

- Problem: a telephone call gives you narrowband G.711 audio at 8 kHz, variable
  delay, and no control over the recipient's device, so translated speech played
  to the line returns and is recognized again. Eight of ten ungated test calls
  looped until interrupted by hand.
- Architecture: two directional sessions, one per speaker, with a three-stage
  filter between them. The echo gate replaces returning audio with valid mu-law
  silence rather than dropping packets, because dropping them reads to the far
  end as a dead stream.
- Reported: 555 ms median caller-to-callee latency over 814 turns; 2,684 ms in
  the other direction over 744 turns, where speech recognition dominates; zero
  echo-induced loops across 147 completed calls; USD 0.28 per minute on the
  evaluated provider stack. From 155 Korean-English calls over a live network.
- Stated limits: the field study is Korean-English over PSTN, not a broad
  language matrix; the callee-to-caller P95 is not yet acceptable for every
  conversation; evaluation used offline LLM references rather than human
  translations; cost reflects one provider configuration and pricing period.
- Venue: ACL 2026 System Demonstrations, accepted, pp. 336-344. An invited oral
  talk and poster followed at IWSLT 2026.
- Report: https://tech.wigtn.com/wigvo/

### WigtnOCR, a 2B document parser

- Method: a 30B teacher generated structured Markdown for 4,501 pages offline,
  a text-only judge scored and filtered it, and a 2B student was LoRA fine-tuned
  on what survived. 2,667 training pages, 294 held out.
- Reported on OmniDocBench: Table TEDS 0.649 against the teacher's 0.523. The
  teacher still leads formula CDM F1.
- Reported on KoGovDoc-Bench: Hit@1 0.739, Hit@5 0.855, MRR@10 0.788, first of
  six parsers on those three. It loses nDCG@10 to the untuned 2B base.
- Stated limits: the released artifact is a LoRA adapter, not merged weights;
  the student retains a skip rate; no controlled cost or throughput study has
  been published.
- Open: adapter, evaluation set and code, Apache 2.0.
- Report: https://tech.wigtn.com/wigtnocr/

### WIGTN plugins, multi-agent software workflows

- The published agent, skill and design-style counts describe the package
  surface. They are not a productivity result and are labelled as such.
- A controlled evaluation on two SWE-bench Verified tasks found GPT-5.6 Sol
  resolved 4/4 runs in each condition, with and without the legacy harness,
  while the harness increased median wall time by 151.7%, output tokens by
  141.2% and command count by 32.0%. The system was redesigned around selective,
  task-dependent intervention because of that result.
- Report: https://tech.wigtn.com/codex-selective-harness/

### WIGSS, visual editing that rewrites source

- Detects which of five CSS strategies a project uses and writes the change into
  that one: Tailwind utilities, CSS Modules, plain CSS or SCSS, HTML with a
  linked stylesheet, or inline React styles as a fallback.
- Every save produces a rollback token and compares the resulting DOM against
  the edited one within a two-pixel tolerance. A mismatch is surfaced as a
  warning rather than reported as a successful refactor.
- No controlled speed or code-quality benchmark has been published.
- Report: https://tech.wigtn.com/wigss/

## Track record

- ACL 2026 System Demonstrations: WIGVO accepted and demonstrated at booth D3.
- IWSLT 2026: invited oral talk and poster.
- Build with TRAE Seoul (ByteDance): Grand Prize.
- Snowflake AI & Data Hackathon Korea 2026: 2nd place, Tech Track.
- OBA Weekendthon: Top 6.

## Links

- [Website](https://wigtn.com)
- [Notice](https://wigtn.com/notices): the release record, every version dated
  from the registry that serves it
- [Story](https://wigtn.com/story): one row per event, each opening the full
  story with its photographs
- [WIG-log](https://tech.wigtn.com/): technical reports
- [GitHub](https://github.com/wigtn)
- [HuggingFace](https://huggingface.co/Wigtn)
- [npm](https://www.npmjs.com/org/wigtn)