This is the first post on the Suikou AI blog. It exists to verify that the MDX content pipeline renders end-to-end — frontmatter parsing, the table of contents, reading-time estimation, and the Article JSON-LD all run from this single file.
Why Suikou AI
Most AI humanizers were built for English. When you paste a Japanese academic
draft into them, you get back English-shaped Japanese: monotone sentence
endings, broken は / が alternation, kanji-ratio that screams machine. We
built Suikou AI because the morphology of Japanese gives humanizers a job that
LLMs trained on English-majority corpora keep botching — and because doing that
job well in Japanese is the same job in Korean and Traditional Chinese, where
the same gap exists.
How the pipeline works
- Three parallel rewrites at temperatures 0.6 / 0.8 / 1.0 via DeepSeek-Chat.
- AI-likelihood scoring of every candidate with Qwen-72B (via OpenRouter) under a Japanese-morphology-aware rubric.
- The lowest-AI-score candidate wins. Pro users can opt in to a
second-pass academic polish via claude-3.5-haiku that preserves
[CITE:1]style citation markers.
What's next
This blog will document the engineering tradeoffs (why DeepSeek not GPT-4, why Qwen-72B for the detector, how to detect citation drift in a polish pass) and the operating tradeoffs (one-person dev, hidden HK entity, why we don't do team plans yet). If you found Suikou AI useful, the best thing you can do is tell one other graduate student.
— Ryota