Build AI that
understands Japan.

We cover formal and informal Japanese across kanji, hiragana, katakana and romaji, including honorific registers, regional variation, slang and industry jargon. Normalization rules for punctuation, emoji and full-width characters are agreed during kickoff so tokenization stays consistent.

Turnaround depends on volume, linguistic complexity and QA depth. We scope a pilot first, measure the real throughput on your data, and then commit to a delivery plan rather than a guess.

No fixed minimum. We support small pilots and scale to high-volume production programs; practical minimums depend on task complexity and tooling setup.

Yes. We work inside client-provided tools and platforms, or in our own secure labeling environment when you would rather not provision accounts.

JSONL, CSV, CoNLL-U and custom schemas defined at kickoff. Delivery includes the label taxonomy version and a change log for every batch.

Revisions run through the same QA pipeline as first-pass work. Change logs, QA reports and targeted re-labeling keep datasets consistent across versions instead of drifting between batches.

NDAs and data processing agreements, role-based access with least privilege, isolated working environments, and retention and deletion policies aligned to your requirements.

Yes — instruction tuning data, preference and ranking labels, safety and red-teaming annotation, evaluation sets, and human-in-the-loop review of model output in Japanese.

Yes — for projects that need it, we build teams around the required domain expertise, which can include people with relevant qualifications or practical experience in fields such as law, tax, accounting, labor and social security, or corporate registration. We do not keep a standing roster; availability is assessed per project against its scope, volume and timeline, and we confirm what is achievable before starting.

Yes. Expert annotators can score generated responses for factual accuracy, reasoning quality, professional appropriateness and compliance with your criteria — and can produce preference rankings for RLHF, benchmark sets, RAG evaluation and red-teaming work.

Guidelines are drafted at kickoff, stress-tested during the pilot on real edge cases, then locked as a versioned document. Every disagreement resolved in review is written back into that document.