What We Ship

Japanese document extraction API. Sign up and call it in a minute.

Japanese invoices, receipts, and forms go in; schema-validated JSON comes out — document type, page count, issuer and bill-to, subtotal, tax, total, and per-page line items carrying name, quantity, and unit price. Built for the layouts that defeat generic OCR: vertical text, hanko seals, and kanji mixed with full-width numerals. Fields the model is unsure of come back flagged in low_confidence_fields rather than silently guessed, so your reconciliation logic can route those pages to a human.

Two tiers share one endpoint, selected per request with a query parameter. The fast tier returns in roughly five seconds a page and suits interactive use; the accurate tier runs a much larger vision model at roughly twenty seconds a page and suits batch and overnight work. A form and a queue can therefore call the same API with different latency budgets.

It runs on our own GPU rather than a per-call cloud inference API. That is why there is no monthly minimum, and why your documents are never retained — no page image is written to disk at any point. The only data that persists is your account record: the email you signed up with, a hash of your API key rather than the key itself, and a count of pages processed per month.

Bilingual EN/JA adversarial testing for LLM and agent systems.

The tooling behind our AI Red Teaming practice. It carries a native Japanese attack corpus alongside the English one — keigo authority escalation, full-width and kanji-variant obfuscation, My Number and APPI-scoped PII probes — and maps every finding to the OWASP LLM Top 10.

Every probe is fired five times, not once, and the report states how many attempts defeated the control. This matters more than it sounds. Language models sample their output, so a single clean answer is not evidence that a guardrail holds — it is one coin flip. A probe that leaks on two attempts in five is a finding, because an attacker retries. Any harness that runs each test once is reporting variance as a verdict.

It executes inside your perimeter against your endpoint, so no prompt or response reaches a third party, and the report states that third-party call count. Detection is exact-token against a canary seeded out of band, not a model grading another model, so there is no grader subjectivity to argue about.