Two-column addressee and sender
When the invoice's addressee sits on the left and the sender on the right, we read them column by column so they don't get mixed up.
Fuseji finds personal data such as names, addresses, phone numbers, and account numbers in Japanese PDFs, and redacts it by burning it into the page so it cannot be recovered. Your documents never leave your computer.
In a PDF where a black box was simply placed on top, selecting everything and copying brings the hidden text right back. Fuseji rebuilds the page and removes the text underneath.
A black box placed on top
Name:
Select all → Copy → Paste
Name: Taro Yamada
The hidden text comes right back
Fuseji (burned in)
Name:
Select all → Copy → Paste
Name: (nothing comes out)
Every export is searched automatically to confirm nothing remains
Fuseji uses rules and a Japanese ML model to find personal data such as names, addresses, phone numbers, My Number, and bank accounts, and shows them as candidates. The model is our own, trained only on synthetic data with no real people's information. You decide what to redact while looking at them on screen.
Redaction is burned into the page. Text, images, metadata, hidden characters, annotations, and attachments are all removed. Nothing survives a copy-paste, and nothing peeks out from under a black rectangle.
Right after export, Fuseji runs a full-text search on the new file and writes a verification report of what was removed and what was checked. On Pro and Office, strict verification also reads the page images with OCR.
The Fuseji app does not send your documents, images, detection results, or reports anywhere. The app has no server, no account, and no analytics.
Opening, detection, redaction, verification, and the report all happen on your computer. The original file is never modified; Fuseji always writes a new one.
The app connects to the internet only to activate, deactivate and weekly re-check your license key, to check for updates once a day (you can turn this off), and when you press “Send a case that didn’t work” or “Send feedback”. The case report replaces every character of the document with its kind (kanji, kana, digit…) and contains no document, words or file names. Feedback contains only the text you write and, if you enter one, a reply email address — no documents or file names. You see the full contents before sending either, and you can turn both off in Settings.
The app has no analytics, no crash reporting, and no install IDs. The Settings screen lists when the app goes online.
We also issue offline license keys that work without an internet connection (Pro and Office plans). When your computer is online, the daily update check also downloads a list of revoked keys and compares it on your computer (nothing is sent; turning off the update check stops this too).

Click a term in the right-hand list to jump to its next occurrence. Drag on the page to add a region (seal, photo, signature) as a redaction candidate. Add a term to apply across the whole document in one click.

After export, Fuseji runs a full-text search (plus OCR on Pro and Office) to confirm the redacted terms do not appear anywhere, and writes a report.

Choose from four redaction styles: black, label (【氏名1】 etc.), white-with-border, or underline. Click “Preview” to see exactly what the exported file will look like, right on screen.

Keep the reason for withholding information, such as “Article 5, item 1,” inside the black box and in the report. The Freedom of Information Act's clauses are built in, and you can register your own ordinance clauses too.
Detection is a set of candidates. Toggle each on or off in the list, and mark each page as confirmed. The report records your confirmation state.
When a page has white text, tiny text, off-page text, or text hidden under images, Fuseji flags the page and rasterises it on export so the text cannot survive.
Pages without a text layer are read with on-device text recognition (Japanese), then run through the same detector.
Open a PDF redacted with another tool, and Fuseji checks it for residual text, hidden characters, and metadata. This is free.
Real Japanese documents aren't always one tidy line of text. Address fields are narrow and the address wraps, the addressee and the sender sit side by side. Fuseji works through the kinds of layout quirks that show up in real paperwork, and fixes what we miss.
When the invoice's addressee sits on the left and the sender on the right, we read them column by column so they don't get mixed up.
Even when the only label on the table is "Representative name" or "Address," we still pick up the values in that column.
"Ya ma da Ta ro" — common on forms and in scanned PDFs — is read as a single name.
When a narrow field makes the address run into the next line, we pick up the building name and room number too.
Names like "カ)〇〇" or "〇〇(カ" are read as a single account holder, and we tell company from person.
If the visible Japanese can't be pulled out as text, we read the page as an image.
We still miss things. So Fuseji always asks you to confirm on screen before redaction. If a document didn't work well, let us know from 「うまくいかなかった例を送る」 (Report what didn’t work) in the app. The report replaces characters with their kind, and the document's text is never sent.
Drop in a PDF you want to verify. Fuseji reads it locally and checks for residual text, hidden characters, annotations, attachments, and metadata. Detection, review, export, and the inspector are all available in the free version — forever.

Sales and payments are handled by Polar (Merchant of Record). USD, all taxes shown at checkout. The Pro and Office plans support bank-transfer invoicing.
Beta pricing: 50% off until December 31. The discount is applied automatically when you buy below (code BETA50).
Referral code : 20% off. The discount is applied automatically when you buy below.
one-time, 1 computer
Buyone-time, 2 computers
Buyone-time, 5 computers
BuyNeed 6 or more seats? Volume licenses are ¥7,000 per seat (¥6,000 from the 21st seat, tax included). Contact us for a quote, or create one yourself (Japanese only).
After the first year, every version you could already use keeps working. Volume licenses are available for 6 or more computers.
Anything else? Email [email protected]. Please don’t send the documents themselves.
Yes. Redaction is burned into the page, so text, images, metadata, annotations, attachments, and hidden text are all unrecoverable. After export, Fuseji re-runs a full-text search (and OCR, on Pro and Office) to confirm the redacted terms don't appear anywhere.
It means removing the text under the redaction and rebuilding the page. Many PDF redactions just place a black box over the text, and the text is still in the file, so it can be copied or revealed by moving the box. Burning in removes the text itself, so it cannot be recovered.
A redaction where a black box sits on top of text that is still in the file. It looks removed, but copy and paste or an editing tool can reveal it. Personal data has actually been exposed this way in published documents.
It checks a PDF redacted with another tool for text under black boxes, hidden text, and metadata such as the author. It is free. Use it on PDFs you receive or are about to send.
Yes. Text outside the redactions can be kept for search and copying (an export option). You can also export every page as an image with no text at all.
We don't quote a number. Automatic detection surfaces candidates, but it isn't a guarantee. Make the final call on screen. Fuseji records your confirmation state but isn't liable for what you didn't catch.
No. Documents, images, detection results and verification reports are never sent. The app only connects to activate, deactivate and weekly re-check your license key, to check for updates once a day (you can turn this off), and when you choose to send a case that didn’t work or send feedback. The case report replaces the document’s characters with their kinds and contains no documents or words; feedback contains only the text you write and an optional reply email address. You can read either in full before sending.
No real people's information. We fine-tuned a Japanese model (BERT) published by Tohoku University, using only data we built by combining fictional names, places, and sentences. The model is bundled under CC-BY-SA-4.0, and the credit is listed in the app's “About Licenses” screen.
Yes. Choose the “Contract template” analysis type when you open a PDF, and Fuseji also surfaces amounts, dates, and periods as candidates alongside company names, personal names, and addresses, with underline set as the default style. You can turn a signed contract into a fill-in-the-blank template.
Yes (from 0.1.10). Open a PNG, JPEG or HEIC image (PNG and JPEG on Windows) and redact names, phone numbers and so on the same way as a PDF. The exported image is read back with text recognition to confirm the removed words are unreadable and that metadata such as location is gone before it is saved (the free version checks too).
Mostly not. The text recognition used for scans and photos is designed for printed text, and handwritten names or numbers often cannot be read. Add a range by hand on the review screen for handwritten parts (drag with the mouse, or draw with Apple Pencil on iPad). Ranges you add are burned in the same way as printed text. The post-export verification (checking that the removed words can no longer be read) also relies on text recognition, so it does not cover handwritten text. Please check the exported file on screen.
On Mac: Apple silicon (M1 or later) with macOS 13 Ventura or later; Intel Macs are not supported. On Windows: Windows 10 or Windows 11 (x64, 22H2 or later); Windows on ARM is not supported.
Detection, review, export, the inspector, and text masking all work in the free version. The free version adds a watermark to exports and masks up to 2,000 characters at a time. Saving the verification report and the redaction ledger start with Personal; batch processing, shared dictionaries, strict (OCR) verification, and redaction reasons start with Pro.
We are improving detection by collecting and fixing cases where it did not work on real documents. During this period it is 50% off until December 31. The beta has no feature limits, and the one year of updates included with your purchase covers the full release. If a document did not work well, let us know at [email protected]. Please don’t send the documents themselves.
Versions released within one year of purchase keep working with no time limit. From the second year, an optional “One Year of Updates” (25% of the purchase price, tax included: Personal ¥995, Pro ¥3,200, Office ¥9,950) gives you every version released during that year. It renews yearly and can be stopped at any time; versions you already had keep working. During that year we keep updating detection and redaction to follow changes in laws and document formats. We will tell you how to buy it before your first year ends.
From January 1, 2027, customers who bought the desktop version will see a referral code and link under Settings → License in the app. Anyone who buys through that link gets 20% off, and you get one free year of “One Year of Updates” for each referral (with no limit). The code is derived inside the app from your license key; the app sends nothing for referrals.
Open the customer page and enter the email address you used to buy. Sign in with the code we send, and your license key is shown there at any time. The key is also in the email you received when you bought (sent by Polar). If you cannot find it either way, email [email protected] with the email address you bought with or your order number. Please don’t send the documents themselves.
Yes. Within 14 days of purchase, for any reason. Refunds are processed by Polar.
If something didn’t work, use うまくいかなかった例を送る (Report what didn’t work) in the app (only a diagnosis with the characters masked by type is sent). For anything else, email [email protected]. Please don’t send the documents themselves.
Automatic detection surfaces candidates; it does not guarantee nothing was missed. The final check is yours.