Features

Identify

Identify the documents in a folder

Drop a folder: each file gets a proposed type, its date, its issuer and the person it concerns. Nothing is renamed, nothing is moved: this is a reading, and you can contradict it row by row.

• 100 types identified: payslip, tax assessment, identity document, company registration extract, bank details, bank statement, tenancy agreement, rent receipt, employment contract, judgment, notarial deed, insurance certificate, loan offer, inventory of fixtures, health-insurance entitlement statement, prescription, public-tender forms, claim notification, annual accounts, formal notice, and the rest.
• Three confidence levels, and the difference matters: identified, to confirm, unsorted. A document the tool is unsure about is never filed with confidence.
• The date that names the document, and its role: the issue date of an invoice, the period of a payslip, the expiry date of a certificate.
• The issuer and the person, when the document says so: that is what allows sorting by client rather than by heap.
• The clues are shown: click a document and the tool says what identified it. When it found nothing, it says that too.

What it does not do: the tool certifies no document genuine, and it never says a file is complete or acceptable. It proposes a type, you approve it.

Identify one document

A single document, and everything the tool infers from it: the type, the date, the issuer, the person, and the list of clues that led to that answer.

• The answer, and its confidence level. An identified document, a likely one and an unclassifiable one are not stated in the same way.
• The clues, in plain words: the words that counted, the codes read, the layout. That is what lets you contradict the answer.
• The dates found, with their role: the one that names the document, the period covered, the expiry date.
• The codes: 2D-Doc, machine-readable zone to the international civil aviation standard, QR. A code either reads or it does not; when it reads, it outranks everything else.

What it does not do: an answer is not a certificate. The tool says what type a document looks like, not that it is genuine or valid.

Name

Rename a folder of files

scan0012.pdf becomes 2026-03-14_invoice_EDF.pdf. The table shows every row before anything is written, and once applied, it all comes undone.

• A token-based naming pattern: {date}, {type}, {emetteur}, {personne}, {original}, {numero}. Six ready-made patterns, and yours if you prefer.
• The before and after preview, row by row, with the reason for each name. Nothing is written until you approve.
• File-system constraints are respected: forbidden characters, Windows reserved names (con, nul, lpt1), paths beyond 260 characters, case collisions, macOS decomposed accents.
• A verified copy, not a blind move: every file written is read back and compared against its digest before the original is touched.
• A log, and the undo. On a real folder, 174 documents were sorted, verified one by one, then put back in 149 milliseconds.

What it does not do: the tool does not invent a name when it has identified nothing: the document keeps its original name and goes to 00-a-classer/. An invented name would be worse than an ugly one.

Naming pattern

A naming model is judged on a real folder, not on an example. Choose your tokens, drop your documents, and read the “after” column.

• Six ready-made patterns: by date, by person, by issuer, numbered for a list of documents, and the house pattern.
• Empty tokens disappear cleanly: when the issuer is unknown, the name does not keep an orphan separator.
• The pattern is checked before it is applied: an unknown token is reported, not ignored.
• The original name stays available with {original}, when traceability matters more than readability.

What it does not do: a naming pattern does not create missing information. If the document states no date, {date} stays empty, and the tool will not go looking elsewhere.

Sort

Sort a folder into a tree

Six composable trees, and a 00-a-classer/ folder for whatever the tool could not identify. What is not identified is never filed at random.

• Six trees: by document family, by person, by year, by issuer, by type, and the composition of several levels.
• A level that cannot answer disappears: a document with no identified person does not create an “unknown” folder.
• 00-a-classer/ at the top of the tree: what was not identified is visible first, not diluted at the bottom.
• The tree is shown before writing, folder by folder, with how many documents each holds.

What it does not do: the tool deletes nothing, ever, and it merges no files. A document it cannot file stays a document to file, not a lost one.

Find duplicates

Two files identical byte for byte are a certainty. Two scans of the same document are not: those are proposed, and you decide.

• Byte-for-byte identicals, by SHA-256 digest: certain, and instant. 142 megabytes of digests in 1.3 seconds.
• The same document under two files, recognised by content, and only proposed.
• The same scanned pages, brought together by an image digest used only as a pre-filter: the text decides.
• Series, which are not duplicates. Three payslips from three months look very much alike. Mistaking them for duplicates would delete two payslips out of three.

What it does not do: nothing is deleted. The tool does not copy certain duplicates into the sorted folder, and it records in the log what it did not copy.

Check what is missing

Check against a list of expected documents

The question everyone asks in front of a folder: what is missing? Choose a list, drop the folder, and read the answer document by document.

• Five ready-made lists, each with its source and review date: mortgage, rental, public tender, new employee, insurance claim.
• The count, not just the presence. “The last three payslips” is not “a payslip”, and the tool says “2 of 3”.
• The continuity of a series: three payslips from January, February and April, and it returns March, because March is what you must ask for.
• Freshness: a three-month-old company extract may be a rejected one, and it is better to know before sending.
• Per person: a two-borrower file complete for one says nothing about the other.
• What may not be demanded: French law forbids asking a rental applicant for their bank details. When such a document is in the folder, the tool flags it.

What it does not do: the tool never says a file is acceptable. It says what is missing against a list you chose, and the list carries its origin and its date.

Mortgage application file

The list is shown on this page, and you need drop nothing to read it. If you do drop the folder, the tool says what is missing, for each borrower.

• The expected list: identity, proof of address, last three payslips, latest tax assessment, last three bank statements, bank details, and depending on the situation rent receipts, property tax or a company registration extract.
• Per person: two borrowers, two columns, and the tool does not mistake one complete file for the other.
• Series are checked month by month: if August is missing between July and September, August is what is asked for.
• A chasing message ready to send, with no amounts, no reference numbers and no file names.

What it does not do: every bank has its own requirements, and this list is a baseline, not a contract. The tool does not say a file will be accepted.

Rental application file

French law sets out exhaustively what a landlord may ask for. The tool checks what is missing, and flags what should not have been asked for.

• The permitted list: an identity document, proof of address, proof of employment, proof of income.
• And the forbidden list, which is the real value of this page: bank details, bank statements and the URSSAF compliance certificate may not be demanded of an applicant.
• The flag is visible: when a forbidden document is in the folder, the tool shows it with the text that forbids it.
• A chasing message that asks only for what may lawfully be asked.

What it does not do: the tool is not legal advice. It quotes a public text and dates it; the decision, and the responsibility, stay with the landlord.

Public tender application

An application is rejected for an expired certificate as often as for a missing one. The tool looks at both.

• Certificates and their freshness: company registration extract less than three months old, valid URSSAF compliance certificate, tax compliance certificate, insurance certificate.
• An expired document is not a present document: it is shown separately, with its date, so you know which one to ask for again.
• Bank details, and what the list does not yet cover.

What it does not do: the French application forms (DC1, DC2, the letter of engagement) are not yet identified: they are listed as extra documents, not as expected ones. That is the next piece of identification work.

New employee file

An onboarding file is assembled in a week and chased over four emails. The tool says what is missing in one go.

• Onboarding documents: identity document, bank details, proof of address, signed contract, and depending on the role the qualification and the occupational health appointment.
• What an employer may not demand: an employee's tax assessment and bank statements are none of the employer's business. When they are in the folder, the tool flags them.
• A chasing message ready to send to the employee, quoting no sensitive data.

What it does not do: the tool does not check that a residence permit allows work, and it replaces no filing obligation.

Insurance claim file

A claim file rarely drags on for a single document: it drags because nobody has the list in front of them at the same time.

• Depending on the claim: motor (accident report, vehicle registration) or home (proof of address).
• Quotes and invoices, with the expected count: two quotes are often required.
• The initial medical certificate, when there is personal injury.
• Photographs are listed, not interpreted: the tool counts and files them, it does not say what they show.

What it does not do: the tool values nothing and assesses no loss. It says which documents are there and which are missing.

Chasing message

Chasing a document takes three minutes of writing every time, and one always gets forgotten. The message comes out written, with the exact list, and the precise month when it is a series.

• The exact list, and nothing else: what is missing, and for whom when the file has several people.
• The precise month of a series: “July 2026 is missing”, not “a payslip is missing”.
• What the message never contains: an amount, a reference number, an address, a date of birth, a file name. It goes to someone, often in copy.
• Documents to check are kept apart from missing ones: you do not ask again for something that may already be there.

What it does not do: the message is not sent by the tool: it is copied, and you send it with your own mail client.

Check twenty files at once

Twenty applications for one property, ten new hires in a month: the same check, done twenty times from memory. Drop the folder that holds them all, and read one table.

• One subfolder per applicant is the only rule. That is how applications arrive, and that is what the tool takes for a file: it does not guess groups from the names read inside the documents.
• Incomplete files come first. They are the only ones with anything to do about them; complete files read at a glance at the bottom.
• One chasing message per applicant, ready to copy, with their own list. You do not chase twenty people with one message.
• The tracking table as CSV: one row per file, its state, what is missing. It carries no file names, because it circulates.
• Situation boxes apply to the whole batch (employed, with a guarantor): it is the same listing and the same property.

What it does not do: no document passes from one applicant to another, and a test checks it. In exchange, a document left at the root of the drop belongs to nobody: it is counted in no file.

Produce

List of documents as PDF

The document you attach to the file: every item numbered, with its type, date, issuer and page count. And, if a check was run, the list of what is missing.

• The numbered table, in the order of the sorted folder.
• The missing documents on the same sheet: a list that hides the gaps is not a list.
• SHA-256 digests, optional: they state that the file listed is that one.
• As PDF, and as CSV to reuse in a spreadsheet or in Klargrid.

What it does not do: a digest proves a file has not changed, not that the document is genuine. The list certifies nothing.

Split a stack

Split a scanned PDF into documents

The feeder scanned everything at once: thirty documents in a single file. The tool proposes the cuts, shows them on a strip of thumbnails, and each one is corrected with a click.

• No separator sheet is needed. The tool looks for what CHANGES between two pages: internal numbering restarting at 1, a header with not one word in common, the document type announced at the top.
• And for what proves a document continues: a sentence cut in two, a footer that repeats, a section heading (“PAR CES MOTIFS”, “Article 3”). No issuer ever opens a document with a closing formula.
• Every cut says why, in one sentence, under the thumbnail. A bare score cannot be argued with.
• One click on the line between two pages adds or removes a cut. That is the only gesture on the screen.
• Blank pages and separator sheets are set aside, shown struck through, never deleted silently.
• The documents that come out go through everything else: type identification, naming, filing, list of documents.

What it does not do: the tool does not guess what a document means. It spots breaks between two pages, and when two documents follow one another with no signal at all it misses one: that is measured, and it is fixed with a click.

Remove blank pages

A double-sided scan of a single-sided file gives one blank page in two. They are spotted, shown struck through on the strip, and you see which ones before they go.

• A blank page is shown, not deleted. It stays on the strip of thumbnails, struck through, with its reason, and the box can be unticked.
• Nor does it cut wrongly. A blank page inside a document is common, and the tool treats it as a clue, not as a rule.
• The original file is never modified: the documents that come out are new files.

What it does not do: an almost-blank page (a stamp, a page number on its own) is not counted as blank. Better one page too many than one page lost.

Put a scan back in order

A single-sided feeder does not turn the sheets: you scan all the fronts, flip the stack, scan all the backs, which therefore come out in reverse. Every page is straight and legible, and the document is unreadable.

• The tool compares two orders, the file's and the one obtained by interleaving the two halves, and measures which holds together: repeated footers, headers that follow on, numbering that runs on, sentences that continue.
• Both conventions are tried: backs in reverse (the whole stack flipped) and backs in order (each sheet turned).
• The swap is proposed, not decided. Reordering a stack that was already in order would be the worst way to lose a file: the gain in coherence is shown as a number, and the box can be unticked.

What it does not do: the tool does not deskew a slanted page and does not rotate an image. A page fed in crooked keeps the rotation the scanner wrote.

Separator sheets

Splitting without separators finds 98.3 % of the cuts. With a sheet slipped between two documents, it finds them all: that is the only way to be certain, and it costs one print.

• The sheet carries its title in large type: that is what the tool recognises, even on a poor scan.
• It is removed from the stack, shown struck through on the strip, and enters no document.
• It can be reprinted and reused at will: it carries no date, no file number, nothing that goes stale.

What it does not do: the tool does not yet read a barcode on the sheet: it recognises its printed text. A sheet photocopied too faintly may go unnoticed, and the cut is then found like the others, by the signals.