Your files already know. Nobody ever asks them.
Every invoice, contract, payslip and scanned receipt lands in the same place: a downloads folder, a Drive inbox, a phone camera roll. Filing them is the chore you notice. The real loss is quieter: two years of documents sitting in one place, and not one of them ever asked what they add up to.
Orden reads what you drop into one folder and files it where you already keep that kind of thing. Then it does the part filing was only ever a prerequisite for: it tells you this electricity bill is 41% above its own twelve-month average, that the accountant's invoice has not arrived and is nine days late, and what the year has cost you by category. Nothing is guessed. Every figure was read off a page you already own.
Runs on Google Cloud Run. Reads with Gemini 3.5 Flash, screened by Gemma 4. Your documents stay in your Drive.
It is four hundred, accumulated over two years, in eleven different naming conventions and none of them yours.
A supplier changes their invoice template and the rule that matched it stops firing. Nothing tells you. The folder just quietly stops filling.
A folder can only show you what is in it. The bill that never arrived, and so never got paid, is invisible to every filing tool ever written. It is the one that costs you money.
The answer to what this year cost, and to which supplier quietly raised their price, is already sitting in scan_0043.pdf and four hundred like it. Reading them is a day of work nobody has.
What changes
Filing one document by hand
- Open it and read enough to know what it is
- Work out who sent it and when
- Decide which folder it belongs in
- Rename it to match everything else in there
- Move it, and remember you did
Filing one document with Orden
- Drop it in the inbox folder
Drive tells Orden the moment it lands. It reads the document, files it, and writes a searchable description onto the file. If it is below your confidence threshold it waits for you in the queue instead.
Then it starts answering questions
A filed document is worth little on its own. A year of them, read and recorded, answers questions no single file can. Orden works this out by arithmetic over figures it already read at filing time, so it costs nothing to run and there is no model left to hallucinate.
It knows what this issuer usually charges, because it read every earlier one. When the next breaks the pattern by more than 15%, that is the thing worth your attention, not the filing.
Three documents establish a rhythm. Once Orden knows a bill is monthly, its absence becomes visible. No other filing tool can see a file that was never dropped in.
Totals by category, over twelve months, off the amounts read from the invoices and receipts themselves. Payslips are filed but never added in, because that sum would mean nothing.
A passport that expires, a contract that auto-renews, a bill due. Orden writes them to a calendar of its own, so a date buried on page four stops depending on you remembering it.
And the direction this points: once an agent knows the rhythm of your archive and has been right about it for months, the next step is acting on it rather than reporting it, chasing the invoice that has not arrived and flagging the renewal before it charges you. Orden does not do that yet. It is earning the right to.
Why not a rules engine
It reads, it does not match
Hazel and its kin match filenames and literal strings, which means one rule per vendor and silent breakage when a template changes. Orden reads the document, so an issuer it has never seen needs no configuration at all.
Everything, not just PDFs
Documents, spreadsheets and raw email. Photos and screenshots, with the text read straight off the image, so a phone photo of a receipt is filed as a receipt. Audio, transcribed and searched for commitments. Saved links.
In whatever language the document is written in, reporting back to you in yours.
It shows its work
An agent that moves your files unasked has to be answerable for it. Every decision opens a panel: the sentence that decided it, the facts read off the page, and which of your own past corrections applied.
The history reports how often you left things where Orden put them. That number is the one the feedback loop exists to move.
It learns from being corrected
Move something Orden filed and it notices. Correct the same sender twice and the correction is promoted to a stated rule, written in plain words on the screen rather than buried in a prompt you never see.
Architecture
Scroll the diagram sideways to follow the whole path.
Serving from Cloud Run · europe-west1 · revision orden-00019-qrk · gemini-3.5-flash + gemma-4-26b-a4b-it · state in Firestore
Pinned, not a -latest alias. Structured JSON out, thinking off: filing is extraction, not deliberation.
A second model, screening junk before the expensive read. Can only ever veto, never decide.
The core exposed as tools, so the agent chooses what to do rather than following a fixed script.
Built from source, state off the instance, and an hourly sweep in case a push notification is ever missed.
Point it at one folder and let it read the rest
Orden asks for access to your Drive, watches the folder you name, and leaves everything else alone. Nothing is moved without a record, and anything it moved can be put back. What it learns along the way is the part you keep.
See it work