Before anything can be found,
it has to be readable
You have twenty years of paper. Some of it is still paper. Most of it is on a disk somewhere — scans, old Word files, spreadsheets, photographs of tickets taken on a phone. Some of it opens. Some of it does not. Nobody can search a box.
It never loses a file, and it is never stopped by one
Strategic Agent starts here. Not with questions — with reading. Conversion takes each file you already have and makes a plain, readable copy of it that says where it came from. The original is never touched.
What conversion is
one jobThe copy is text. Any machine can open it. The original stays where it was, unchanged, and the copy sits beside it carrying a note of which file it came from and how it was read.
That is the whole job. A scanned ticket becomes a page of text, with a note saying which scan it came from. A spreadsheet becomes rows of text, cell by cell. An old Word file becomes the words that were in it.
It reads the common kinds — text files, Word files, spreadsheets, and PDFs that already have text inside them. Scanned PDFs, which are only pictures of pages, go through a separate and slower reading pass. It does not read everything, and the page on what it cannot read lists what it does not, and why.
Nothing is made up in this step. Nothing is summarised. There is no language model in it at all. If the ticket says 376 m, the copy says 376 m.
The promise, in two halves
both are requiredIt never loses a file
Every file it walks past gets a written status. Read fine. Spreadsheet, read as rows. A picture only. Empty. Could not be read, and here is why. Not found on the disk. There is no bin marked "skipped" that nobody ever looks in. If a file could not be turned into text, the list says so, by name.
It is never stopped by one
Some files hang the program that opens them — a bad PDF, a corrupt old document. So each file is opened in its own small box with a clock on it. If the clock runs out, that one file is set aside, marked, and the run carries on. One bad file costs seconds, not the day. And a run stopped halfway picks up where it left off rather than starting again.
Why both halves matter
A tool that stops on a bad file gets switched off within a week. A tool that quietly skips one is worse, because you will keep trusting it — and you will not find out what is missing until the day you need the thing that was missing.
What it has been run over
counts, not estimatesIt has been run over 56,687 real files from the founder's own company archive. Of those, 53,776 came out as readable text — 98 per cent clean. Two files hung the reader and were set aside for a person to look at. Two, out of 56,687.
It got there by failing first, and the failures are worth telling because they are what the product is made of.
The first version opened old documents with a heavyweight office program, and 578 of them froze it. Swapping to a small reader that opens one file and quits took that to three. The first version restarted the whole job every time one file hung, and managed about three files a minute. Opening each file in its own box took that to around 2,250 a minute, with no good file wrongly set aside.
One counter once showed 15,800 files done when only about 4,400 really were — the same files being retried over and over, each retry counted as progress. It now counts each file once. A number that goes up is not progress unless it counts things you actually have.
And after a folder was moved, 1,097,899 of the notes saying this copy came from that file pointed at nothing. They were repaired, to zero. Every copy still points at its original.
What this page does not claim
These counts come from the founder's own records, which is the only body of material this has been run over at scale. There are no customers yet. The photographed documents in that archive are still not searchable, and a queue of scans has not been read. Both are listed on what it cannot read and on the record.
That is the filing cabinet. It is not clever. It is careful, and it writes down what it did.