The noise nobody sees
If you have ever put a folder of PDFs into Gemini Notebook, a Claude Project, Copilot or a company chatbot, you may have watched it quote “Page 3 of 10” or “Confidential — Internal Use Only” back at you. That text was never content. It is the header, footer and page number printed on every page — but when text is pulled out of a PDF, it comes out with everything else.
A tool that reads a PDF gets the words on the page, not what they are for. The running header, the footer and the page number all come out as ordinary lines. A saved web page brings its menu, its cookie notice and its footer along. Then the AI tool cuts the text into pieces and searches them when you ask a question. The same footer ends up in piece after piece, and a piece that crosses a page break has a header sitting in the middle of a sentence.
None of that is dramatic on its own. But it is in almost everything, and you never see it — because you look at the PDF, not at the text the tool actually got. That gap, between the document you see and the text the machine reads, is the problem Michael Shadpour kept running into. He makes Agnos at Gidion, and he built it to close that gap.
What Agnos does, and what it shows
Agnos is an app for Windows 10 and 11. You drop in a folder of text-based PDFs, Word, Excel or PowerPoint files, or web pages, and it takes out the parts that are not content: repeated headers and footers, page numbers, cookie banners, boilerplate and duplicate passages. It keeps headings, tables, code and numbers. Then it exports the clean text — documents to upload, chunks for a RAG index, or training data.
The part Shadpour cares about most is that Agnos does not remove anything silently. Review lists what was removed from each file, up to 500 removals per file, with the rule behind each one. If you disagree, you can keep a piece in one click where its rule allows, or switch the rule off and clean the file again. The cleaner is accountable to the person using it, line by line.
That accountability is the product's real argument. Plenty of tools clean documents; Agnos is built for people who need to trust the cleaning — people feeding client contracts, staff handbooks and internal policies into AI tools, where a silently dropped table label or a quietly kept footer is not a cosmetic issue but a wrong answer waiting to happen.
Publishing the misses
Because cleaners get things wrong, and that includes his — Shadpour says so himself. He tested Agnos on 25 public PDFs from US federal agencies, 1,342 pages in all, and published everything. It removed 2,063 lines as page furniture. When every one of them was read, 200 looked like content: table cells, column headings that repeat on each page, table labels and phone numbers.
It also kept things it should have removed. The biggest was a four-line NASA footer on all 89 pages of one document, which sits just past the lines Agnos checks at the bottom of a page. Shadpour could have left those results out. He put them near the top instead. If a tool can be wrong, you need to be able to see where — that, he writes, is the whole point of Review.
The study itself is public and reusable: every document, the method, what Agnos got wrong, a CSV with 46 columns per document under a CC BY 4.0 licence, and two short Python scripts to repeat it. There is also a free browser tool that finds the repeated lines in a PDF without uploading it. It is an unusual way to sell software — lead with your error rate — and it only works because the product's answer to the error rate is built in: here is every removal, here is the rule, change it if we are wrong.
Local-only, pay-once
The documents people put into AI tools are often the ones they are most careful with: client contracts, staff handbooks, internal policies. Uploading them to one more service just to clean them seemed backwards to Shadpour. So Agnos does all its cleaning on your computer, and it works with the network off. It goes online only to activate a licence key, check it about once a week, or deactivate a device. It never sends a document or a file name.
The pricing follows the same logic. Most people he pictured using Agnos clean a batch of documents, upload them, and come back weeks later when the files change — a monthly bill for that felt wrong. So it is 14 days free with everything, no card and no account. After that, Pro is $99 once for one person on up to two devices, and a Company Pack is $399 once for up to five people at one company, on up to ten devices. Nothing renews, and a 2.x licence does not expire; a future major version may be a separate purchase, and whether to buy it is up to you.
Shadpour is equally plain about what Agnos is not. It is Windows 10 and 11 only, with no Mac version. It handles text-based PDFs only — scanned pages are skipped and named in its warnings, not cleaned badly. Taking out a “Confidential” footer is not the same as taking out confidential information, so it is not a redaction tool. And it is not a library for developers; if you write Python, he points you to open-source tools like Docling and documents where each tool is the better pick. It takes out one kind of noise, he writes. It will not make a chatbot right about everything. If you try it on your own documents, he would like to hear what it gets wrong — he answers every reply himself.
The playbook
- Show your work, including the misses — Agnos lists every removal with its rule, and its founder published a study showing 200 of 2,063 removed lines looked like content. Transparency is the feature.
- Price for the usage pattern, not the fashion — Batch cleaning every few weeks does not justify a subscription, so Agnos charges once — $99 for an individual, $399 for a company pack.
- Keep sensitive work off the network — Contracts and handbooks should not be uploaded just to be cleaned, so Agnos runs entirely on the user's PC and works offline.
- Say what the product is not — Windows only, text-based PDFs only, not redaction, not a developer library — clear boundaries that keep the wrong customers away.