Document prep for your own AI
Not a chatbot. You upload documents. We prepare them. You download a ZIP. You keep it and point your AI at it — the one on your computers, or the stack you already run. A text pile often finishes in minutes. A scan-heavy pile can take about an hour. Every brick is $10, and only if the ZIP ships. If something in the files was too messy to trust, we list it — we don’t hide it. That is a different bill from paying a model to reread the files on every question, or paying people for weeks to build a search pile (why $10).
Pay on success You keep the ZIP No training on your data
Why this instead of the usual way
Most people hand their files straight to an AI. That is the usual option. It is also how confident, wrong answers get made. We prepare the files first, on purpose, to cut those failures down.
A chat will answer even when the file is a scan, a mess, or missing the page. The voice stays calm. The sentence was not in your document. If we cannot read a stretch, we leave it out and write down that we left it out. We do not fill the hole with a guess and call it your manual.
A caption, a header, or a broken table can be treated as the procedure. The answer cites “the document” and is still false. We keep the real passages and mark the parts that are not safe to answer from, so your AI is not treating junk as authority.
Drop a long pile into one chat and the model keeps the beginning and the end. The step in the middle disappears. We pack a bounded book, not your whole archive in one window, so the page you need is still a page, not a blur in a giant paste.
Preparing the files makes those failures harder to produce. We cannot promise that your AI will never be wrong. We can hand it a package that does not invite the mistake, and a list of what we refused to pretend. You keep the package. You point your own AI at it.
Proof
Public sample ZIPs, mute lists, and a plant write-up — open them yourself; don’t take our word for it.
A real portable ZIP from the public shelf — manifest, chunks, mute flags, and a craft brief you can open before you pay.
More details →Same figure-shell chunk: dump-RAG would index it; the plant mutes it and writes the reason into the package.
More details →Receive → Extract → Structure → Package, with a closed fail path and a real manifest snippet — manufacturing, not a chat demo.
More details →Plain-English Notes for humans and models, plus a small retrieval set on Hugging Face if you want to measure.
More details →Why a mill
We prepare a portable package you keep — the text, a list of what we left out, and the files your stack needs — so your own model or search has something it can use. The computer and the model matter. The documents decide whether the answers hold up after the box is online.
Audit-oriented package: manifest, hashes, mute list, craft brief — your stack, your models. No hosted lock-in.
Weak or unreadable stretches are muted and listed off the answer path, so you can see what we refused to pretend was knowledge.
Card authorized at Go; captured only if a usable package ships; voided if the plant fails. Refuse-before-run = no charge.
One price
A text pile and a scan-heavy pile are the same charge. A scan-heavy pile can take about an hour. A text pile often finishes in minutes. That difference is time, not price. Cap: 15 files / 150 MB, one job at a time. Over that → refuse, no charge.
The usual alternatives are paying a model to read the same files again on every question, or paying people for weeks or months to build a search pile by hand. A hundred bricks here is about a thousand dollars. That is a different bill, not a smaller version of a staffed build.
How we handle your files
HTTPS in, ZIP out — we don’t host your knowledge base. 72 hours to download if you haven’t yet; about 30 minutes after your first successful download (re-download refreshes that grace), then we purge our copy. Failed jobs clear in 24 hours. Named operator; named processors on the diligence pages.
How it works
The hopper is the product — not a lead form for a sales call. Your model and stack stay yours.
.eml, contracts, and forms are in the plant magazine
(see file types); full mailbox dumps are not.
Unsupported types refuse before processing.
Try your documents
Look at a sample package first, then try your own files. A text pile often finishes in minutes. A scan-heavy pile can take about an hour. Policy-locked work: who this is for · ask about an isolated path.
Who mills this
For a card hold on a job that can take an hour, you should know who is on the plant. Questions, invite coupons, and restricted / on-box conversations go to the same person.
Contact operator
No account. No support portal. Invite coupon requests welcome. Same inbox as the operator card: [email protected].