Skip to content

Exactly what happens to your document

The processing pipeline in detail: validation, private storage, deterministic extraction, model analysis, page citations, retention and deletion.

Your document is validated, stored in private encrypted object storage with no public URL, checked for structural safety, and read. Facts that can be read deterministically — application number, dates, examiner, art unit, section headings — are extracted by our own code. Selected pages are then sent to a language model through OpenRouter for summarisation and edge-case classification, under an arrangement that excludes training on your content. Every fact the model reports must cite the page it came from, or it is discarded. Nothing is published, and deletion removes the file and every artifact derived from it.

An AI is going to read my patent application. What does that actually mean?

It means a language model receives selected pages of your document in a single request, returns a structured summary, and is given no authority over anything that matters.

It does not calculate your deadline, it cannot report a fact without citing the page it came from, its uncertainty is shown to you rather than smoothed over, and a person checks its output before you see it.

The pipeline, step by step

  1. Validation

    The file must genuinely be a PDF, within the size and page limits, unencrypted, structurally sound and free of embedded active content. Anything else is refused with a specific reason and is not processed.

  2. Private storage

    The document is written to encrypted object storage with no public access. There is no URL that serves it to the public web, and the application never returns document bytes to a browser.

  3. Deterministic extraction

    Our own code reads the text layer for the facts that have reliable printed forms: application number, mailing date, examiner, art unit, statutory headings, claim numbers and cited references. Each fact records the page it was found on.

  4. Model analysis

    Selected pages are sent to a language model through OpenRouter for the summary and for classifying cases the deterministic pass could not settle. The model is configurable and the API key is held server-side only.

  5. Grounding and confidence

    Every fact returned by the model must resolve to a page in your document. Facts that do not are discarded rather than shown. Low-confidence facts are marked uncertain, and where confidence is too low overall we abstain and route the document to a person.

  6. Deterministic deadline calculation

    The reply period is calculated by rules-based code from the mailing date and the action type. No model is involved in that calculation, in either direction.

  7. Human review

    A reviewer checks the output against the document before it reaches you.

What we do not do

We do not use your document to train or fine-tune any model.

We do not make your document reachable by any public URL.

We do not send your document to a provider by handing them a link to fetch it; the pages we send are transmitted directly in the request.

We do not let a model calculate your deadline.

We do not let a model draft a response to the USPTO.

We do not show you a fact that cannot be traced to a page in your document.

If automated analysis is switched off

The service has a switch that disables model analysis entirely. With it off, you still get the deterministic extraction, the reply period, and human review — the summary is shorter and more literal. This is a supported mode of operation rather than a failure state, and the interface tells you when you are in it.

Deletion

Ask, and we delete the stored file, the extracted text, any page images, the analysis output and any cached artifact derived from your document. We record that the deletion happened and when. What remains is the case record without document content, so that we can show what we told you and when.

Ready when you are

Questions people ask

The model is configurable and is reached through OpenRouter rather than being wired directly to one provider. That means we can change model when a better or cheaper one becomes available without changing how your data is handled, and it means we are never locked into a single vendor.

Language models can produce confident wrong answers, which is why the design does not trust the model for anything structural. Deterministic code extracts the facts that matter, every model-reported fact must cite a page, ungrounded facts are dropped, and a person reviews the result before it reaches you.

Only the pages needed for the analysis, selected after the deterministic pass, and never more pages than the configured limit. The pages are sent in the request itself rather than as a link the provider fetches.

Yes, with the version of every component that produced it — pipeline, extraction rules, deadline rules, model and prompt version — so that a result can be explained and reproduced later. Deleting your document deletes the analysis with it.