Selected work

Flagship product · Functional desktop application

Stet

Word-native manuscript review automation for medical writers.

Stet reads reviewer comments and their context, drafts a revision and reply, gives the writer control over every suggestion, then exports the result as a real Word document with tracked changes and threaded replies.

546+

automated tests

102

comment stress test

.docx

native Word output

The problem

Review happens inside Word.

A medical writer handles reviewer comments one thread at a time: understand the request, inspect the surrounding manuscript, consult a source where needed, write a tracked revision and respond in the comment thread.

Generic AI chat breaks that workflow. Context has to be copied out manually, and pasting the answer back into Word does not create the revision history and comment structure the review process depends on.

Some documents also cannot be sent to an unapproved external API. The useful solution therefore has to address the document format, the human review loop and the deployment policy—not only produce plausible text.

The workflow

Automation with the writer still in control.

  1. 01 / OPEN

    Parse the manuscript

    Stet reads the OOXML package, including comments, existing revisions, paragraph structure and reference-manager field codes.

  2. 02 / DRAFT

    Generate in context

    The model receives the comment, target text, surrounding paragraphs and any source file attached to that thread.

  3. 03 / REVIEW

    Accept, edit or retry

    The writer sees a red-and-green diff, can refine the response and decides what enters the in-memory document model.

  4. 04 / EXPORT

    Return to Word

    Stet creates a new .docx containing native tracked revisions and replies nested under the original comments.

Context

The comment is only part of the request.

Each workbench card brings together the reviewer thread, target paragraph and expandable surrounding context. Supporting PDFs, spreadsheets or Word files can be attached to an individual thread.

Stet workbench showing a reviewer comment, manuscript context, a proposed text revision and a drafted response
The writer reviews the suggested revision and reply before accepting either change. Select the image to inspect it at full size.
Stet interface presenting a structured manuscript table alongside reviewer instructions and proposed changes
Structured document context is retained instead of flattening a table into an unrelated chat transcript.

Structure

Complex Word content remains document content.

The application works with the manuscript’s internal structure so tables, formatted runs, images and citation fields can survive the round trip through review and export.

Output

Review-ready Word output.

The exported document contains real w:ins and w:del revisions that Word can accept or reject, plus replies threaded under the original reviewer comments.

Microsoft Word document containing tracked revisions and a threaded reviewer reply created by Stet
A new Word file is produced; the uploaded source file is not overwritten.

The technical work

Document fidelity, not a chat wrapper.

Stet uses a custom parser and serializer built with lxml and zipfile. It works directly with the OOXML parts that represent the manuscript, revision markup, comment threads and citation fields.

Tracked revisions

Insertions and deletions are written as native OOXML revision elements rather than visual approximations.

Threaded comments

Replies are written across the Word comment parts and anchored back into the document.

Reference fields

EndNote and Zotero field codes are protected through paragraph editing and export.

Sequential context

When multiple suggestions are accepted in sequence, later comments can see the accepted text that came before them.

Evidence

Functional and demonstrated.

The original card-based application has been packaged as a Windows executable and demonstrated to industry. Automated coverage exceeds 546 tests. Stet has been tested using manuscripts with demanding reviewer-comment workloads, including a 102-comment test document.

These are technical validation points, not claims of measured customer time savings. A production benchmark remains to be completed.

Deployment

Local where policy requires it.

Stet can run against a local model through Ollama, keeping manuscript content on the machine. OpenAI is also supported where the client’s policy, configuration and contractual controls permit its use.

Retention, model training and data-handling characteristics therefore depend on the selected deployment—not on a universal product claim.

Ongoing development

Extending the workflow carefully.

A second-generation workspace is in development, built around folder-level reference files and a tool-calling document agent. Stet is now looking for its first production deployment: a live manuscript taken through to closeout with an agency partner.

Have a document workflow with similar constraints?

Stet is one example of designing around the document format, the reviewer and the organisation’s data policy together.

Discuss a document workflow