AI Article Gen — Document-to-Publish Pipeline

Preparing an article for publication meant retyping eight derived fields by hand for every piece, then hoping the table-of-contents anchors lined up with the headings they were supposed to jump to. It was slow, it was the same work every time, and the anchors were a recurring source of broken pages.

Built a six-phase pipeline in Python + FastAPI. Intake is deliberately deterministic — mammoth converts the .docx to semantic HTML with no LLM involved — then editorial guidelines resolve from SQL, and three JSON-mode LLM calls handle structure, body and metadata for a short article. The body pass is chunked by character budget so a long research paper can't exhaust the context window. Render writes the DOCX, PDF and HTM artefacts.

• 01Editor's job goes from retyping eight derived fields to upload, review, save
• 02Table-of-contents anchors match the contents list by construction — both derive from the same document walk, so the model is never asked to keep them in sync
• 03Editorial rules live as SQL rows resolved exact-type → any-type default → built-in prompt, fail-open on a database fault, so guidelines change without a code deploy
• 04Three LLM calls for a short article, each retried and budget-counted