6,000 PDFs that nobody could safely edit

Problem

A global pharma client’s standard operating procedures lived as PDFs on SharePoint across five regions. Updating one meant finding the file, editing it offline, and chasing approvals over email, with no reliable record of what changed, who approved it, or which version was current.

Approach

I led the four-person cross-functional team that built the platform, spanning generative AI, API work, data setup, and UI/UX.

We ingest the existing PDFs in parallel, parse them, and store them section by section so edits are surgical rather than whole-document. An author describes a change in plain English; the system rewrites only the affected sections and shows a redline preview (old text struck through, new text in red) before anything is committed.

An automated check screens the draft against basic compliance rules and returns it to the author if they aren’t met, then routes it to the correct approver for that region.

Outcome

Approved versions publish back to SharePoint, which is still the client’s system of record, with the version in the filename, while the platform retains version metadata and full change history. The design worked inside the client’s existing governance rather than asking a regulated organization to abandon it.

Technical detail

Section-level storage

Section-level storage is what makes plain-English editing tractable. The model operates on a bounded section rather than a forty-page document, which keeps rewrites reviewable and diffs meaningful.

Region handling

Region handling was scoped deliberately. Regulatory requirements and document templates were shared across the five regions, and only the approver routing differed. So that became the one region-aware piece, instead of five separate workflows to maintain.

Rejection paths

Rejection paths exist at two points: automated compliance screening and human approval. Both return to the author with the redline intact, so nothing has to be reconstructed from scratch after a rejection.