Pharma PDF → Word Extraction & Audit Traceability Platform

Overview

An end-to-end Python platform that extracts structured data from regulatory PDFs — CSRs, AE summaries, CoAs, stability reports, WHO monographs — and populates Word templates with inline source citations. Every extracted value is tagged with its source filename, page number, and section header. The entire data journey is reproducible, auditable, and GMP-ready.

System Architecture — End-to-End Pipeline

Each stage produces verifiable artefacts (provenance records, audit JSONL, populated .docx) so the entire data journey is reproducible and inspectable.

flowchart LR
    A["[1] Ingest\nSHA-256 hash\nSession ID\nPage type detect"] --> B["[2] Multi-Engine Extract\n10 engines in parallel\npdfplumber · PyMuPDF\nCamelot · Tesseract · Claude AI"]
    B --> C["[3] Provenance Record\nkey · raw_value · page\nsection · bbox · confidence\nengine · source_doc_id"]
    C --> D["[4] Field Mapping\n90+ regex rules\nClaude AI Auto-Map\n match"]
    D --> E["[5] Word Population\npython-docx\nInline citations\nformat preserved"]
    E --> F["[6] Audit & Package\nSHA-256 hash-chain\nJSONL audit log\nZIP artefacts"]
    style B fill:#0E7490,color:#fff,stroke:#0E7490
    style C fill:#065f46,color:#fff,stroke:#065f46
    style F fill:#7c3aed,color:#fff,stroke:#7c3aed

Stage Details

Provenance Record Schema

Every single extracted value is a structured, traceable record:

ProvenanceRecord {
  field_id         : UUID    # unique per extraction
  key              : str     # field name (e.g. 'Total Impurities')
  raw_value        : str     # extracted text exactly as in PDF
  normalized_value : str | None  # cleaned / unit-converted value
  page_number      : int     # 1-based PDF page
  section_header   : str     # nearest heading above the value
  bounding_box     : {x0,y0,x1,y1}  # PDF-point coords for highlighting
  confidence_score : float   # 0.0 – 1.0
  extraction_engine: str     # pdfplumber | camelot | claude | ...
  source_doc_id    : str     # SHA-256 of source PDF
  record_type      : str     # kv | table | acroform | llm | ocr
}

Extraction Quality — Citation Examples

Certificate of Analysis

Stability Report

Before & After — Word Template Population

Batch Number:      
Assay Result:      
Total Impurities:  
Retest Date:       
Approved By:       

Batch Number:     AXV-2024-0042 [Src: CoA_Batch2024.pdf, p.1]
Assay Result:     98.7 % [Src: CoA_Batch2024.pdf, p.3, Sec: 2.1]
Total Impurities: 0.42 % [Src: CoA_Batch2024.pdf, p.3, Sec: 2.2]
Retest Date:      2026-08 [Src: Stability_AXV101.pdf, p.2]
Approved By:      Dr. J. Smith [Src: CoA_Batch2024.pdf, p.1]

Technology Stack

What This Delivers

Why This Approach

• Already built — not a prototype. A deployed FastAPI + React platform with 10 extraction engines, tested on real regulatory documents.
• Pharma domain knowledge — understands ICH E3, 21 CFR Part 11, CoA structure, analytical report sections, and GMP audit requirements.
• Full citation chain — every value carries filename, page number, and section header out of the box.
• Handles edge cases — scanned PDFs (OCR), AcroForms, two-column layouts, bordered tables, and hybrid documents.
• LLM-augmented — Claude and GPT-4o fill gaps where regex/heuristic extraction fails, with confidence scoring on every result.
• Clean, tested Python — type-annotated, documented, and structured for handoff.

Get In Touch

Interested in this system for your regulatory document workflow, or have a similar problem?

Email me LinkedIn