Document Automation

JENNY converts draft SOPs into properly formatted federal compliant documents. An LLM extracts structured content from the draft, then a deterministic Python pipeline handles all template mutation with a structural integrity gate. The pipeline is model-agnostic -- any LLM that produces a valid config can drive it.

Upload       User uploads the template (.docx) and source draft (.docx or .pdf)
Extract      LLM reads the draft and produces a structured config
Review       User verifies and edits the config in the browser UI
Generate     Pipeline mutates the template, validates output, produces .docx

Two extraction paths:

• Keyed build (JENNY_keyed.exe): Anthropic API key embedded -- extraction runs automatically
• Keyless build (JENNY.exe): User copies the extraction prompt into any LLM (ChatGPT, Gemini, Copilot, etc.) and pastes the output back

See QUICKSTART.md for end-user instructions.

From the source draft, JENNY extracts or derives:

LLM-derived sections (Scope, Roles, Materials, Guidelines, Hierarchy) require SOP owner review.

JENNY ships as a single .exe -- no Python, Node.js, or environment setup required.

Place the .exe in its own folder and double-click. A browser window opens at http://localhost:5000. The app creates uploads/ and jobs/ folders next to the exe at runtime.

See Efficiency by Design: Reducing AI Cost Through Smart Architecture for a detailed look at how JENNY's split-phase architecture achieves a 15-25x cost reduction over LLM-heavy approaches by limiting the model to text comprehension and handling all document construction in code.

Current prototype validates against 82-84 checks and scores 100% on tested SOPs:

• Full extraction-to-docx pipeline for both .docx and .pdf drafts
• Image extraction, positioning, and embedding (PDF and DOCX)
• Hyperlink preservation
• Highlight detection from .docx and application to config
• Cover page uses full title, all other locations use short title
• External LLM flow: copy prompt with draft text + image markers, paste config back
• Config review UI with sequential numbering, section labels, editable fields
• Validation gate with score reporting
• Single-file executables (keyed and keyless)

• PDF text extraction depends on PyMuPDF -- scanned/image-only PDFs not supported
• Multi-procedure SOPs (structure_type: "multi") not yet implemented
• External LLM cannot detect highlights -- JENNY splices them from the uploaded draft
• Step hierarchy is an LLM judgment call and may need manual correction

Backend:

pip install -r requirements.txt
set ANTHROPIC_API_KEY=sk-ant-...
cd backend
python jenny_backend.py

Frontend:

cd frontend
npm install
npm run dev

Build executables:

pip install pyinstaller
python build.py

JENNY/
  README.md
  QUICKSTART.md                         End-user guide
  requirements.txt
  build.py                              Build script for executables
  jenny.spec                            PyInstaller spec
  .env                                  API key (not committed)
  backend/
    jenny_backend.py                    Flask API + session management
    jenny_pipeline.py                   Deterministic template mutation
    unpack_docx.py                      Unpack .docx to XML
    pack_docx.py                        Pack XML to .docx
    JENNY_Phase0_Extraction_Prompt.md   Extraction prompt (single source of truth)
  frontend/
    src/App.jsx                         React UI
    index.html
    package.json
  data/                                 Test drafts and templates
  dist/                                 Built executables (not committed)

• No exec() -- all config parsing uses ast.literal_eval
• Debug mode disabled by default
• 50 MB upload size limit
• Server-side error logging only
• Path traversal protection on image serving endpoint