SHED is a local web tool for extracting structured data from construction documents (Delivery Notes, Invoices, Certificates, and more) using AI vision models. Upload a scanned PDF or image, run the extraction command, and get all fields populated as structured JSON — ready to review, edit, and compare side-by-side with the original document. Field schemas are configurable per doc type in the FIELDS tab; a 57-field Delivery Note schema is included as the default.

Each document passes through a preprocessing pipeline before any AI is involved, improving accuracy and minimising token cost.

• 01
  Upload

  Drop a PDF, JPG, or PNG onto the left panel. The file is saved to uploads/ and appears in the document list immediately. PDFs are rasterised at 200 dpi before processing.
• 02
  Smart PDF Pipeline PDFs only

  Before image enhancement, multi-page PDFs go through two automatic checks:
  
  Boundary detection — a sample of page images is sent to the configured prescreen model (same provider as the main prescreen) with a prompt asking it to identify where one document ends and another begins. If boundaries are returned, the PDF is split into _part1.pdf, _part2.pdf, etc. Each part is tagged auto-split, gets meta.split_from set to the original stem, and a note recording its page range. Each part then continues through the pipeline independently. The original PDF is deleted. If detection fails for any reason, the file is treated as a single document (silent fallback).
  
  Page-limit chunking — if a document exceeds the active provider's page limit (Claude API: 20, OpenAI: 10, Gemini: 16, Mistral: 8; CLI: unlimited), it is split into ≤limit-page chunks. Chunks are written to a system temp directory — never to uploads/ — so they are invisible to the browser and batch runner. Each chunk is extracted independently, then the results are merged: top-level fields come from the highest-confidence chunk, materials are concatenated in page order, costs are summed. The merged document is tagged auto-chunked with a note listing the segment page ranges.

  Gemini provider exception: When the active provider is Gemini, PDFs are passed natively as application/pdf bytes directly to the Gemini API — rasterisation, image enhancement, and OCR are all skipped. Gemini handles text reading internally with no page limit in this mode.

• 03
  Image Enhancement

  Skipped for Gemini when processing PDFs natively.

  The document is deskewed (rotation corrected), converted to grayscale, then noise-checked: a median filter is applied only if the image is detected as noisy (RMS pixel deviation from local neighbourhood > 4.0) — this avoids blurring fine text on clean scans while still removing salt-and-pepper artefacts from poor-quality ones. Adaptive contrast is then applied (histogram stretch to full range), followed by sharpen ×2. These steps improve both OCR and AI vision accuracy across the full quality range.
• 04
  Resolution Cap (1200px) — for Claude

  The enhanced image is capped at 1200px on its longest side before being saved for Claude. A 200 dpi A4 scan goes from ~1654×2339px to ~848×1200px. Claude's vision model reads text clearly at this resolution, and image tokens scale with pixel count — this saves ~60% of image tokens with no quality loss.
• 05
  OCR + prescreen — in parallel

  Once the enhanced image is ready, OCR and prescreen run simultaneously. The OCR path depends on the document type:

  • Digital PDFs (text layer detected by pdfplumber — avg ≥ 10 words/page): text is read directly from the PDF layer, Tesseract is skipped entirely. OCR hint limit: 2000 chars. Faster and more accurate for computer-generated documents.
  • Scanned PDFs and images: Tesseract runs on a 2× in-memory upscale (dan+eng; classical OCR needs pixel density). OCR hint limit: 800 chars, covering the most useful content near the top of the document.

  In parallel, the prescreen model classifies page 1 (document type, confidence, reason, supplier guess). Running both paths in parallel saves ~8–15 seconds per document. The prescreen model is configurable in ⚙ Settings — default: Claude Haiku via CLI.
• 06
  AI Extraction — prescreen + main model

  First, the prescreen model classifies the document type (Delivery Note, Invoice, Certificate, etc.) and guesses the supplier name. The matching field schema from the FIELDS tab is then loaded for that doc type — falling back to the built-in 57-field Delivery Note schema if none is configured. The extraction model (vision) is called with the capped image, OCR hint, and resolved schema. It returns a single JSON object with all fields. Each found field is an inline object {"value": "...", "conf": 0.97} — confidence baked in, no separate block. Not-found fields are null. No explanation, no markdown — JSON only. Both the prescreen and extraction models are independently configurable in ⚙ Settings.
• 07
  Post-processing — Danish Numbers, Meta Stats & Webhook

  Numeric fields (quantity, weights, volumes) are normalised deterministically in Python regardless of what format the model returned. Danish convention: comma = decimal separator, period = thousands separator. "40,000" → 40.0, "1.200" → 1200.0. Applied after extraction so it never depends on the model getting it right. The script also computes fill_rate (fraction of fields with a value) and avg_confidence (mean conf across all filled fields) and injects them into meta. Finally, a webhook notification is POSTed to the configured URL on extraction complete (configured in ⚙ Settings).

The extraction model and prescreen model are independently configurable in ⚙ Settings. Each slot can use a different provider — useful for mixing a free-tier Gemini Flash prescreen with a Claude CLI extraction, for example.

Image tokens dominate the cost per extraction. Four targeted optimisations bring the total down by roughly 70–75% compared to a naive implementation, with no loss in extraction quality.

After extraction, every field can be reviewed and annotated directly in the browser without touching the JSON file. Documents move through workflow phases as they are reviewed and approved.

The ANALYTICS tab gives a live overview of extraction quality across all loaded documents. All metrics are computed client-side from the JSON already in memory — no extra server calls. Four sub-tabs:

• OVERVIEW — Workflow counts (pending/extracted/in review/approved), quality summary bars (fill rate, confidence, accuracy), document type breakdown, document format distribution (PDF/JPEG/PNG), and a "Needs Attention" list of documents with low confidence or annotated errors.
• PROVIDERS — Side-by-side comparison of all AI models used, grouped by provider (Anthropic, Google, OpenAI, Mistral). Shows avg fill rate, confidence, accuracy, and cost per document. Trend arrows (↑↓→) show whether recent extractions are improving vs. the previous batch.
• FIELDS — Per-field quality table (fill rate, avg confidence, error count) grouped by schema section. Filter by document type and sort by worst performers.
• INSIGHTS — Cross-dimensional heatmaps: doc type × provider and supplier × provider, with a metric toggle (fill rate / accuracy / confidence). Also shows field performance per provider — highlighting fields where one model struggles compared to others.

CSV exports (docs and materials) have moved to the TOOLS tab.

• —
  Flask (Python)

  Lightweight web server. Serves the UI, handles file uploads, and exposes REST endpoints for document listing, extraction retrieval, and saving edits.
• —
  Pillow + deskew + NumPy

  Image preprocessing pipeline: rotation detection and correction (deskew), contrast/sharpness enhancement, Lanczos upscaling.
• —
  Tesseract OCR (pytesseract)

  Open-source OCR engine with Danish (dan) and English (eng) language packs. Provides character-level hints to the AI model.
• —
  pdf2image + Poppler

  Converts PDF pages to high-resolution images (200 dpi) before the preprocessing pipeline runs.