Aperture PR-01: Sub-Millimeter Geometry Meets Vision-LLMs

Extracting structured data from messy enterprise documents has traditionally forced operations teams into an uncomfortable compromise: either rely on rigid regex-based OCR templates that shatter when a vendor alters their margin by 5 millimeters, or pass raw images to general-purpose LLMs that hallucinate totals and fail to provide auditable source coordinates.

With Aperture PR-01, ParseRelay bridges this divide.

The Problem with Brute-Force Vision Parsing

General-purpose vision models downsample input documents to standard patch grids (e.g., 768x768). On a dense 4-page international freight invoice or a 50-row customs declaration, this compression obliterates sub-millimeter character boundaries. Small decimal points vanish, and multi-column tables merge into ambiguous textual blobs.

More critically, general-purpose models cannot produce verified bounding boxes tied to native PDF vector paths. When an auditor asks why line item 14 was parsed as $14,200 instead of $1,420, an ungrounded LLM offers zero geometric proof.

The Aperture PR-01 Hybrid Architecture

Aperture PR-01 operates via a three-tier pipeline:

1. GPU Patch Slicing & Vector Geometry: High-resolution PDFs are rendered at 300 DPI on our dedicated GPU inference cluster. Embedded vector fonts are extracted directly from PDF streams, while raster areas pass through token-level geometry classifiers.
2. Normalized Coordinate Bounding Boxes: Every recognized scalar field and table cell returns exact [ymin, xmin, ymax, xmax] normalized coordinates (0 to 1000) relative to the source page canvas.
3. Structured Schema Grounding: Extracted tokens are mapped strictly against your defined 7 typed primitives (string, number, date, currency, table, key-value, boolean) without free-form hallucination.

// Sample Aperture PR-01 bounding box payload
{
  "field": "totalAmount",
  "value": 4312.50,
  "currency": "EUR",
  "confidence": 0.994,
  "geometry": {
    "page": 1,
    "boundingBox": [842, 650, 868, 792] // [ymin, xmin, ymax, xmax]
  }
}

99.4% Accuracy on Degraded Source Files

By combining deep token geometry with specialized vision-LLM attention, Aperture PR-01 achieves 99.4% field recall across degraded thermal receipts, skewed smartphone photographs, and multi-page carbon copies.

Most importantly, when confidence drops below your defined threshold (e.g. 85%), ParseRelay automatically routes the document into the human review workspace, highlighting the exact bounding box for 1-click verification.