Sep 15, 2026
v1.4
Aperture PR-01 Hybrid Vision Engine
Shipped our flagship fine-tuned vision-LLM engine combining GPU high-resolution patch slicing with deep OCR token geometry for sub-millimeter bounding box precision.
New
Aperture PR-01 model architecture deployed to production GPU cluster with <1.2s median inference.
New
Sub-millimeter [ymin, xmin, ymax, xmax] normalized bounding boxes returned with every scalar and table cell.
Improved
Token recall on degraded thermal receipts and skewed carbon-copy bills increased to 99.4%.