Employment Project
Jun 2025 - Present
ArchGlyphOCR
Dense Architectural Blueprint Callout & Vector OCR Pipeline
Organization: Attentive.AI
Key Result & Impact Benchmark
Boosted vector callout parsing accuracy by 7% while reducing annotation turnaround time to 35%.
Technical Overview
ArchGlyphOCR is a novel document AI architecture designed at Attentive.AI specifically for dense architectural blueprints and CAD construction plans. Traditional OCR solutions (such as Google DocAI and Gemini) struggle on large-format engineering drawings with rotated callouts, leader lines, and dense spatial annotations. ArchGlyphOCR combines a vector path pretraining framework (PDF → SVG tokenization) with a DEIMv2 (DINOv3 backbone) multi-class segmentation network to propose text regions and extract dimension callouts with 7% higher macro-F1 score.
Key Technical Highlights
- Outperformed Google DocAI and Gemini 3.1 Pro on dense architectural plan callout extraction benchmarks.
- Reduced total annotation time (TAT) for HVAC ducting plans to 35% of original baseline.
- Eliminated expensive third-party vision API costs through self-hosted on-prem deployment.
Technologies & Frameworks
Vector Path Tokenization
DEIMv2
DINOv3
Text Region Proposal
OpenCV
PyTorch