Attentive.AI

Role Overview & Impact

• Architected a vector path tokenization framework (PDF → SVG) as a pretraining objective for zero-shot symbol detection and vector path annotation, increasing SOTA macro-F1 by 7%.
• Trained and shipped DEIMv2 (DINOv3 backbone) for multi-class vector path segmentation and geometry dimension estimation, reducing total annotation time (TAT) for HVAC to 35% of original duration.
• Designed, built, and deployed ArchGlyphOCR: a dense plan text-region proposal + prediction pipeline outperforming Google DocAI & Gemini 3.1 Pro on callout extraction while eliminating API costs.