An auditable AI workflow that extracts bar-bending schedules from structural drawings while preserving source evidence and engineer review at the highest-risk stages.
Extracting a bar bending schedule from a structural drawing set is a full day of manual counting, and the output feeds straight into procurement and pouring. An error is expensive and effectively permanent. A naive OCR pass does not work: the values only make sense in the context of the region of the drawing they sit in.
I built a multi-stage pipeline that starts with page overviews, detects regions with computer vision, then inspects detail and context crops. Every extracted value links back to its source, while typed review gates let engineers correct decomposition and classification without restarting a run.
In an internal evaluation across 40 drawing sets, reduced an approximately eight-hour workflow to under 30 minutes at 94% field-level accuracy while preserving engineer oversight.
My Role: AI Engineer