HHD Real Estate AI Brief

Handoff-H1 scores 81.6% composite on the Construction Blueprint Takeoff Benchmark, above independent estimators at 77.6%.

Benchmark covers 1,348 primary-tier materials across 10 real residential blueprint sets with 2,009 verified line items.

Seven frontier and open-weight models tested on the same benchmark span composites of only 35–61.

Paper draws on two interdisciplinary civil-engineering and computer-science research projects targeting building-operation optimisation.

Buildings differ from digital environments: a bad control decision wastes energy irreversibly, violates occupant comfort, or accelerates equipment wear.

Authors propose a foundation for software engineering for AI (SE4AI) in systems where failure carries physical consequences.

Mask R-CNN base model achieves bbox mean average precision (mAP) of 0.7596 and segmentation mAP of 0.7111.

At IoU threshold 50, bbox mAP reaches 0.9850; at threshold 75, segmentation mAP reaches 0.9219.

Model trained on COCO-format annotated floor plan images; no commercial deployment named.