HHD Real Estate AI Brief

DrawingVQA covers 33 'Issued for Construction' drawings and 92 expert-curated question-answer pairs.

Benchmark tests three reasoning depths: perceptual, contextual, and domain-expert.

Evaluated models show largest performance gaps at the domain-expert reasoning depth.

Study tested 484 PDFs across four university buildings, covering 153 credit-level decisions.

A 4-billion-parameter model (gemma3:4b) outperformed an 8-billion-parameter model (llama3.1:8b) on text-only verification.

Adding low-resolution drawing images consistently reduced accuracy across all tested configurations.

AI firms' total footprint in Seattle and Eastside reaches 855,000 sq ft — double the 2024 level.

Databricks signed the quarter's largest deal: 142,000 sq ft at Four106 in downtown Bellevue.

Region posted 372,000 sq ft of positive net absorption in Q2, ending a multi-quarter reversal.