DAST: Bridging Specialized and General-Purpose AI for 20,000 Pages of Luxembourg Notarial Archives

What happens when Transkribus meets Claude API? The DAST project chains specialized HTR (under 3% CER) with large language model intelligence to process 20,000 pages of Luxembourg notarial archives — extracting structured data, summarising in three languages, and linking 52,000+ person mentions across sources. The key finding: specialised and general-purpose AI are stronger together than either is alone.