Deed extraction · Doma Technology
DeedIQ-NextGen
A new, updated version of the DeedIQ running in production: faster, cheaper, and more accurate, without losing what already worked.
My contribution
I led DeedIQ-NextGen, built on Azure Content Understanding with GPT-5.5, and the testing that cleared it for production.
The problem
DeedIQ was already running in production, so an updated version had to beat it on cost, speed, and accuracy without regressing on anything title teams rely on. Some documents from data providers also turn out not to be deeds at all, and extracting from them wastes time and creates false positives.
How it works
- Build extraction on Azure Content Understanding with GPT-5.5, packaged as a plugin with its own repo and contract.
- Run a full parity test against the production DeedIQ, then test on more than 2,000 documents before deploying.
- Add registered-land (Torrens) detection, so title teams know early when a Torrens process applies, and a Decisions API classifier that rejects non-deeds before extraction runs.
What changed
DeedIQ-NextGen was just deployed to production after testing on more than 2,000 documents with zero issues. It is built for a lower cost per document and faster turnaround, and filtering out non-deeds before extraction saves processing time and cost and means fewer false positives. Production numbers will follow once it has run for a while.