← Back to selected work

Document intelligence · Doma Technology

DeedIQ

Making a manual deed-analysis workflow dependable enough for production.

My contribution

I architected the system and led its delivery within Doma's applied-AI team.

99%+accuracy in production

The problem

Deed analysis combines inconsistent documents, detailed extraction, and decisions that need to be checked. A useful system has to make its mistakes visible and give reviewers a way to correct them.

How it works

  1. Use LLM agents and neural search to extract the relevant information from a deed.
  2. Run independent extraction teams and cross-validate their results rather than trusting one pass.
  3. Keep human review in the workflow and feed corrections back into a continuously learning knowledge base.

What changed

The published portfolio results are 99%+ production accuracy and a reduction in cost per deed from $12 to $0.20.

Talk about this projectExplore the other work →