The challenge
Construction-defense cases run on discovery, and discovery runs on people reading pages.
- The volume is huge. A single production can reach hundreds of thousands of pages, some of them scans no search box can read, plus thousands of site photos.
- Keywords miss the point. Lawyers need ideas, not just words: a search for a “glass shoe” should also surface the design drawings and contractor queries about the railing.
- The data can’t leave. Client documents must stay inside the firm’s control, and every answer has to trace back to its page.
What we’re building
- Strategy: a tool shaped around how the firm’s lawyers actually review a production, not a generic search box.
- Party pull: every page that involves a party, including name variants, with its page reference ready to export.
- Concept search: finds pages by meaning. In a demo on fictional data, it found 414 relevant pages where an exact-word search found 32.
- Checked answers: every quote is checked against its source page before a lawyer sees it, and nothing leaves the system without a lawyer’s sign-off.
- Security by design: it runs in the firm’s own cloud account. Each case is sealed off with its own storage and key, and deleted cleanly when the case closes.
- Honest gaps: any page the tool cannot read is listed for human review, never dropped.
What will be measured
Before any client document is loaded, the system must pass security testing, including a test that one case cannot see another. Then it is measured on three things the firm cares about: time saved per production, cost per page, and search accuracy against pages the firm’s own team checked by hand.
What this means for you
If your most expensive people spend days on work a well-built tool could do in minutes, start with a Fast Start to prove the case in two weeks. A Custom Build delivers the tool inside your own systems, under your control.
Under the hood
- Status: designed and prototyped on fictional data (about 3,300 pages in the demo). Validation on real documents is the first milestone.
- AI: Claude through Amazon Bedrock, running inside the firm’s own AWS account in a US region.
- Pipeline: text reading for scanned pages, page-level indexing with page references, party and concept search, and photo search.
- Answer checks: a verifier that refuses any answer whose quote doesn’t match its retrieved page, with 45 test cases on fictional data. In the demo, 30 of 30 quotes passed.
- Isolation: one store and one encryption key per case, with tests that try to cross between cases.
- Deletion: case data deleted and its key destroyed at case close, with a deletion certificate.
- Infrastructure: defined as code, with security-sensitive changes approved by a second reviewer.