For businesses across the US · Team based in Dallas, Texas · On-site visits nationwide

Home / Work / AI document intake

Legal tech · United States

A case assistant that turns lien paperwork into a draft case, with every value traced to its page

Law firms on the platform used to read contracts, invoices and change orders and type the details into a new case by hand. Now they describe the job to an assistant or drop in the documents, and review a draft instead.

Client
US construction-lien SaaS (name private)
Our role
Case assistant, extraction engine, review screens, AWS deployment
Engagement
Dedicated team
The product

Real screens, client details replaced

Start from a sentence or the documents. The user types what the job is about in plain words, or drops in the contract, invoices and change orders.
Step 01

Start from a sentence or the documents

The user types what the job is about in plain words, or drops in the contract, invoices and change orders.

Real product screens. Names, addresses and amounts replaced with sample data.

01 · The problem

Accuracy mattered more than speed

Each new lien case starts with a bundle of documents: a contract, a few invoices, maybe a scanned change order. Someone had to find the property owner, the contractor, addresses, amounts and the months of work, then enter them correctly, because a wrong date or name in a lien filing has real consequences. The platform wanted AI to prepare that first draft, but it could not be allowed to make things up.

02 · How the workflow runs

Rules first, AI only where needed

  1. The user describes the job in plain words or drops in documents
  2. Up to 10 files per upload: PDF, Word, Excel, images, or pasted notes
  3. Digital PDFs are read as text; scanned pages go through OCR
  4. Rule-based extractors fill every field they reliably can
  5. An LLM is called only for what the rules leave open
  6. Each quote the model returns is checked against the source text
  7. A review screen shows every value with its confidence, page and quote
  8. A person confirms the case before anything is saved
  9. Legal deadlines are calculated by the application, never by the AI
03 · What we did

Our part of the project

  • Built the case assistant users talk to when opening a case
  • Rebuilt the extraction engine and the review screens
  • Replaced a separate Python service with an extraction engine running inside the Java application, behind an interface the screens already used, so nothing changed for users
  • Set up the deployment on AWS
04 · Measured during development
~$0.006AI cost per 3-page document, measured during development
≤ 1AI call per upload
0AI calls when the rules fill every required field

Measured during development. The client is kept anonymous until they agree to be named.

Technology
Java 21Spring Boot 3.5React 19MySQLTesseract OCRLLM APIAWS

Have a document-heavy process like this?

We’ll look at a few of your real documents on the free assessment and tell you what AI can and can’t take off your team’s plate.

Get a free AI assessmentSee pricing