Document Intelligence Prototype
Answers plain English questions about a property's documents with the source page shown next to each answer, and checks a draft subcontract's scope against a required scope checklist to catch gaps early. Built for a client pitch and run on synthetic documents.
Period
- Jun 2026
Role
- Sole engineer on a client prototype
Built for
- BlueSteel AI Solutions
Type
- Runs locally
Tools
The idea
A real estate developer deals with a stack of documents per property: contracts, addenda, specs. Two questions come up constantly. Where does it say that? And does this subcontract actually cover everything it needs to? This was a quick prototype for a pitch that shows both.
How it works
There's no vector store. The documents are small enough that each page goes into the prompt whole, labeled with its page number, and the model has to cite the title, page and exact sentence it used, or say plainly that the answer isn't there. The citation renders next to the answer so you can check it.
The scope check is mostly not AI. A required scope checklist of CSI MasterFormat codes is the answer key. The model maps what the draft subcontract claims onto those codes, then plain set math finds what's missing, what two trades both claim, and what caused change orders before and is missing again. Each flag gets a keyword check and a model second opinion before it's shown, and the model never makes the final missing scope call. A draft tab also builds a .docx subcontract from a template and leaves visible placeholders where no clause fits.
Status
It was built in one sitting against a deterministic stand in for the model, with 38 tests covering the parsing and the gap math. It never ran against the real model or real documents, and it didn't go further than the pitch. The client's name stays off this page.