Case study · AI agents on company data
Anchorsite
An AI pipeline that reads every document on a construction project, so an agent can answer questions across all of them.

At a glance
~20,000documents per project
$0.002average cost per document
$0 to $0.20range, by document type
The problem
A construction project produces emails, PDFs and technical drawings by the thousands. The answer to a question usually sits across several of them, and reading every document with a large model is too expensive.
What I built
- A data pipeline that processes about 20,000 documents per project, including emails, PDFs and technical drawings.
- A router that sends each document to the cheapest model tier that can read it, based on static analysis of its PDF structure. Some documents cost nothing; complex drawings cost up to $0.20.
- A structured memory built from the pipeline's output, so an AI agent can query across every document in a project through MCP.
Python · LLMs · MCP · document processing