The Shovel

What it does

• Answers questions from your documents by retrieving from the project corpus, citing the source document with a confidence score.
• Extracts a structured Bill of Quantities from PDF, XLSX or CSV.
• Runs quantity take-off on drawings (PDF and DXF), including schedules and title blocks.
• Generates CPM-validated work breakdown structures — 200+ activities with early/late start and finish and the critical path.
• Reconciles drawing quantities against BOQ totals with exact variance maths and recommendations.
• Answers engineering questions from a curated construction knowledge base (concrete, buildings, roads, earthworks, geotechnics, procurement). Every answer is a cited rule with provenance and a credibility tier.

What it refuses to do

• Invent a figure that is not in your documents. Refusal is one of the end-to-end tests.
• Leak one project’s documents into another. Project isolation is tested on every change.
• Answer from memory when it could cite. Citations resolve to real ingested documents.

How it is built

FastAPI backend, React frontend, Postgres with pgvector, and hybrid retrieval (BM25 and vector search fused by reciprocal rank fusion, then re-ranked with a hard project-scope filter). An agent runtime routes each request to the right block: project Q&A, BOQ, drawing take-off, scheduling or cost.