Deterministic pricing plus an AI scoring layer that quotes page referenced evidence, calibrated against a real evaluation panel.
Tender evaluation meant reading hundreds of pages per bidder.
Bills of quantities arrived in inconsistent formats and had to be priced by hand. Scoring criteria were worked through by hand too.
The process was slow, inconsistent between evaluators, and hard to audit after the fact.
A two stage scorer, with a hard line between the stages.
Stage one is code. It prices every bidder it can from the bills of quantities, and states honestly why the others cannot be priced. No silent failures, no guessed prices.
Stage two is an AI layer. It scores each criterion, and it must quote the exact passage and page number behind every score. Code verifies every quote against the source documents before anything is shown to a human.
Four rules keep the scorer honest.
The scorer was calibrated against a real evaluation panel, on a real tender with five bidders.
The AI layer's total matched the panel's score exactly for one bidder, and came within a point for another.
It also surfaced evidence the panel had not considered. That evidence was presented as additional consideration, never as a panel error.
Model spend for a full run: under a dollar.