The category runs on documents, but operates on judgment
Pre-construction lives in PDFs, scans, and emailed addenda. Bid tabs come back as flat tables of line items and dollars. Historical pricing exists, but as exports nobody trusts. The work product of estimating — the actual thinking about a bid — is captured in spreadsheets named after the project, the estimator, and the date.
None of this is broken in a way operators don't already know. It's broken in the way things that have always worked are broken: everyone has adapted to the friction, so the friction stops being visible. New tools layer workflow on top of the friction. Almost none of them go underneath it.
What “intelligence layer” actually means
We use the phrase carefully. It is not a dashboard. It is not a chatbot bolted onto the same data. It is the part of the stack that:
- Reads the documents — bid tabs, addenda, historical scans — at scale, accurately.
- Normalizes the entities — contractors, projects, lettings, line items — so the same thing in two places is one record.
- Connects the patterns across sources, time, and geography.
- Answers the questions estimators actually have, against a typed ontology of what the category is made of.
Workflow tools assume the data is already legible. We're starting from the assumption that it isn't. The hard part isn't the UI on top — it's the layer underneath that makes any UI worth building.
Why now
Three things changed at once. Document understanding stopped being a research problem and became a procurement decision. Vector search and structured extraction are commodities. And the cost of running the work is now low enough that a vertical-and-deep system is buildable by a small team against public data, before a single contract is signed.
The window is to build the answer in the open, against real public sources, before the next dashboard-with-AI-bolted-on raises another round and crowds the conversation.
What we're not
We're not a CRM. We're not a procurement workflow. We're not a bid scorer. Each of those products has buyers and merit. None of them is what's missing.
We're also not building “AI for construction” in the generic sense. The category already has surface-level AI overlays. The deep work is vertical: knowing what a unit-price line actually means, knowing how a particular DOT writes its specs, knowing why a contractor won three jobs in a row and lost the next two.
What we're building first
Public bid data is the obvious starting wedge. It's real, it's messy, it's cross-state, and it rewards entity resolution and pattern recognition immediately. We're ingesting state DOT bid tabs, resolving entities across sources, and surfacing the answers most estimators currently get by phone calls and gut feel.
We publish what we ship at /changelog. If what's shipping looks like it solves a problem you have, the waitlist is the door.
Cassandra of Troy: the prophet whose true predictions were dismissed. The brand promise is that we see what the documents are already saying — and surface it before the bid is locked in.