Inspiration: Sparked by a messy vendor-comparison spreadsheet during a PM internship — a five-minute decision that took four days because nobody's system understood what the company actually valued (budget, risk, past lessons). Also noticed procurement is the least "glamorous" corner of construction tech, so almost nobody's building real AI for it. What we learned: The hard part wasn't the AI reading documents — it was realizing "useful" differs by company, which is what led to the custom policy engine. Also learned timelines shouldn't be flat averages of past projects; better to weight past projects by similarity to the current one: How we built it: Four layers on one shared data model — document intelligence (parses contracts/POs/invoices), the policy engine escalating past a threshold, a live supply-chain tracking layer, and an AI-summarized communication layer. Built the policy engine and parsing first since it was the riskiest assumption. Challenges: Tuning the policy engine so it was specific without being rigid; handling messy/noisy historical timeline data without faking false confidence; and scoping down a wide idea to what actually strengthens the policy engine, cutting features that didn't.

Built With

Share this project:

Updates

Submission history