Cracked City

Overview

A road-damage triage tool for the City of Toronto that ranks every open street defect by what it costs to leave broken, and verifies each report with computer vision. Built in three days, on public data only. It won first place.

Challenge

Three days, and one hard rule: public APIs only, nothing private, nothing behind a contract. A city already knows where its roads are failing. The reports exist. They just arrive as a flat queue with no way to tell which crack becomes a pothole in six weeks and which becomes a claim. The problem was never collecting more data, it was ordering what was already there.

Approach

We trained our own crack-detection model on NVIDIA hardware, a desktop-sized box on the table beside us, and hosted it on Hugging Face so the app could call it live. Around that I designed one screen that serves two very different people: a 3D map any resident can read, and a ranked dispatcher queue that puts a dollar-per-day cost of delay against every ticket. My job was to make three days of machine learning feel like something you would actually open.