EN FR
Editorial illustration of a simplified architect plan and elevation with window openings scattered across the walls, one opening tucked near the roofline marked in terracotta

A window-and-door takeoff tool I'm building for a Quebec installer

A tool that reads an architect's PDF and produces a takeoff spreadsheet in minutes instead of an hour.

A residential installer who does exterior windows and doors needs a tool where he drops in an architect's PDF and gets back a takeoff spreadsheet — every window, every door, every spec — ready to price in the manufacturer's software. An hour of counting and spreadsheet work becomes a few minutes. I designed it around the installer's actual workflow.

Forty-plus openings on a big house, scattered across elevations and plan views. Get the dimensions wrong for one tucked under a roofline and the installer eats the cost.

Stylized architect plan view and south elevation with dozens of window and door openings rendered as gray rectangles; one amber rectangle marks the easy-to-miss window under the roofline

PDF in, takeoff spreadsheet out. An hour of work becomes a few minutes.

Four-stage horizontal flow showing an architect PDF becoming openings detected, then a spec table, then a takeoff spreadsheet highlighted in amber

The detector works well after several rounds of training on a large set of semi-synthetic data.

Left panel shows a grid of semi-synthetic training plans with labeled openings feeding into several training rounds; right panel shows a real architect plan where every opening is accurately detected after training

The checkout page is wired. The billing endpoint exists. I'm currently working on refining the detection model to increase visual recognition accuracy for this task.

Building this tool has given me deep expertise in image recognition for architectural takeoffs — how to train a visual recognizer to find openings in dense plans, how to handle the gap between clean synthetic drawings and messy real-world architect PDFs, and how to get the specs right on every unit. That expertise applies to any visual recognition task where you need structured data out of technical drawings.

Let's talk →

Related projects