DUWAMISH RIVER — Craig Campbell launched a website called Past Maps in 2022 after turning down venture capital offers and selling his previous e-commerce tool for Shopify businesses. The site, which overlays historical maps onto modern-day maps with adjustable opacity, grew from 20,000 to over 300,000 monthly active users by its third year.

Campbell, a former Meta engineer, initially created the mapping tools to support his metal detecting hobby by locating old structures and trails. He shared the tool on Reddit with fellow enthusiasts, sparking early interest that evolved into broader demand. Past Maps now serves users engaged in genealogy research, historical exploration, and mapping old oil wells, drawing its map data from publicly available sources like the U.S. Geological Survey.

Google Search is the site’s primary traffic source. Campbell optimized Past Maps’ webpages and map metadata for search visibility, a move he said led to steady growth. “As I started exploding out this data and making it finally available to Google and giving it a place on the web, traffic just started to build.” He described the project as a return to foundational internet principles: “This is how the web is supposed to work. This is actually the old school web.”

The site operates on a freemium model, offering limited access for free and full features for $9 per week or $52 annually. Revenue supports Campbell and his wife, who assists with the business. To manage customer service efficiently, Campbell runs a local AI agent on his desktop that processes emails hourly when his laptop is on, filtering spam, flagging urgent messages, and drafting replies. This system has reduced his daily customer service time from one to two hours to about 10 minutes.

Campbell is also developing an AI-powered OCR tool to extract text from historical maps, which often feature curved labels, inconsistent spacing, and overlapping content that challenge conventional optical character recognition systems. Off-the-shelf OCR tools failed to parse historical maps effectively, but modern large language models showed more success when combined with human input.