Scott Newman is an enterprise solutions architect who runs Devonport City Guide, a travel site for Tasmania's north west coast, on the side. Mersey is the travel agent he built for it, named after the local lighthouse, embedded on the site and given its own tab in the mobile apps. This is a recorded build session, so you get the dashboard, the tool configuration, and a live demo on the site alongside the walkthrough.

The value of it is that he leads with the failure. His first version pointed a web knowledge tool at the site and scraped it, and it broke in ways familiar to anyone who has tried that. So he rebuilt the foundation underneath it, siloed the data, and wrote custom tools that pre-process before they query. Watch this if you are past your first agent and hitting the limits of scraped content. His summary is the whole lesson: bad data in, bad output, no matter how good the model is.

What this video covers

• What broke in the first version, a crawler that had to be re-triggered by hand so the data went stale, an article about an event that was indistinguishable from the event itself, and no ability to reason about dates
• The rebuild, with places, events, guides, and reviews siloed into separate databases instead of one pile of scraped text
• Three custom tools, each calling a function that pre-processes before it queries, and how he writes their descriptions so the agent routes an open-ended question to the right one
• How the place tool injects the current date and time and answers whether somewhere is open, while the event tool resolves this weekend or in two weeks into an actual date range before it searches
• How he judges whether the agent is working, reading chat logs by hand and scoring accuracy with AI, because the view he wants does not exist yet
• His lessons from the build, start simple, expect to keep tuning the prompt, keep the data clean, and give the agent a personality people will actually click on

Chapters

• 0:00 Introductions
• 2:12 What Devonport City Guide is
• 4:04 Meeting Mersey, the travel agent
• 5:10 The first version, built on a crawled knowledge base
• 5:53 Where the scraped setup broke down
• 7:21 Rebuilding onto separate databases
• 8:04 Three custom tools and how the agent picks one
• 9:32 Answering what is happening this weekend
• 10:40 Live demo on the site
• 12:36 Writing tool descriptions the agent can route on
• 21:10 Keeping listings and events current
• 24:16 Lessons from the build
• 28:57 How he judges whether the agent is working
• 35:09 What building a custom tool was like