Five deterministic checks over a live short-term rental, ranked by severity and delivered to Telegram every morning. What it took to make the alert worth reading, including two bugs it shipped while every test stayed green.
I connected the Hospitable MCP server to Claude and pointed it at my real Airbnb. Two weeks of letting an agent actually look turned up stale door codes, missed conversions, a closing review window, and a year of complaints it clustered into three themes in seconds.
A scheduled pipeline that sends the same soccer-news prompt to four frontier models and shows their outputs side by side, with live voting. What I built, what I learned from running it, and the AWS deployment debugging in between.