WanderLust: A travel agent that interviews you to suggest your next destination

WanderLust is an AI travel agent that interviews you through a short story, works out what kind of traveller you are, and picks one small, specific town or neighbourhood for you anywhere in the world. You get a 48-hour plan, a postcard, and a soundscape for the place it picks.

 

It's for anyone who wants their personality and preferences matched to a real, specific place.

 

The whole thing runs on AWS, built with Amplify Gen 2, so the backend is TypeScript in the same repo as the Vite app. Amazon Cognito hands every visitor anonymous guest credentials. AWS AppSync exposes two GraphQL queries, interview and predict, plus the usual CRUD for saved trips. Both queries hit one AWS Lambda function that is the agent itself. Saved journeys go into Amazon DynamoDB, the frontend is on AWS Amplify Hosting, and the LLM key lives in SSM Parameter Store as an Amplify secret.

 

The agent keeps a profile of ten traits. Some come from tourism research, like Plog's scale of how much someone craves the unfamiliar and the sensation-seeking literature. Others are practical: pace, comfort, food, how much you want nature over streets. Each trait has an estimate, a certainty, and the exact words from your answers that moved it.

 

The interview is adaptive. Before each scene, the Lambda finds the two traits it's least sure about and asks the model to write a scene that tells people apart on exactly those. Each of the three answers carries trait signals, so tapping one updates your profile on the spot. You can also type your own answer, and the model reads it and quotes it back as evidence. While you're reading a scene, the next scene for every option is already being written in the background, keeping the pace fast.

 

Each scene gets a random theme or twist, an overnight train, lost luggage, a dinner invitation from a shop owner, and guardrails keep every situation unique. The agent also reads light context the browser already knows: your time zone, language, local time, whether you're on a phone. A 2am session gets night scenes, and December in Nairobi gets treated differently from December in Oslo.

 

My new account's Amazon Bedrock quota for Nova models came in at zero tokens per day. So text generation runs through an OpenAI-compatible endpoint, OpenRouter and Groq. Bedrock still powers Nova Canvas postcards, with generated SVG art as a second option. I also built an offline engine, a hand-curated atlas of 28 micro-destinations and a bank of written scenes, ranked with the same maths, keeping the app running through model or network issues.

 

The one thing I care about is that it feels like it's actually listening. A live radar on the side changes shape as you answer, showing which traits the current scene is probing. The interview runs five to eight scenes, however many it takes to reach high confidence. The progress bar tracks that certainty score, and when it's confident enough it says "I've got a read on you" and lets you reveal or keep going.

The reveal shows its working: a Fit score and a separate Certainty score, a radar of you against the place, and the answers that drove the match. It also tells you where the place will stretch you.

How I know it worked: I played it as two opposite people, a planner who books the spa when it rains, and someone who orders whatever the locals are eating and says yes to a sunrise boat trip from a stranger. The two runs produced completely different picks, and each reveal could point to the exact answers behind it.