Travel Buddy

Travel Buddy is an Android app for planning trips. You enter a destination, your dates and a budget, and the app suggests an itinerary, places to eat and somewhere to stay, with an estimate of the costs. Trips you have made are kept in a history, you can save favourites, and the home screen shows trips that other users created recently.

My part

We were a team of four developers. I designed and built the back end and connected the app to the AI model on Google Cloud Vertex AI, which generates the trip suggestions. I also worked on parts of the front end and on testing. Places and photos come from the Google Places API, and the app uses Firebase and MongoDB for its data.

Working with the model

The model’s answers had to come back in a structure the app could use, and every call costs money. For the structure, I gave the model a JSON schema for its response, validated each response in the back end before using it, and asked again when a response was unusable. For cost, I cached results so the same request did not reach the model twice, chose a smaller model that was good enough for the task, and limited when the model was called at all.

What I learned

A language model is an unreliable data source until you treat it like one: define the format you expect, check what comes back and plan for the answers you cannot use. Cost is a design constraint from the start, not something to optimise afterwards.