AI Trip Planner

Planning a trip means holding a dozen browser tabs in your head at once — flights, neighbourhoods, opening hours, what is actually walkable from where you are staying. The information exists, but assembling it into an hour-by-hour plan is tedious work that nobody enjoys.

The app collects four inputs — destination, number of days, budget tier, and who you are travelling with — and asks Gemini for a structured itinerary. Each suggested place is then resolved against the Google Places API so it carries a real photo, rating, and location rather than a model-generated description. Itineraries persist to Firebase so a plan can be revisited and shared by link.

Early versions asked the model for prose and parsed it, which broke constantly — the shape of the response changed between runs and a single malformed day took down the whole render. I moved to a strict JSON schema in the prompt, validated the parsed response before it ever reached React, and fell back to a partial itinerary when a single day failed validation rather than discarding the entire result. Treating the model as an unreliable API instead of a trusted one is what made the feature shippable.

A trip plan that took an evening of tab-juggling now takes about thirty seconds, and every recommendation is backed by real Places data instead of a plausible-sounding hallucination.