Nova Travel | Devpost

Inspiration

Planning a trip is exhausting. You open five tabs, compare flights on one site, hotels on another, search for things to do on a third, and by the time you've figured it out you're drained before you've even left. We wanted to build something that felt like having a personal travel agent in your pocket — one you could just talk to.

What it does

Nova Travel is a voice-first AI travel planner powered entirely by Amazon Nova. You speak your destination, dates, and budget — and within seconds the app generates a complete, personalized day-by-day itinerary including flights, hotels, real local events, restaurants, and free attractions like parks and beaches.

The experience is fully conversational. If you don't like something — "I don't eat seafood" or "I want to fly at night" — just say it. Nova Travel updates only the affected parts of your plan instantly, remembers your preferences for the rest of the session, and reads the updated itinerary back to you via voice. When you're ready, it books your flight and hotel automatically.

How we built it

The backend is built in Python with FastAPI. The orchestrator agent uses Nova 2 Lite to parse voice transcripts into structured trip intents and classify mid-session feedback into one of five types: preference update, schedule change, budget shift, swap request, or full restart. Based on the feedback type, only the affected specialist agents re-run — keeping the experience fast.

Four specialist agents (flight, hotel, events, food) run in parallel using asyncio, pulling from Ticketmaster and Google Places APIs simultaneously. The itinerary agent then uses Nova 2 Lite to synthesize all outputs into coherent day-by-day plans at both budget and premium tiers. Nova 2 Sonic handles voice transcription and text-to-speech readback. Nova Act automates the actual booking flow on airline and hotel websites. The frontend is built in React with Vite and Tailwind CSS, with both frontend and backend deployed on Railway.

Challenges we ran into

Getting all four specialist agents to run truly in parallel while maintaining session state was the core technical challenge. We built a thread-safe session store that persists user preferences, conversation history, dirty agent flags, and itinerary state across the full conversation loop. The feedback loop was particularly tricky — classifying natural language corrections into actionable agent re-runs required careful prompt engineering with Nova 2 Lite. On the infrastructure side, deploying a multi-module FastAPI backend required careful path management and a clean mock/live toggle system that let us build and test everything locally before AWS credentials were finalized.

Accomplishments that we're proud of

The feedback loop is what we're most proud of. Saying "I don't like seafood" and watching only the food agent re-run — with flights, hotels, and events staying cached — while preferences persist for the rest of the session feels genuinely intelligent. It's not a chatbot. It's a system that reasons about what changed and responds proportionally. We're also proud of the parallel agent architecture — four external API calls completing simultaneously gives the app a responsiveness that feels closer to magic than software.

What we learned

Building multi-agent systems requires thinking carefully about state. Every agent needs to read from the same source of truth, and feedback needs to flow back into that source cleanly. We also learned that a mock/live toggle pattern — where every agent has a USE_MOCK flag — is essential for hackathon development. It let us build, test, and demo the full pipeline before a single real API key was confirmed, which saved us hours on deadline day.