Skip to content
All projects

2026

PixelTrip

AI group trip planner: an 8-bit party quest guided by a multi-agent travel guide

View the repository (opens in new tab)Open the live demo (opens in new tab)
AWS Kiro BuildFest 2026
Winner
Vercel + AWS
Live
Developers
3
  • Next.js
  • TypeScript
  • Tailwind CSS
  • Supabase
  • FastAPI
  • LangGraph
  • Gemini
  • Docker
  • AWS EC2

Summary

PixelTrip is a real-time collaborative travel planner for groups of friends. Everyone joins a shared room with a guest name, builds a pixel character that carries their travel preferences, and plans the trip together through a chat with an AI guide: discovering destinations, voting, searching flights, generating an itinerary, and negotiating changes until the party approves.

My role

Led backend development in a team of three: designed the PostgreSQL schemas, built the FastAPI endpoints, and built the agentic AI with LangGraph and Gemini.

Problem

Planning a trip with friends turns into a negotiation spiral: budgets differ, interests clash, and dates never quite overlap. Existing tools handle flights or itineraries on their own, but none of them mediate the group.

Approach

A Next.js front end on Vercel talks to the backend only through typed service contracts, so the UI could be built against mocks while the backend was still in progress. Supabase Realtime keeps every member in sync for presence, chat and votes. A FastAPI backend, running in Docker on a single AWS EC2 server behind Caddy, hands each turn to a LangGraph orchestrator that routes to specialist nodes for discovery, logistics, itinerary and negotiation. Those nodes call Gemini and live flight, places and weather tools.

Party2 to 6 friends, guest namesNext.js appVercel, App RouterSupabase RealtimePresence, chat, votesFastAPIDocker on EC2, Caddy HTTPSSupabase PostgresRooms, votes, itinerariesorchestratorLangGraph state graph, routes each turndiscoveryDestination optionslogisticsFlight searchitineraryDay-by-day plannegotiationFeedback and revisionsmodel and toolsGemini, flight APIs, places, weather
  1. A party of friends joins a room in the browser. The Next.js app subscribes to Supabase Realtime for presence, chat and votes.
  2. The UI calls FastAPI through typed service contracts, so it runs against mocks or the real backend. FastAPI runs in Docker on one EC2 server.
  3. Each message goes to a LangGraph orchestrator, which decides what the party needs next.
  4. Specialist nodes take one stage each: destinations, flights, the itinerary, and negotiating changes.
  5. The nodes call Gemini and live tools for flights, places and weather, then write results back for the whole party to see.

Results

  • Won AWS Kiro BuildFest 2026.

Screens

Back to all projects