TurnUp

AutonomousHacks, hosted by GDG Gandhinagar and Google for Developers. The brief: ship something autonomous that solves a real problem. Three teammates, 36 hours, no slides.

We picked trip planning — a fragmented problem where five apps (chat, maps, splits, notes, calendar) handle one job badly. We wanted to fold all five into one app, and let agents do the boring planning work.

One Node + Postgres backend serving both web (React) and mobile (React Native via Expo). Shared types, shared models, two clients. Real-time was Socket-io for chat + location pings.

The agent layer was the differentiator. Small, single-purpose agents — one for itinerary drafting, one for restaurant clustering, one for expense allocation — coordinating through a shared trip state. Cheaper than a single big-context call, easier to debug, and weirdly fun to watch.

Splitting backend ownership early. One of us owned the API surface, one owned the agents, one owned the clients. We never blocked each other after the first three hours.

Also: live location worked the first try. Nothing demos better than a moving dot.

The expense-split agent kept hallucinating decimals at hour 28. Replaced it with a deterministic function, demoed it inside the agent flow, and flagged the swap in the writeup. Lesson: not every problem is an LLM problem.

Shipped a functional prototype with all four pillars working — chat, live location, navigation, splits — plus agent itinerary generation. Beyond the hackathon, the codebase is now the starting point for a serious rewrite.