Crashly

Crashly

Inspiration

As students traveling for hackathons like YHack, finding a place to stay quickly became one of the hardest parts of the experience. Hotels were expensive, often required guests to be 21+, and coordinating with friends or mutual connections was difficult and unreliable.

We kept thinking: why isn’t there a smarter way to use the college networks we already have?

So we built Crashly – an AI-powered lodging agent that automatically taps into your personal network, extends to friends-of-friends, and falls back to verified housing options to find you a place to stay.

What It Does

Crashly is an AI-powered housing agent for college students traveling to events, conferences, and competitions.

Users enter their destination and dates, and the agent:

Searches direct connections in your network
Expands to friends-of-friends with trust-gated access
Falls back to Airbnb, Craigslist, and housing groups
Ranks options based on compatibility, trust, and preferences
Generates a unified, explainable summary
Enables in-app outreach with structured request messages

How We Built It

We built Crashly as a mobile-first, AI-driven system.

Frontend: React Native + Expo for an application that is iOS-focused, with custom animations and a minimal design system built with StyleSheet.
- Expo extensively used to simulate a mobile app, giving us a single TypeScript codebase that runs natively on both iOS and Android, from the animated onboarding flow all the way to the AI-powered search
Backend: Supabase with Postgres database for profiles, a social graph with BFS-derived connection degrees, trip storage, and in-app messaging. No separate server.
AI Agent: Three-stage pipeline via Anthropic API for degree 1 networking, degree 2 networking, and open web synthesis. Agent reasons over compatibility, host ratings, and user preferences at each stage.
Web Search: Tavily API runs six parallel searches across the web for lodging and housing groups. Results synthesized by Claude.
Location: Google Places API for city autocomplete, Google Maps for destination pinning.
Infrastructure: Supabase Edge Functions (Deno) handle all agent logic, Tavily calls, and database queries in a serverless pipeline.
Google Places API: Handles location search and mapping

Challenges

Expo versioning: Our original build pulled a canary build incompatible with Expo Go. Getting a stable build required constant package conflict debugging and version pinning.
Managing GitHub collaboration across teammates (merge conflicts, broken features)
Edge Function debugging: Deno-based functions with no local runtime meant every bug required a full redeploy and log cycle. A header type mismatch in the Anthropic fetch call took significant time to isolate.
Integrating multiple APIs without causing inconsistencies.

Accomplishments

Built a trust-layered housing network that searches 1st-degree contacts, 2nd/3rd-degree connections, and the open web in parallel using a Claude AI agent
Gated the entire platform to verified .edu emails only, ensuring every user is a real, enrolled student
Designed a multi-step onboarding flow in Expo + React Native that dynamically adapts based on whether the user is a host, guest, or both
Shipped a working AI ranking system that scores results by trust, cost, and convenience, not just linearly by availability

What We Learned

Through building Crashly, we deepened our ability to debug complex systems and integrate multiple APIs into a single cohesive pipeline. It highlighted the importance of designing reliable fallback logic in AI products to ensure users always receive useful results.

At the same time, we learned how to balance technical complexity with user experience, making sure the product remains minimal and intuitive while handling sophisticated logic along with the fallback system. Most importantly, turning a real problem we personally experienced into a meaningful, tangible solution made the entire project especially impactful for us.

What’s Next

Crashly is just getting started. The foundation is built, now we scale the trust layer and deepen the network.

Stronger identity verification: student ID upload, university SSO login, and optional social graph linking (Instagram, Snapchat) to surface warm connections before a stay request is even sent
AI matching 2.0: move beyond availability and price to match on personality signals, study habits, sleep schedules, and past stay reviews
Event-mode: surge matching for conference weeks, spring break, and recruitment seasons when students need short-term stays the most
University partnerships: work directly with housing offices to surface off-campus options and integrate with existing student networks
Expand beyond the US: Crashly's model works anywhere students travel; next targets are Canada, the UK, and Australia