Wander

Wander's single, quiet surface — just ask where to go next.
Ask for Paris: rich property cards served from Amazon Aurora.
One tap books it: an ACID transaction in Aurora, a real-time event in DynamoDB, and Slack notified via Vercel Connect.
Your confirmed trips, read straight from Amazon Aurora.
Traveler stats, aggregated live across your bookings.

Inspiration

What it does

How we built it

Challenges we ran into

Accomplishments that we're proud of

What we learned

What's next for Wander

Inspiration

Planning a trip still means juggling a dozen tabs and forms. We wanted the opposite: a travel concierge you simply talk to. For the H0 hackathon — "Hack the Zero Stack" — we set out to prove a real, transactional product can run on a near-zero stack: AWS databases straight from the Vercel Marketplace, zero static credentials, and zero backend glue.

What it does

Wander is an AI-native travel concierge. You chat, and it searches stays, checks availability, books with a human confirmation step, tracks your trips, optimizes itineraries, converts currencies live, and remembers everything. Every answer comes back as a card — and every card can be saved and shared.

How we built it

Frontend + agent on Vercel — a single Next.js app. The agent runs same-origin via eve, Vercel's open-source agent framework (tools in TypeScript, instructions in Markdown, durable by default).
Two AWS databases from the Vercel Marketplace:
- Aurora PostgreSQL — the relational, ACID source of truth (properties, bookings, payments). A booking writes booking + payment in one atomic transaction.
- DynamoDB (single-table) — the real-time hot path: availability cache, activity feed, saved cards.
Zero static credentials — every AWS connection authenticates through OIDC federation (RDS IAM for Aurora, a scoped role for DynamoDB). No AWS keys are ever stored.
Vercel AI Gateway runs Claude (via OIDC, no API key).
Vercel Blob stores each generated card as a public, shareable artifact.
Vercel Connect posts a Slack notification the moment a booking is confirmed — scoped token, no Slack secret stored.
Vercel Firewall rate-limits the agent and login at the edge.

Challenges we ran into

Marketplace resource roles can only be assumed from the production runtime, so seeding and verifying happened in-prod. We mapped four logical stores onto DynamoDB's single-table PK/SK design, and made every authenticated user resolve to a valid UUID so transactional bookings never fail.

Accomplishments that we're proud of

One booking that, in a single tap, writes an ACID transaction to Aurora, updates DynamoDB in real time, saves a shareable artifact to Blob, and notifies Slack via Vercel Connect — end to end, verified in production, with zero stored credentials.

What we learned

How far the "zero stack" idea goes: provisioning real databases from the Marketplace, wiring them through OIDC, and shipping an agent + UI as one deployable unit.