ArtDex

Snap real art at a museum — AI on Amazon Bedrock identifies it, Amazon DynamoDB collects it.

Comment

Inspiration

Most people walk past a masterpiece, take a blurry photo, and forget it by the time they reach the gift shop. You can't buy the Mona Lisa — but you can keep the moment you stood in front of it. Pokémon GO got people to walk miles to "catch" digital creatures; we asked what if the rarest thing you could collect was being there, in front of the real Starry Night, with a record that can't be faked. The interesting part isn't the photo — it's the data model that makes where and when meaningful.

What it does

  • Snap → identify → collect. Point your camera at an artwork; AI matches it and adds it to your personal art "Dex" with a rarity reveal.
  • The database is the game. GPS finds your nearest museum → Amazon DynamoDB returns the works currently exhibited there today → the photo is matched against only those few candidates, which is what makes recognition reliable.
  • Artworks travel — and the app shows it. Exhibitions are a time-bounded artwork↔museum relationship, so a piece's location is dynamic (The Starry Night: London → Paris → New York). Each artwork's detail view surfaces this exhibition history straight from DynamoDB, and meeting the same work again in a new city strings your encounters into one memory timeline.
  • Location-gated legendaries. The rarest works can only be sealed when you're physically within 150 m of the holding museum — enforced server-side with a haversine check.
  • Keepsakes. Each capture saves a selfie-with-the-art to Amazon S3 and a "moment" (with GPS coordinates) to DynamoDB; revisit a work and it becomes a reunion.
  • World map + progress. Collected pieces pin to a living world map; per-artist progress ticks up.

How we built it

  • Frontend: Next.js 15 (App Router) + React 19 + Tailwind 4 + TypeScript, deployed on Vercel. Live camera via getUserMedia; world map via react-leaflet + OpenStreetMap.
  • Database — Amazon DynamoDB (@aws-sdk/client-dynamodb + lib-dynamodb), PAY_PER_REQUEST, region us-east-1. Five tables — artdex_artworks, artdex_museums, artdex_artists, artdex_exhibitions, artdex_collections (per-user append-only "moments"). Catalog: 60 artworks · 14 museums · 26 artists · 66 exhibition rows, with several masterpieces touring three cities to exercise the temporal model.
  • Recognition — Amazon Bedrock, claude-haiku-4-5 vision. We send the photo plus the GPS-narrowed candidate list and ask for an id or "none."
  • Images — Amazon S3 presigned PUT/GET for selfie keepsakes.
  • Auth: cookie-based anonymous id — one Dex per browser, zero friction for a demo.
  • Pure domain layer: rarity, the 150 m location gate (haversine), candidate filtering, recognition parsing, and progress math are pure, fully unit-tested functions (56 tests).

Challenges we ran into

  • Aurora → DynamoDB pivot. We specced on Aurora PostgreSQL + PostGIS, but the free AWS plan gates Aurora's Data API behind a paid tier. We moved the data layer to DynamoDB and pushed the geospatial work (nearest-museum, the 150 m gate) into a pure haversine in the app layer — the temporal + geospatial model stayed intact.
  • Reliable recognition. Open-ended "what painting is this?" is unreliable; candidate-set narrowing (museum → today's exhibits → match against only those) is what makes it work.
  • Demoing GPS from a desk. A NEXT_PUBLIC_MOCK_LOCATION override lets us demo the on-site legendary gate without traveling.

Accomplishments that we're proud of

A database schema that is the gameplay: time-bounded exhibitions and a 150 m geofence turn a CRUD app into a place-and-time game. End-to-end on real AWS — deployed and working.

What we learned

DynamoDB's single-table, append-only "moments" shape fits an event-sourced collection log cleanly, and narrowing the problem (candidate sets) beats reaching for a bigger model when you need reliability.

What's next for ArtDex

Friends and shared world-map footprints (aggregated in DynamoDB), a museum "passport" sub-goal, and museum B2B partnerships. (Achievements and a global collectors' leaderboard already ship.)

Built With

Share this project:

Updates