ArtSeekers
Welcome to The Gallery. Let's ask an AI curator about an artwork, and see how it looks at the picture, not only the title.
The story
ArtSeekers is a website for exploring art from The Met, a museum in New York. You can browse by department and keep a list of favourites.
Its special part is an AI curator. Open an artwork and ask: the AI is sent the picture itself, not only the title, and writes two or three sentences about what stands out. Each answer is saved, so the next visitor reads it right away.
My role
I built the backend of the first version, a 4Geeks Academy group project with two classmates, put it online and connected the external database. Later I refactored and updated it on my own with AI coding tools, including the AI curator.
Overview
ArtSeekers explores The Met's collection with an AI curator that looks at each artwork's image. One Flask app serves the React frontend and the API, and the artworks are stored in PostgreSQL (Neon).
My role. The first version was a 4Geeks Academy group project with two classmates, programmed from the bootcamp's template: I built the backend and helped on the frontend. I put it online and connected the external database. Later I refactored and updated it on my own with AI coding tools: a faster artwork list, sessions, the AI curator and a responsive layout.
Architecture
- Frontend: React and Bootstrap 5.
- Backend: Flask and SQLAlchemy, with JWT login, on Heroku.
- Curator: the backend looks for a saved answer. If there is none, it sends the image to Gemini and saves the reply.
Key decisions
- Look at the picture, not the title. If the image cannot load, the curator returns an error instead of guessing.
- One analysis per artwork. Saved answers mean one AI call per artwork (Stop 4).
- A faster list. The artwork list builds a department lookup once: O(N×M) down to O(N+M).
- Try it free, with a limit. Visitors get three new analyses a day, counted on the server without storing the IP.
- The server decides what the AI sees. The browser only says which artwork, so nobody can slip in another image or fake text. I also removed two unused routes that exposed users' emails.
Testing and evaluation
There are no automated tests yet; a health check watches the live site. Nobody has measured whether the curator's descriptions are accurate.
What I'd do next
- Add tests with a fake Gemini and a fake Met API.
- Check the curator against artworks with known facts.
- React
- Bootstrap 5
- Context API
- Webpack
- Python
- Flask
- SQLAlchemy
- Gunicorn
- JWT
- PostgreSQL (Neon)
- The Met Museum API
- Google Gemini API
- EmailJS
- Heroku