This app was built by James Rushford.
For questions or partnership opportunities, connect with me on LinkedIn: linkedin.com/in/rushfordj.
I'm currently looking for a job. If you're hiring, let's talk.
Many museums have limited labels, no audio guide, or an audio guide that is outdated. This app helps you explore what you're seeing by letting you ask questions in the moment and listen to richer explanations while you walk.
Museums are the authority on their collections. Museum Guide sits alongside that work and makes it easier for visitors to explore, learn, and share curiosity with other visitors.
Museum Guide is built on three layers of information. Each layer has a different purpose and a different level of authority.
1) Wikipedia and other canonical sources
When an artefact has a Wikipedia page, we use it as the foundation for that artefact's context in this app.
Wikipedia is not perfect, but it is public, editable, and built around transparency, citations, and correction. When the Wikipedia article improves, this app can improve too. This layer is treated as the closest thing to a shared public reference point.
2) Knowledge text (the artefact record)
Not every museum object has a Wikipedia page. When there is no suitable Wikipedia article, the app creates a knowledge text for the artefact.
Knowledge text is meant to behave like a small, museum-scoped version of Wikipedia. It is a canonical description for the artefact inside this app. It can be corrected and improved over time when information is missing, unclear, or wrong. The goal is clarity and accuracy, not a perfect story.
If we discover an error in what the app says about an artefact, the correct place to fix it is in the knowledge text or in the Wikipedia source it is based on. That way the correction becomes part of the foundation.
3) AI-generated introductions and answers
On top of those sources, the app generates introductions and answers to questions using an AI model.
These are designed for listening and exploration, like an unofficial tour guide. They are helpful and often very good, but they are not treated as the permanent truth record. We do not edit these outputs directly. If an answer contains a mistake, we correct the underlying source layer instead, then regenerate. That way improvements come from better foundations rather than patching individual conversations.