Gaid

Inspiration

In Lagos, a gallery’s catalogue may be a PDF shared on WhatsApp, while a wall label has room for only a few dozen words. The fuller story of an artwork often stays with the artist or curator, leaving visitors with only a small part of what there is to learn. Galleris already gives galleries their own viewing room, opened from a code on the wall. GAID turns that viewing room into something visitors can talk to, so they can keep their eyes on the artwork instead of spending the visit looking at a screen.

What it does

1. Scan and ask — Phone Visitors scan the code beside a painting and press play to hear its story in the gallery’s own words. They can ask who made it, when it was created, what it is made from, and how large it is. As GAID speaks, the phone screen fades to encourage visitors to look up at the artwork. Look Closer explores the painting part by part while the story continues. Where Next recommends another work and explains why it is worth visiting.
2. The room answers — Fire TV A screen beside an artwork displays a QR code. Visitors scan it with their phone or tap Scan inside GAID to pair the devices without typing or installing another app. The TV confirms the connection before the phone does. When visitors ask a question, the TV can show supporting visuals such as an artist profile, a timeline, the artwork’s size beside a person, or the room on a plan. Look Closer moves on the TV and phone together. Each TV is linked to the artwork beside it, so visitors at different works do not interfere with one another. The Fire TV remote also supports a phone-free experience: OK plays the story, Up looks closer, and Play tours the room.
3. It follows you home — Echo Show When visitors leave, GAID gives them three words, one for each artwork they spent time with. Later, they can ask an Echo device about those works and continue the conversation. A single press can also send an enquiry to the gallery through WhatsApp or email.
4. The gallery learns — Curator Questions GAID cannot answer are added to a list for the curator. The curator can use these questions to improve future wall labels and provide information visitors actually want to know. The project is designed not to retain other information about visitors. The demo gallery contains 24 real paintings from the National Gallery of Art’s open-access programme. The works are public domain; the project is not affiliated with or endorsed by the Gallery.

How we built it

GAID uses an MCP server built with the official SDK, the 2025-11-25 specification, Streamable HTTP, and stateless requests. It exposes eight tools: ask_work, find_work, whats_nearby, recall_visit, where_next, look_closer, send_enquiry, and visitor_questions. The phone’s simulated Alexa+ experience calls these tools for each question, in the way a connected assistant could call an external service. One shared answer system serves the phone, TV, and Echo Show experiences. Curator-provided questions and answers take priority and are used word for word. Next, GAID checks the gallery’s published records. If further explanation is needed, Amazon Nova can use the information the gallery has published, including the description, background, credit line, collection, and exhibitions. The model is instructed to say that the gallery has not provided an answer rather than guess. The system is designed not to invent names, dates, prices, or ownership details. Nova can be accessed through Amazon Bedrock or through Vercel’s AI Gateway when an AWS account is not available. GAID is designed to fail gracefully. Languages are offered only when they can actually be spoken, and network, model, database, or TV-link failures should not leave visitors facing an error screen. The server acts as the gateway: database functions require GAID’s own secret, the TV displays only words signed by GAID’s server, and callers are rate-limited. The phone and TV communicate through a real-time channel named by the pairing code, and the connection details are not stored. The link was tested live from a separate machine against the real service, which exposed a fault that local testing had missed. Stories can use recordings from the curator or artist, or a voice copy of the gallery owner. The voice is restricted to words approved by GAID’s server. Galleris galleries can enable GAID on paid plans; it is off by default, and GAID reads only information the gallery has published. The app uses static pages and Node functions on Vercel. The team reports 103 automated tests, each checked by reintroducing the relevant bug and confirming that the test fails.

Challenges we ran into

One of the main challenges was designing a voice guide that is helpful without making up facts. Artworks can raise questions about meaning, ownership, dates, and prices, so GAID needs to distinguish between information the gallery has confirmed and information it has not published. Another challenge was making phone-to-TV pairing simple enough for a gallery visitor to complete without typing, while keeping each screen connected to the correct artwork. The experience also had to work across several device types without making the interface feel complicated or distracting from the art.

Accomplishments that we're proud of

We brought the phone, TV, and at-home experience together around one shared answer system. Visitors can ask about an artwork, explore its details visually, and return to the experience later. We are especially proud of the QR-based pairing concept, which avoids a lengthy setup, and the answer rules that prioritise curator-approved information over generated responses. The curator question list also turns unanswered questions into useful feedback for improving the gallery experience.

What we learned

We learned that a useful voice experience is about more than speech recognition or narration. It needs reliable source information, clear limits on what the AI can say, and a fallback when something goes wrong. We also learned that multi-device experiences need to be designed around the visitor’s attention: the technology should support looking at the artwork, not compete with it. Finally, testing the real service from a separate machine helped reveal an issue that local tests did not catch, reinforcing the importance of testing beyond the development environment.

What's next for Gaid

Next steps include adding recorded voices from Lagos galleries, exploring Yoruba, Igbo, and Hausa narration through Amazon Nova, and adding the Ask button to every paid Galleris viewing room. The longer-term goal is a real Alexa+ connection, so visitors can say “Alexa, ask GAID…” on an Echo device rather than using the simulated web experience.

Built With

• alexa+
• amazon
• api
• bedrock
• context
• echo
• firestick
• galleris
• javascript
• mcp
• model
• nova
• realtime
• show
• speech
• supabase
• tv
• vercel
• web