Kiosk AI Demo Uses Realtime API and MCP: Cool Tech, Predictable UX Pushback

A dev posted a short demo showing a kiosk-style building assistant on an OLED holographic display. The system uses the Realtime API with MCP to fetch the cafeteria menu; the interaction is triggered when you stand on a QR code spot on the floor. The avatar speaks and animates while the system reads the menu aloud. The post and discussion appeared on Reddit.

Top comments and UX reaction

Top comments made it clear the reaction was less about the engineering and more about the UX. One top reply summed it up in meme form: “she looks like she hates her job.” A higher-quality critique argued for a simple principle that kept recurring across threads: users want quick, scannable information, not an avatar recital.

What the demo shows (important tech signals)

• Realtime integration at the edge: streaming audio and visuals to a kiosk, a useful demonstration that Realtime API can power physical installations.
• Triggered interactions: ambient activation via a QR code spot is a pragmatic pattern for discovery (you opt in by standing there).
• Visual avatar plus speech synthesis plus data retrieval (menu API): a standard multimodal pipeline that can be extended to navigate building data, maps, or live signage.

Why the thread matters

This small demo illustrates a recurring trade-off as AI leaves the screen and moves into public spaces: novelty versus usefulness. The tech is now trivial to wire together; the harder part is designing interactions that scale to real-world constraints such as noise, crowds, accessibility, and limited user attention. In short: shipping an animated assistant is easy; shipping a useful one is not.

Community critiques worth noting

• Speed and discoverability: Many users pointed out that reading a menu aloud is slower than glancing at a poster or a printed sign. “I could read a sign way, way faster,” wrote one top commenter.
• Unnecessary anthropomorphism: Several replies asked why a full-body human avatar is needed; suggestions included a head and upper-body or a non-human visual that does not attempt realism.
• Accessibility and inclusion: Callouts for captions, text alternatives, and tactile or visual fallbacks for people who cannot hear, do not speak the same language, or have vision impairment.
• Practical constraints: Ambient noise in malls and queues is cited as a real obstacle to a spoken-only interface.

High-value comments

• rodrigobb: Recommended keeping audio but adding on-screen text, a map, and clear menu info such as opening hours and location. This mirrors HCI best practices: multimodal redundancy increases usability.
• SpiritualWindow3855: An HMI engineer noted the demo is interesting but has poor discoverability; they suggested shrinking or removing the 3D model and pairing the voice interface with a traditional UI.

Design and engineering suggestions (practical)

1. Show text and visuals alongside speech: menu text, pictures, and a mini-map. Allow users to skip narration.
2. Use upper-body or face-only avatars, or symbolic visuals, to avoid uncanny or full-body awkwardness and save screen real estate for useful info.
3. Provide captions, language selection, and a tactile or mobile fallback such as a QR link to view the menu on your phone.
4. Cache menus locally and prefetch to reduce latency in noisy or offline environments.
5. Handle queuing: short responses, single-click to copy a menu link to phone, or send the menu via SMS or QR to avoid blocking.

Conclusion

The demo is valuable because it demonstrates how Realtime APIs and simple triggers like QR spots can produce novel kiosk behaviors. Community feedback is a reminder that novelty does not equal value. Animations and holographic displays will draw attention, but attention alone will not solve the core user problem: getting information quickly and accessibly.

My takeaways

• Realtime streaming for physical kiosks is practical and worth exploring further.
• Prioritize multimodal redundancy: visual plus audio plus mobile fallback, and fast glanceability.
• Avoid gratuitous anthropomorphism; optimize the interface for clarity rather than spectacle.

There is often another side to the story: what looks gimmicky in a single demo can be the seed of genuinely useful products, but only if teams pair the technical plumbing with solid HMI design and accessibility-first thinking.