Ari’el: An Embodied AI Guide

Inspiration

Complex software often assumes that users already understand its layout, vocabulary, and hidden rules. Traditional help systems make people stop what they are doing, search through documentation, and translate instructions back into the interface.

Ari’el explores a more human approach: an embodied guide that enters the software environment, moves toward the control the user selected, explains it in context, and responds through voice, animation, and spatial behavior.

What it does

This prototype places Ari’el inside a VR Quest training interface.

When a user selects a menu, tool, or control, Ari’el can:

• Move toward the selected area
• Visually acknowledge what the user chose
• Explain the feature in context
• Read the explanation aloud
• Synchronize facial movement with speech
• Use motion, proximity, and sound to feel present rather than pasted onto the screen
• Remain available inside the application without forcing the user to leave their task

The result is a help system that behaves more like a knowledgeable guide standing beside the user than a conventional manual.

How we built it

The demonstration is a browser-based interactive prototype built with HTML, CSS, JavaScript, browser speech synthesis, animation, contextual help content, and spatial audio behavior.

OpenAI Codex was used throughout the build to:

• Refactor the interactive HTML, CSS, and JavaScript
• Improve animation behavior and character responsiveness
• Correct menu alignment and zoom-related selection drift
• Integrate speech, mouth synchronization, and movement audio
• Improve Ari’el’s contextual responses
• Test, repair, and package working iterations rapidly

GPT-5.6 was used for product direction, usability review, response writing, feature prioritization, submission planning, and translating testing observations into specific engineering changes.

Challenges

The hardest problem was not simply making a character move. Ari’el had to remain connected to the user’s action.

Her visual position, clickable regions, speech, animation, and explanation all needed to refer to the same interface element, even when the browser was resized or zoomed. We also had to make her feel alive without allowing her to distract the user, block controls, or move beneath the cursor.

Accomplishments

We created a working embodied guide that can move through a real software interface, react to user selections, explain controls, speak, gesture, and use sound to reinforce motion and distance.

Ari’el is not a prerecorded video or static chatbot overlay. She is an interactive interface component designed to become part of the software she is teaching.

What we learned

The interface changes when the AI-guided character is treated as a participant rather than a separate help panel. Context becomes physical: where Ari’el moves, what she looks toward, when she speaks, and how close she appears all contribute to the explanation.

We also learned that embodiment must preserve agency. The guide should support attention, not capture it.

What’s next

The next stage is to connect Ari’el to a true reasoning and context system so she can generate more adaptive explanations rather than relying primarily on authored interaction logic and help content.

Future versions may integrate LIMINA’s broader privacy-respecting architecture, but that integration is not part of this submitted prototype.

The same embodied guidance pattern could support education, accessibility, industrial training, healthcare software, creative tools, and other complex systems where people need guidance without leaving their work.

Built With

• chatgpt
• codex
• css3
• html5
• javascript
• openai
• web-audio-api
• web-speech-api