Adaptable Generative UI

For someone experiencing a migraine with visual aura, a bright screen is a source of pain. Traditional tablet apps expect users to navigate menus, tap buttons, and read text. But what happens when looking at a screen is the thing you're trying to avoid?

This is the core design challenge for Reachy Mini Minder. I needed to build an interface that adapts to the user's current state. The screen is always there, but it's never required. Voice and physical presence handle everything.

Generative UI is a pattern where the AI agent selects and renders UI components as part of its response. Instead of returning text alone, the LLM picks from a registry of pre-built React components and pushes the right one into the conversation stream with the right data.

To be clear: the AI does not create new interfaces from scratch. Every component is hand-built. The "generative" part is the selection: the LLM decides which screen to show based on what the user said, and when.

For a voice-first care companion, that selection is the hard problem. Consider the scenario from the opening: a user mid-migraine says "I have a headache." The robot:

Zero taps. Zero navigation. The user never opens their eyes. A regular voice assistant gives you a text response. GenUI gives you a structured medical record rendered visually for later review, built entirely through natural conversation.

The same pattern scales to simpler interactions. "Play some music" produces a grid of tappable cards with number badges. Say "number two" or tap the screen. Both work.

A user might say "I took my pills," "meds done," or "yeah I had the morning ones." Mapping that range of free-form speech to the correct clinical component using conditional logic means writing fragile intent-matching code that breaks on every new phrase. The LLM handles it naturally. For a fixed dashboard with known user flows, this would be over-engineering. But Reachy Mini Minder's input is natural speech, and the space of possible requests is wide and unpredictable. That is where this pattern earns its complexity.