Google Gemini

in Wearables

I worked within the Google Assistant team in the areas of content and conversation design for multimodality on wearables.

Some of my responsibilities included assisting with fine tuning Gemini responses, creating synthetic data for training, and designing content to deliver customer experiences that were helpful, timely, and relevant.

My Roles

Conversation & Content Designer

Team

Conversation & Content Designer
Product Designer
Product Manager

Years

2023-2024

Contact

email@domain.com

000-000-000

— Instagram

— Twitter

— Facebook

Problem

Google Gemini was expanding across devices where screen real estate is limited and attention is brief. In wearables, users do not want verbose, overly detailed responses. Instead they need precise, scannable, context-aware information delivered in the shortest possible interaction.

When integrated with Gemini, the next generation of Assistant intelligence, responses were too verbose for wearables and often contained irrelevant or tangential information. This caused:

- Frustration due to length and cognitive load
- Increased latency before task completion
- Obstructed key insights in moments where glanceability matters most

Product Opportunity

Wearables demand a tailored response framework that:

- Prioritizes short, signal-first information
- Aligns with user goals within quick interactions
- Maintains relevance without extraneous content
- Scales across audio, text, and visual modalities

- This is a classic content-design optimization problem at the intersection of multimodal UX and conversational context prioritization.

Design Constraints

- Output must fit within limited UI real estate while still being actionable

- Content must be comprehensible in audio-only and screen contexts

- Gemini patterns could not be deeply modified

- Framework needed to inform future generations of wearable UX

Research & Discovery

To ground the framework:

- I analyzed real Gemini responses to identify verbosity patterns
- I mapped common queries to information needs and usage contexts
- I defined edge cases where users required exceptions to brevity (e.g., procedural steps)

These activities made it possible to differentiate noise from signal in AI-generated responses.

Framework

I collaborated with the product designer to create a Wearable Response Framework that:

1. Categorizes queries by intent and urgency

- Immediate action required
- Informational glance
- Follow-on interaction

2. Defines output profiles for each category

- Concise: <15 words, actionable (e.g., “ETA 5 min. Leave now.”)
- Contextual: paired data points with priority hierarchy
- Expanded only on demand: drops secondary detail behind user request

3. Applies synthetic training data

- Created high-precision examples to correct Gemini tendencies
- Ensured multi-criteria outputs (accuracy, brevity, relevance)

4. Supports multimodality

- Adjusted for when audio is primary
- Adjusted for when text + visuals are available

- The framework was used to set guidelines for AI behavior without altering core models.

Systems-Level Takeaways

Device context matters: An assistant cannot use the same output across screens and wearables.
Signal prioritization trumps completeness in quick-interaction interfaces.
Content design shapes AI utility as much as the underlying model

Embedding discipline into LLM output through structured frameworks enables predictable, reliable, and user-centric outputs even from generative systems.