Google Deepmind: Contextual UI
Traditional generative AI is a wall of text. We thought we could do better. For this project, our goal was to catch specific categories of user prompts and, rather than generating raw text or table data, introduce smart interactive UI components. Rather than making people read through a massive paragraph to find what they need, we created a system capable of adapting to provide them with information presented in a familiar, and easy to digest, way.
NDA note: The focus below is on the underlying behavioral architecture and systems strategy
Role: UX Lead, Gemini Team
Context: Confidential MVP Launch
Smart Routing & Safe Answers
Working alongside the prompt engineer, we tuned the model's system instructions so it could correctly read and respond to the user to determine they actually needed. With that determination locked it, we were able to then shift the user into one of two paths:
Quick and easy: For straightforward requests, the system immediately serves up a clean, structured layout so people get what they need without waiting or scrolling through walls of text.
Deep dive: For more complex situations, the model slowed down and acted more like a guide, asking the right questions to gather context before responding with an appropriate UI module contextual to the user’s intent.
We also had to build strict guardrails into the prompting. We set clear boundaries around sensitive topics such as medical advice (critical, as legal kept reminding me, that Gemini never played doctor) and actively trained it to keep responses objective and safe instead of just validating unverified assumptions.
Building the UI Framework
To keep people from getting overwhelmed by wall-to-wall text, I designed a flexible card system that broke complex AI responses into clean, easy-to-read pieces containing pertinent images, information, and links.
In the spirit of engineering efficiency, rather than letting adjacent teams build custom modules for their own responses, I created a single universal container system capable of adapting to whatever content it contained:
Flexible content types: I designed the container structure to gracefully accommodate highly variable text lengths of model-generated content, shifting fluidly from compact headline alerts to deep, multi-tiered instructional steps and conversational macro-data points without breaking the grid or requiring manual designer intervention.
Swappable content: The container was set up to automatically change its layout based on what the model returned, easily swapping between rich editorial photos, structured lists, and interactive buttons.
This universal card component was adopted across internal model teams as standard, saving significant development time by requiring only one core component build to handle an array of possible response variations.
Keeping Everyone on the Same Page
Shipping at this scale meant moving fast while staying careful. I focused on three main areas to make sure everything ran smoothly:
Legal & safety alignment: I worked closely with the legal and safety teams to make sure important disclaimers and guardrails showed up naturally, without getting in the way of a smooth experience.
Keeping design & code in sync: Wrote clear behavior guidelines and teamed up with engineers for hands-on design QA, making sure what we designed was exactly what got built.
Alignment with leadership: Put together interactive prototypes and presentations for big stakeholder reviews, including sessions with Sundar Pichai.
The Impact
In the end, I delivered a successful high-context MVP that defined how Gemini structures agent-driven content, establishing the foundational patterns and layout logic for the platform's next-generation adaptive interfaces.