Currently @ Google DeepMind leading design for multimodal conversational experiences in the Gemini App (1B+ users)
01Product + Model Design2026
Multimodal image generation
Turning a wild idea into a tangible proof of concept right at the moment you're experiencing it by integrating Nano Banana into Gemini Live
I wasn't quite sure that Gemini Live was the right place for Nano Banana. But then I built a prototype, and it clicked – not just for me, but for my team. And it ended up clicking for users too, becoming the most used tool in Gemini Live a within the first month of its launch.
Highlights
Prototyped the core interactions and flows in AI Studio on the actual APIs, which secured funding and headcount and drove product & technical requirements. Tested and refined interactions in Antigravity, with Claude models.
Rewrote the system instructions governing the handoff and the stale, robotic completion phrasing that followed it to provide clearer transparency of system status and maintain overall conversationality.
Identified potential collision with another feature – visual cues – and implemented clear decision-making heuristics to address the overlap
Defined the eval rubric & set, analyzed losses, and adjusted the design accordingly. For example, the loss analysis indicated the default size of the generated images made it difficult to quickly preview and refine the image.
Knowing if, when, and how to show and tell in Gemini Live by introducing audio-visual responses in Gemini Live
Visuals are helpful, but they're not always necessary. We shouldn't show for the sake of showing. To validate this hypothesis, I took a “fast and slow” approach – learning quickly through a focused set of deterministic outputs and developing a foundational framework, eval rubric, and training data to shape model behavior upstream.
Highlights
Defined a clear, actionable roadmap, framework, and eval rubric adopted by the foundational audio model team and other design teams across devices and product areas
Navigated pushback on my original thesis because it added rigor that read as unnecessary complexity. Held the line to avoid adding extraneous cognitive load and breaking the core premise of audio interfaces, using prototypes to illustrate my points.
Shipped a deliberately narrow first version at I/O 2026 to learn from real usage
Built a tool in Antigravity with Claude models to quickly visualize raw outputs from the model to facilitate faster analysis and revisions to the guiding system instructions
03Design Lead2025
Video and screen sharing
Transforming Gemini Live into a true multimodal assistant that can see and understand your physical or digital world as you move through it
It started as a research prototype. There wasn't much else to go off of. At I/O 2024, Project Astra introduced the future capabilities of a universal AI assistant capable of understanding the world around you. With limited foresight into how users would actually use this feature, I defined the core interaction framework and taxonomy, outlining the solution space and guiding the team to productionize this research prototype months ahead of schedule, winning a “Best of MWC” and “GLOMO” award and becoming the top driver of new users and key retention lever.
Highlights
Addresed the shortcomings of our previous eval approach – a random spread of prompts – by introducing a taxonomy that represented the actual range of use cases based on user insights and model capabilities. This enabled us to better test and hillclimb on quality, and inform which features were most important. The taxonomy format became the “gold standard” across the team.
Directed the team through a ship date moved up by months, cutting scope to land the core experience sooner.
Introducing a more natural way to communicate with AI through Gemini Live
OpenAI’s Advanced Voice Mode launched in May. We needed to respond… and quickly. I had just joined the Gemini Live team a month prior. Before I could effectively lead anyone, I had to get my arms around work to-date, structure the thinking, scope the work, and mobilize a team of 10 people who hadn’t worked together, balancing oversight and empowerment to move fast and as one.
Highlights
Drove prioritization of Live MVP requirements and tasks with XFN leads to ensure team focused on the top priorities based on user needs, business strategy, and success criteria
Created and managed the Gemini Live MVP design system, ensuring quality and cohesion across the E2E experience
Pre-reviewed all UX changes from the design team, identified and resolved gaps, and worked with designers to clarify designs and rationale. This operationalized clear communication and ensured smooth, high-velocity collaboration to meet the MBG launch deadline.
Oversaw integration into the larger Gemini Design System, demonstrating system-aware craft by advocating for the core design intent of Live and preserving key decisions to deviate from the main app where necessary.
Connected emerging AI trends to specific business opportunities and user needs to identify key opportunity areas for Live and organized PM & UX leads around a post-launch strategy.
Making Gemini more helpful and accessible by integrating directly with your device
From pitch deck to the app store; taking an app from 0 to 1 (billion) When Gemini first launched, it was capped by the browser. I was on a team of 5 designers to pitch execs on a different bet: a mobile app that would unlock new capabilities, increase usage, and facilitate the transition from Google Assistant. The Gemini mobile hap became the fastest growing app in Google history, surpassing 1 billion users in 2 years.
Highlights
Owned design for two of the app’s highest visibility surfaces: onboarding and home page. Led a team of 3 designers, reviewing and steering work, managing changing feedback from leadership, and jumping in as needed.
Navigated device manufacturer constraints, regulatory compliance, multiple platforms, and branding changes without slipping on timeline.
Identified a fundamental conflict between 2 key user journeys (input vs browse) on the home page, an advocated for an opinionated design rather than trying to equally serve both goals, which would compromise coherence and clarity.
I've always been fascinated by the “and“ — by working in the seams. It first led me to study human-computer interaction and now to work between the model and the interface, understanding and shaping emerging capabilities in ways that work for and with people, ensuring that we don't outsource what makes us, at our core, human.
Outside of work, I am:
Plotting on my next travel adventure (last = Puerto Rico; next = Greece)
Tinkering in the kitchen (was recently afflicted by the Covid-era bread obsession)
Finding the best slice in NYC (current = Chrissy's)
Wrangling a happy hour & trying to perfect the art of “splitting the G”
