All work
Case study · 01 Google DeepMind 2022 — 2025

Gemini App — Foundation, Gems, Agent.

Founding designer on the Gemini App. I defined the response model that shipped at scale, built Gems so anyone could create their own AI expert, and led the agentic framework that made Gemini Agent act — not just respond.

Role
Senior UX Lead — Founding designer
Team
10 designers · 6+ XFN partnerships
Surface
Bard → Gemini · Gems · Gemini Agent
Outcome
10+ launches · 260K Gems · Agent shipped

Chapter 1 · The foundation

Founding the response model you use today.

I joined Gemini (then Bard) before there was a product. Conversational AI didn't have a shared vocabulary for trust, control, or continuity — every team was inventing primitives in parallel, and every launch risked contradicting the last one.

I led the design of the response model — the way Gemini speaks, reasons in the open, cites its sources, and invites you to push back. Then I helped ship it ten more times: Gemini Advanced, Apps, Gemini in Chrome, Gemini Live, the rebrand from Bard to Gemini, Search, Scheduled Actions. The patterns I shipped here became the substrate everything else was built on.

If we don't define how Gemini behaves, every team will define it differently.

The framework

Four pillars. Every Gemini surface builds on them.

A pattern language is only useful if every team can build the same idea the same way. I anchored the system on four pillars — short enough to remember, sharp enough to argue against. They carried the response model, then Gems, then Gemini Agent.

01 Trust

Show your work. Make AI legible — the model's reasoning, sources, and confidence are part of the interface, not an afterthought.

02 Control

User stays in the driver's seat. Pause, redirect, or take over at any step. Autonomy is granted — never assumed.

03 Continuity

Context flows across surfaces and sessions. The agent remembers what matters and forgets what shouldn't be remembered.

04 Composable

Patterns scale to dozens of teams without bespoke variants. Trust compounds when builders share a language.

Scale

10+ major launches. One coherent system.

From Bard's first conversation to Gem Sharing Enterprise — every launch built on the same foundation. Each one stress-tested the patterns, each one made them sharper.

Hover any tile for my contribution.

01

Founding designer on the launch surface — defined the response model, reasoning trace, and trust UI that became the baseline for every Gemini surface after.

Bard
02

Led the premium experience design — long-context handling, advanced settings model, and the upgrade flow that scaled the framework to power users.

Gemini Advanced
03

Designed the extensibility layer — connector grants, scope visualization, and the at-mention pattern that shipped to every Gemini surface.

Apps
04

Partnered with Workspace to bring Gemini into Docs, Gmail, and Drive without breaking each app's existing model. Cross-org pattern reconciliation.

Gemini in Chrome
05

Voice-forward variant of the framework — adapted Trust, Control, and Continuity to ambient interactions and barge-in behaviors.

Gemini Live
06

Helped lead the system-wide rebrand from Bard to Gemini — visual identity migration, surface taxonomy, and entry-point unification across products.

Rebrand
07

Pitched and shipped from zero. Created the agent-creation pattern language — instructions, knowledge, identity, scope — that scaled to 260K+ Gems.

Gems
08

Designed the recurring-task primitive — confirmation, edit-in-place, and rescue patterns for agents that act outside the user's session.

Scheduled Actions
09

Brought Gems into Workspace admin — share scopes, governance, and the admin-review pattern for company-wide agent rollouts.

Gem Sharing
10

Pitched and led from zero. The autonomous leap — Gemini that acts on your behalf across multi-step tasks, with confirmation gates and full interruptibility built into the framework.

Gemini Agent

Chapter 2 · Gems

Custom AI experts, built by anyone.

Gems ↗ are configurable AI experts — career coach, brainstorm partner, coding helper. Pick a premade one, or write a few sentences of instructions and ship your own. Configurable, scoped, deterministic. Not autonomous: a Gem stays inside the chat and follows the rules you gave it.

I pitched Gems from zero and shipped it at Google I/O. The pattern language for AI configuration — instructions, knowledge, identity, scope — didn't exist. I had to invent it. Three principles drove every decision:

Approachability Clear language, intuitive, helpful.

A grandparent should be able to make a Gem. The first-run experience hides every setting that isn't load-bearing.

Flexibility Advanced options for power users.

The depth is there when you want it — instructions, knowledge, tone, default tools. Layered complexity, not gated.

Transparency What the model knows and remembers.

Every Gem makes its scope visible. No invisible state, no hidden context. Trust is the substrate.

At launch · first six months

80KWeekly active users
260KGems created
1.6MChats with Gems
Paid retention impact

Proof in the wild: YouTube creator Paul (AI Advantage) walks through building Gem after Gem — research assistants, writing coaches, code reviewers — in minutes each. It's the clearest demonstration of the principle: when the creation surface gets out of the way, people build the agents they need, not the ones I imagined.

"One of the most used Gemini Advanced features."
9to5Google ↗

Chapter 3 · Gemini Agent

Gemini needed to act, not just respond.

Gemini Agent ↗ is the autonomous leap. Not a configurable persona — a system that handles complex, multi-step tasks from start to finish. It does the research, browses the live web, compares options, drafts the email, schedules the meeting. You stay in control: it confirms before critical actions, and you can stop or take over at any time.

This is the chapter Gems couldn't be. Gems are deterministic — they follow the instructions you give them inside a chat. Gemini Agent acts on your behalf across surfaces, sessions, and apps. Nobody had designed that before. I pitched the framework, defined the strategy, and led the work from zero.

The four pillars carried over — Trust, Control, Continuity, Composable — but each one had to be redesigned for an AI that does things instead of describing them:

Trust Show the work before you do the work.

Every step the agent plans is visible before it runs. Reasoning trace, tool calls, intermediate results — all surfaced. Trust is earned action by action, not granted up front.

Control Confirm critical actions. Always interruptible.

Sending email, making a purchase, modifying calendar — the agent stops and asks. The user can interrupt mid-task and take the wheel. Autonomy is granted in increments, never assumed.

Continuity State that survives across surfaces.

An agent that acts has memory consequences a chat doesn't. Designed the patterns for context that persists across sessions, apps, and devices — and the controls to scope, edit, and forget it.

Launch · Google I/O 2025

Three years of work. One keynote.

Sundar Pichai unveiled Agent mode to a live audience at Google I/O 2025 — the framework I'd been building from zero, now shipping to hundreds of millions of users. The moment a response model became an execution layer.

"They might've just changed the game for AI agents."
Everyday AI · Medium ↗

Design thinking · Gemini Agent

Designing for autonomy — the hard problems.

Agentic design isn't interface design with a new coat of paint. These are the specific problems I worked through — each one required rethinking a pattern that worked fine for chat.

Problem: agents acted without warning on irreversible tasks. I designed the confirmation gate — a consistent pre-action surface that shows the plan, lets the user edit or abort, and builds trust incrementally.

Confirmation gate

Problem: once an agent started, there was no graceful way to stop it. I designed mid-task interruption as a first-class pattern — pause states, partial undo, and clean handoff back to the user at any step.

Interruption + takeover

Problem: model reasoning was hidden behind toggles, eroding trust. I pushed for an inline reasoning trace — a scannable step-by-step format visible in the response itself, not buried in a debug panel.

Reasoning trace

Problem: multi-step tasks crossed surfaces and sessions with no shared state model. I mapped every transition end-to-end — confirmation points, failure modes, and handoff patterns across research → draft → send.

Multi-step task flow

Decisions

Three decisions that shaped the arc.

One per chapter — the moments I had to push back, advocate, or change direction across Foundation, Gems, and Gemini Agent — and what shipped because of it.

Foundation Reasoning trace in the response, not buried in a panel.

Engineering wanted to hide model reasoning behind a "more details" toggle to keep the chat clean. I pushed for an inline, scannable trace as part of the response itself. My position: trust isn't earned by tucking the work away. It's earned by showing it.

→ Shipped inline. Became the trust pattern every later surface inherited.
Gems Names and identity vs. neutral labeling.

Engineering and PM argued for neutral framing — "Custom AI" — citing anthropomorphization concerns. I pushed for full identity: name, description, avatar, tone. My position: named experts create habits. Unnamed tools get forgotten.

→ Shipped with identity. Return usage notably higher for named Gems.
Gemini Agent Confirm-before-action vs. silent autonomy.

Some on the team argued the agent should "just do it" — fewer interruptions, faster perceived value. I pushed for explicit confirmation gates on critical actions (send, purchase, calendar writes). My position: an agent that surprises you once loses your trust forever.

→ Shipped with confirmation gates and full interruptibility. Became core to the framework.

Impact

A framework the org could build on.

Three years. Three chapters. One pattern language that carried from a blank-slate response model through Gems — 260K created, 1.6M chats, 2× paid retention — to a product category redefined at I/O. The four pillars weren't a deliverable. They were the operating system other teams used to ship.

260KGems created
1.6MChats with Gems
Paid retention impact
10+Major launches

Next case study

Case study · 02 — Intuit

Financial Intelligence — Intuit × OpenAI & Anthropic.