Product Design, Engineering
Role
Sr. Product Designer
Team
AI/ML
Engineering
Product
Data Science
Overview
Experiments is an internal AI-powered decision intelligence product for channel sales teams and geo partners. I led design across analytics, experiment creation, and agentic AI.
The challenge: teams needed to learn from completed work and use it to decide what to do next. My work connected the experiment lifecycle: intent, prior work, development, and outcome.
Problem
As Experiments grew, the Library became a record of completed work, not yet a source of intelligence. Teams could access past experiments, but couldn’t carry their learnings forward.
Experiments accumulated — knowledge stayed buried
Similar work existed — discovery was difficult
Outcomes were captured — learnings didn’t shape what came next
Our goal: turn the Library from an archive into the intelligence layer of Experiments, so past work informs decisions and what comes next.
The bet
The design direction was to make the agent useful across the experiment lifecycle. That meant more than answering a prompt with a draft. The agent had to weigh the user’s goal, missing context, and whether relevant work already existed.
I designed how it interpreted intent, asked clarifying questions, surfaced similar experiments, and offered alternatives before generating.
The ambition: “less chatbot, more product that understands the experiment lifecycle.”
Designing an agent that knows when not to generate
I designed the agent’s interaction model around one goal: knowing when to ask, recommend, or generate.
I defined the behaviors a useful agent needs, then criteria to judge them: whether it understood the request, surfaced the right evidence, asked the right questions, and knew when not to generate.
Describe your experiment design
The agent also needed to offer off-ramps. I designed how it presents them before the user moves on.
Open
Here is one of many ways the agent can present a simplified artifact to the user.
Tradeoffs
A central design question was when the agent should generate and when it should do something else.
Asking for context, surfacing a similar experiment, offering an off-ramp, or producing a draft: I treated that choice as part of the product experience.
It also shaped the evaluation criteria. Understanding the request and knowing when not to generate mattered as much as producing content.
Staying useful when users are away
Experiments needed to stay useful even when users weren’t actively in the product. I defined notification patterns around key lifecycle moments — pending actions, status changes, and results becoming available — so users could re-enter the experience with clear context and know what required attention.
Activity
Experiment completed
Just now
Checkout CTA test has finished collecting results.
Results Needed
Yesterday
Mobile navigation test is outperforming control.
Comment Added
3 mins ago
Sarah left a comment on your PDP experiment test.
I designed Activity as a persistent record of important experiment updates, helping users quickly see what changed and what needed their attention.
One visual language for outcomes
I defined a semantic color system to make experiment outcomes and states legible across the product — creating a consistent visual language for dashboards, Library, and Activity.
Impact
My work evolved the platform from experiment reporting into AI-powered decision intelligence.
I helped establish:
Decision-oriented analytics
AI-assisted experiment creation
Reuse of organizational knowledge
Reusable AI interaction patterns
A foundation for agentic workflows
How we measured success: time to create an experiment, throughput, recommendation use, and time from insight to action.
Results
My design work spanned decision-oriented analytics, AI-assisted experiment creation, and reuse of organizational knowledge, plus reusable AI patterns for broader agentic workflows.
The measures identified for success were:
Time to create an experiment.
Experiment throughput.
Recommendation use.
Time from insight to action.
Reflection
This work brought the moments around generation into focus: understanding intent, finding evidence, asking for missing context, and helping users choose a next step.
Those moments became explicit parts of the interaction model and the criteria for judging the agent.
Back
/ Apple