Google · Product Designer · 2020 – 2021

Building foundations for AI at scale

How I unified design across Google's experimental AI and ML teams — creating foundational patterns that scaled from prototype to production as these products matured.

50M+
users at scale
5
ML/AI teams unified
85%
pattern adoption

The challenge

Five different teams. Five inconsistent approaches. An engineering-led org where design took a backseat to functionality.

Google's emerging AI products division was brilliant at engineering but struggling with design coherence. Each team had developed its own patterns, components, and visual language in isolation — users were confused moving between products, and the brand experience was fragmented.

We needed to prove that great design could enhance technical excellence, not slow it down — while navigating the ambiguity of AI products still finding their market fit.

Research & discovery

Understanding the fragmentation before building unity.

23
different button styles found
5
conflicting design approaches
25+
stakeholder interviews
2 wks
embedded shadowing ML teams
📊

Component inventory

Catalogued UI elements across 5 ML products, identifying patterns unique to AI interfaces — confidence UI, model states, training views, pipeline visualization.

🔍

Workflow analysis

Shadowed teams for two weeks to understand how ML engineers actually interact with their tools, and established baseline metrics for design velocity and consistency.

The journey

Leading through ambiguity to create order from chaos — 14 months.

Months 1–2

Discovery & alignment

Interviewed 25+ ML engineers and designers. Discovered deep mistrust between design and engineering. First design review was brutal — “this will slow us down by months.”

Months 3–4

Failed prototype

Created a rigid design system. Teams rejected it immediately. I was solving the wrong problem — they needed flexibility for experimental AI features, not rules. Pivoted to a modular approach.

Months 5–8

Building trust

Embedded with ML engineering teams. Co-created components WITH them, not FOR them. First successful implementation in the AutoML team.

Months 9–11

Scaling success

Other teams saw AutoML's velocity increase and adopted voluntarily. The system evolved from real usage — engineering became its biggest advocates.

Months 12–14

Full implementation

Rolled out across all 5 teams with a design review process and self-service tools. Teams shipped 45% faster with better quality.

The audit — five teams, five languages

The component inventory made the fragmentation impossible to argue with: 23 button styles across five products, then one shared foundation.

ml-foundations · audit-and-system.figDrafts · Design
68%
beforeButton audit — 5 products
Train
TRAIN MODEL
Run training
→ Train
Start
…18 more variants
afterFoundation — one system
Train model
Cancel
1 button · 4 sizes · 3 emphases
85% adoption across 5 teams
AI-specific components
Confidence indicator · 87% ▮▮▮▮▯
● Training● Deployed● Failed
Pipeline view · ingest → train → eval
Threshold slider · precision ⇄ recall
Recreated for this portfolio — original files remain on Google infrastructure. The audit findings and component set mirror the real system.

The transformation

Before: 5 design languages

  • 5 inconsistent approaches
  • Limited component sharing
  • Siloed team decisions
  • Inconsistent user experience
  • 2-week design cycles

After: 1 unified system

  • Shared design foundation
  • 67 foundation components
  • Weekly design syncs
  • 85% core pattern adoption
  • 1-week design sprints

“This design foundation let us experiment rapidly with AI interfaces while maintaining quality.”

— Design Lead, Google Cloud AI

Key learnings

Start with empathy, not enforcement.

Understanding why teams made different choices was crucial. Authority comes from value, not title.

Make it easier, not mandatory.

The system succeeded because it made teams' lives easier. Adoption through attraction beats compliance.

Embrace engineering culture.

By speaking their language and understanding their constraints, design became a valued partner in ML innovation.

Design for the unknown.

Building flexibility into the system let it accommodate AI features we couldn't yet imagine. Ambiguity is a design constraint.