I led design of this AI product, Taiwan's first Google Gemini AI in retail
Type
E-commerce, Conversational AI
Role
UX Designer
UX Researcher
Team
UI Designer
UX Researcher
Front-End Engineer
Back-End Engineer
PM
Duration
2023.09 - 2023.12
(Design Phase)
Client
Carrefour
Platform
Mobile Web App
Stakeholder
CTO
Data Team Lead
Dev Team Lead
My design delivered measurable business impact
Wine sales Increase after 2 months of launch
Users added wine to cart in ≤3 searches
Google Gemini AI application in Taiwan retail
"AI sommelier can give you the right wine in just a few seconds."
Henry Ting
Digital Technology Officer, Carrefour Taiwan
Research
To ship fast, I explored mature AI applications and market demands.
AI application assessment
I analyzed 14 deployed AI applications to extract workable design and interaction patterns.
Sommelier and seller interviews
We interviewed 2 liquor store owners and 1 Carrefour sommelier to understand wine selection challenges.
Online forum threads analysis
I analyzed wine community forums to uncover beginner barriers and purchasing behaviors efficiently.
Key Findings
We uncovered an opportunity and a pain point
Opportunity: Sommelier's 3-step guidance framework
User pain point: Knowledge barriers blocked shoppers
Shipped in 2025
Sales Cloud System
I shipped key quarterly updates with engineers and accelerated team efficiency by normalizing design consistencies
Role
Product Designer
Team
PM
Design Director
Senior Designers
India-Based Engineers
Platform
Salesforce Sales Cloud System
Duration
2025.05-08 (3 months)
Key Feature 1
Expert-informed AI conversation flow
Before
Customers overwhelmed by selection and buy nothing
Carrefour's extensive wine selection intimidated beginners who lacked the knowledge to navigate options. Without accessible expertise, customers left stores without making purchases.
Challenge
How to provide personalized recommendations, manage AI costs, and meet tight deadlines?
Generating full AI responses and translating them consumed too many tokens, driving up costs and slowing response times. The MVP timeline didn't allow time to extensively test and guarantee AI predictability.
Key Decision 1
I translated sommelier expertise into step-by-step conversation flows
I transformed sommelier expertise into guided conversation flows with choice-based interactions. This approach made professional wine guidance accessible to beginners while ensuring accurate recommendations, controlled costs, and on-time delivery—all without requiring model fine-tuning
Key Decision 2
I designed choice-based interactions to ensure accuracy and delivery timeline
To guarantee accuracy and meet the 4-month deadline, I designed structured choice flows instead of free-form input. Guided selections ensured reliable recommendations without model training, while reducing decision anxiety for wine beginners.
The Design
I shipped expert-guided conversation flow for reliable MVP accuracy
I transformed sommelier expertise into guided conversation flows with choice-based interactions. This approach made professional wine guidance accessible to beginners while ensuring accurate recommendations, controlled costs, and on-time delivery—all without requiring model fine-tuning
Key Feature 2
AI-simplified wine profiles and illustrations
Before
Expert wine language confused and excluded beginners
Technical terms like 'pencil lead,' 'sundried rocks,' and 'medium-bodied palate' excluded wine beginners rather than guiding them. This expert language reinforced intimidation, turning potential customers away without purchases.
Challenge
How to simplify wine language without sacrificing accuracy or credibility
Expert wine language ensures accuracy but excludes beginners. I needed to make descriptions accessible without oversimplifying—maintaining enough detail for trustworthy description while removing intimidating jargon that prevented purchases.
ACCESSIBLE LANGUAGE
Translate expert terminology into everyday words
EXPERT CREDIBILITY
Maintain trust and authority while being approachable
Description accuracy
Ensure descriptions genuinely match wine favors
I translated wine expertise into visual scales and everyday language
Simple descriptors like 'micro-bubbles' and 'fruit aroma' paired with numerical scales transformed complex wine attributes into scannable, relatable information.
The Design 2
Approachable wine profiles with everyday descriptions and visual aids
I combined three elements to make wine expertise accessible: numerical scales (Acidity: 3) for quick comparison, everyday descriptions ('fresh and smooth' instead of 'medium-bodied'), and food pairing icons (beef, cheese, sweets) that relate to real dining experiences
Key Learnings
AI flows can evolve as users became more confident and direct
After launch, user behavior evolved. Initially, beginners needed structured guidance through choices, but as they became familiar with the system, they started entering complete requests upfront. I adapted by allowing direct input alongside guided flows and added 'I don't know' options for uncertain beginners
Multilingual AI requires cultural adaptation, not just word translation
Direct Chinese translation reduced recommendation accuracy significantly. I learned that multilingual AI needs culturally appropriate concepts and examples to maintain accuracy and relevance.
Project 1
🏦 Business Challenge
Clients' Sales Managers spent 100+ hours manually updating and mapping two databases every season
🙌 My Contribution
I shipped intuitive field mapping features
🚀 The Impact
Sales Managers' work now reduced from hours to minutes
Project 2
🏦 Business Challenge
Mission-critical buttons cut off on tablets and appears differently across 18 products
🙌 My Contribution
I identified critical responsive issues and drove system-wide normalization
🚀 The Impact
From 40+ inconsistencies to one unified system, ready for AI integration
Reflection
