
How machine learning helped couples find their wedding aesthetic
An ML image feed that turned first week browsing into a reason to return
I led the design of an ML image inspiration feed for couples planning their weddings. Think Pinterest for weddings: it surfaces personalised visuals from onboarding preferences and natural language search.
Senior Product Designer, partnering with product and engineering.
I owned:
The interaction model for image discovery
Mobile UI and states
User flows and experiment constraints
Aligning the design to retention metrics
UX assumptions for the A/B test setup
[+2.7%] Week 1 retention
[+1.5%] Day 1 retention
[+67%] engagement with the feed
[9.2] images saved per user
Why this mattered
Wedding planning is visual, but our mobile app gave couples no easy way to explore and save inspiration early on. Struggling to put their vision into words, couples engaged less and were more likely to drop off in the first week, hurting retention and later conversions. We bet that a visual entry point on the homepage would help couples picture their wedding sooner and keep them coming back.
What we set out to achieve
Lift Week 1 retention by 2% (primary metric)
Improve Day 1 retention as a leading indicator
Hold or improve sign up to enquiry conversion
Protect First Contact Rate (FCR) quality
Research showed couples:
Feel overwhelmed early in planning
Struggle to describe their vision in text search
Need visual references to decide with confidence
Want to save and organise ideas easily
Couples who use visual inspiration in their first session will retain better because they:
Find value quickly
Connect emotionally with their wedding vision
Have saved inspiration to come back to
Could ML image search cut time to value and drive real engagement in the critical first week?
What I learned before designing
I studied Pinterest, Instagram saves and Houzz, and ran light testing with couples in the consideration phase. Three insights shaped the design:
Couples trust ML results when the first ones feel relevant
Refining visually (similar images) reduces cognitive load
Quick save is critical: couples won’t return to images they can’t find again
I placed the feed at the bottom of the homepage to:
Keep high intent flows at the top undisturbed
Catch exploratory browsing once couples are oriented
Create a Pinterest style scroll for engaged users
What we built, and the limits we built within
An ML image feed at the bottom of the mobile homepage, surfacing images from onboarding keywords and couples’ own prompts. A low effort way into discovery: like Pinterest, but built for wedding planning.
Technical
Mobile only, due to resourcing
UK only (A/B test scope)
Limited vendor profile interaction (MVP)
Design
No preset categories: ML and search only
Vendor cards showed name and location only, keeping sales pressure out of discovery
Had to sit alongside existing homepage modules without disrupting conversion
These limits kept the MVP focused and quick to validate.

How it works for the people using it
Natural language search
Built for conversational prompts like “rustic barn reception with string lights”
A simple field that expands on focus
Skeleton loaders while ML results load
Image cards
Large, thumb friendly tap targets
A persistent heart icon to save
Subtle vendor credit, so discovery never feels like a sales pitch
Similar image discovery
Tap any image for ML suggested similar images
Horizontal scroll for fast browsing
Vendor card at the bottom, opt in only
Visual design
Content first layout led by imagery
High contrast save buttons for accessibility
Optimised image loading for smooth scrolling
Fits the brand system while feeling inspirational



The calls I made, and why
1. Bottom of page placement. Protected high intent conversion paths; couples scrolling past the main CTAs were already in research mode.
2. ML first, no presets. Forced us to get personalisation right and set us apart from competitors’ generic categories.
3. Minimal vendor interaction. Kept sales friction out of the inspiration phase. We optimised for scrapbook saves (future intent) over instant vendor contact.
4. Similar image search. Couples often can’t describe what they want but know it when they see it. Visual refining also improved ML accuracy.
How we proved it
Control: homepage without the feed
Variant: homepage with the ML feed at the bottom
Succeeds at: +2% Week 1 retention
Fails at: a 2% drop in Week 1 retention
Guardrails: sign up to enquiry conversion must not fall; FCR quality must hold
Looking back
What worked
Natural language search met a real need for nuanced queries
Save to scrapbook gave couples a clear reason to return
Bottom of homepage placement balanced discovery and conversion
What I’d do next
Add light preset categories, since not everyone wants to type
Test richer vendor profiles for high intent couples
Explore desktop for couples who switch devices
This project shows how I design data driven AI features that balance user delight with business goals. A tightly scoped MVP with clear success metrics gave couples value straight away and reduced the risk of a bigger bet on ML discovery.
My approach: understand the user problem deeply, design for measurable impact, and ship iteratively.
