Challenge
Fashion discovery is visual and emotional. Most people can't type what they want; they know it when they see it. The platform's keyword-based search assumed the opposite. It didn't go well.
Retail Case Study
WhenaFashionMarketplaceStoppedGuessingWhatShoppersWanted
A curated fashion platform was losing customers not because their clothes were wrong, but because discovery was broken. Shoppers couldn't find what they were looking for in a sea of options. We rebuilt the experience with AI, and conversion surged 45% in three months.
Project evidence
Fashion & Retail
Confidential Fashion and Retail client
Anonymized
6 min read
Solution
We built four interconnected AI capabilities that solved the core discovery failures without requiring a full platform rebuild.
Result
45%
Conversion rate increase
Timeline
16-week delivery
4 delivery phases
Team
5 specialist roles
Cross-functional delivery
Evidence
Anonymized
Project and post-launch operating period
Client Context
Business Context & Telemetry
Our client was a curated multi-brand fashion marketplace with over 200 independent labels and strong Instagram traction. They weren't struggling to attract visitors; they were struggling to keep them. Shoppers would land, scroll endlessly through a disorganized catalog, and leave—not because the right product wasn't there, but because it was impossible to find.
Client Operating Profile
Scope, visibility, delivery context, and trust signals
09 signals
Executive Perspective
“We had great products and people who wanted them. The gap between those two things was just—the experience. Nobody could find anything easily.”
C&
Co-founder & CPO
Client
Confidential Fashion and Retail client
Reach
India-first, with shipping across South and Southeast Asia
Surfaces
3 platforms
Evidence
anonymized
Context Telemetry
Client operating details, platform surface area, and validation signals that shaped the work.
01
Client VisibilityConfidential Fashion and Retail client
Anonymized public case study
02
Company SizeGrowth-stage startup
03
Team Size22 people in-house
04
GeographyIndia-first, with shipping across South and Southeast Asia
05
Core PlatformsWeb, iOS, Android
06
Founded2020
07
Evidence Levelanonymized
08
Measurement WindowProject and post-launch operating period
09
Metrics NoteMetrics are shown as client-reported or operating-period outcomes; confidential identifiers are removed where required.
The Challenge
Great products, trapped in a catalog nobody could navigate.
Fashion discovery is visual and emotional. Most people can't type what they want; they know it when they see it. The platform's keyword-based search assumed the opposite. It didn't go well.
01
Search didn't speak 'fashion'
Typing 'flowy summer dress in earthy tones' returned zero results. The search bar only matched exact words in product titles, frustrating anyone who searched the way they actually think.
02
Random recommendations
The 'You might also like' section was just a bestseller list in disguise. It had no concept of the user's personal style, price range, or browsing history. It was noise, not help.
03
The 'Warehouse Sale' problem
With thousands of SKUs and basic filters, the catalog felt like a disorganized warehouse. Without intelligent style clustering, browsing was overwhelming and fruitless.
04
The 'I saw it on Instagram' gap
Shoppers often arrived with a visual reference—a screenshot, a photo—but had no way to use it. The platform was blind to the primary way its users actually discovered new styles.
05
High intent, zero discovery
Analytics showed healthy traffic and 'add-to-cart' rates, but a massive drop-off at purchase. Exit surveys had one theme: 'I couldn't find what I was looking for.'
Previous Attempts
They tried manual tagging, but it was inconsistent and didn't scale. They also bought a generic third-party recommendation plugin, but it just showed people what was already popular, creating a feedback loop of mediocrity.
"The founders had built a platform with genuine taste and curation. They knew their best products were hidden in plain sight, and it was professionally painful to watch users give up just before finding something they would have loved."
The Real Cost
The Approach
We didn't start with AI. We started by watching people shop.
Before discussing models, we ran user sessions with one simple prompt: 'Find an outfit for an occasion.' We watched silently as almost every participant gave up on the search bar and resorted to endless, hopeless scrolling.
Discovery & Methods
The patterns were immediate. Users used visual, descriptive language when they thought about clothes, but the website forced them to use rigid, keyword-based logic. The product wasn't broken; the language was. Our job was to build a smarter interpreter.
Moderated user sessions with diverse shopper profiles
Heatmap and session recording analysis of the browsing funnel
Synthesis of 1,000+ exit survey responses
Technical audit of existing (and inconsistent) product tagging
Fashion search is translation, not retrieval.
The gap was between how a shopper thinks about style and how a database thinks about products. We needed to build an AI that could translate a vague feeling or a visual idea into a specific set of SKUs.
Design Philosophy
The AI had to feel like a knowledgeable friend, not a sales algorithm. Every recommendation needed to feel genuinely helpful. That distinction drove every design and model choice we made.
Constraints Respected
- Small Team: The solution had to be maintainable by a lean engineering team, not a black box requiring a data scientist.
- Startup Budget: We prioritized fine-tuning open-source models over expensive proprietary APIs.
- Messy Data: The AI had to work with the client's existing inconsistent product data, not require a year of manual cleanup.
The Solution
An AI layer that made 200 brands feel like a personal stylist.
We built four interconnected AI capabilities that solved the core discovery failures without requiring a full platform rebuild.
Architecture Spec
Visual Search ('Search by Image')
Function
Shoppers upload any image—a screenshot, a photo—and the platform instantly surfaces visually similar products from the catalog.
Impact
It meets shoppers where their inspiration actually begins. The image becomes the query, ending the frustration of 'I saw something like this but can't describe it.'
Implementation Note
Fine-tuned CLIP model with a Qdrant vector similarity search layer.Tech Stack
Next.js & React Native
Fast, SEO-friendly storefront and unified mobile experience
Python (FastAPI)
Modular, independently deployable AI microservices layer
CLIP (Fine-Tuned)
Core model for visual and semantic style embeddings
Qdrant
High-performance vector database for similarity search
AWS (ECS, S3, CloudFront)
Scalable hosting with a CDN optimized for mobile image delivery
Design Decision
The Visual Search icon lives inside the search bar.
“If a feature is more than one tap away, it won't get used. Placing it next to the text cursor made it feel like an equal, natural option, not a gimmick to be discovered.”
Design Decision
Every personalized item has a 'Why you're seeing this' label.
“Transparency builds trust. When users understand why something is recommended ('Because you liked...'), they engage more, providing better feedback for future recommendations.”
Execution
Sixteen weeks, four phases, zero big-bang launches.
We deliberately shipped each AI feature sequentially, so the client's team could see real-world impact and provide feedback at every stage, not just at the end.
Delivery Timeline
Operational Log
1
Discovery & Architecture
Weeks 1–2User research, technical audit, and system design. Mapped the messy catalog data to understand what the AI needed to fix.
2
Catalogue Enrichment
Weeks 3–5Built and ran the AI pipeline to retroactively tag every product with clean, structured style metadata. This was the foundation for everything else.
3
Visual & Semantic Search
Weeks 6–10Shipped the new search capabilities to a 1,200-user beta group. Used their real-world queries to calibrate the models before the public launch.
4
Personalisation & Handoff
Weeks 11–16Launched the style feed and virtual try-on. We finished with a two-week knowledge transfer so the client's lean team could own and maintain the new AI layer.
Team Topology
Deployed Roster
1 × Engagement Lead2 × AI/ML Engineers2 × Full-stack Developers1 × Mobile Developer1 × Product Designer
Collaboration
Working Rhythm
We ran two-week sprints with a shared project space. Every sprint ended with a live demo on a real phone—not a PowerPoint. This built momentum and trust, making the client feel like a true partner in the build.
Course Corrections
Diagnostic Log
Friction Point
The source product data was a mess. Some brands had rich descriptions, others had none, relying solely on images.
Resolution
We designed the AI enrichment pipeline to lean heavily on computer vision rather than text analysis, ensuring a consistent baseline of quality even for products with zero description.
Friction Point
Virtual Try-On worked beautifully on dresses but struggled with structured blazers, creating an inconsistent user experience.
Resolution
Instead of shipping a half-baked feature, we launched Try-On selectively, enabling it only for garment categories where rendering quality hit a 95% confidence score. We were transparent about the limitations and built a roadmap to expand coverage.
Measured Impact
Three months later, the platform finally felt like the curated boutique it was always meant to be.
The hard metrics were excellent, but the team was most excited by the qualitative shift. Shoppers were no longer just scrolling; they were exploring, curating, and returning.
Primary KPIVerified Metric
45%
Conversion rate increase
in the first 3 months vs. the prior 3-month baseline
Session duration
3.2×
average time-on-site more than tripled as users explored
Repeat purchase rate
60%
up from 31%, driven by the personalized style feed
Qualitative Objectives Reached
- The internal content team was freed from manual tagging and redeployed to high-value brand curation and editorial work.
- Support tickets related to 'can't find a product' dropped significantly in the first month post-launch.
- Two new designer brands cited the platform's AI discovery features as their reason for wanting to join the marketplace, turning the tech into a B2B asset.
"I was skeptical that AI would make a real difference. What surprised me was how quickly it changed how our customers talked about us. We started getting messages like 'this app understands my style.' Nobody had ever said that to us before. That's what we were trying to build."
Co-founder & CPO
Fashion Marketplace Client
Key Learnings
Insights Gained
Valuable lessons and strategic insights uncovered through this project that inform our future work and architectural decisions.
01
Data quality is the unsexy work that makes AI shine.
The AI enrichment pipeline was the least glamorous part of the project, but it was the foundation for every success. Great AI on messy data produces stylish garbage.
02
Explainability builds trust.
Showing users *why* something was recommended ('Because you liked...') didn't make them suspicious; it made them feel understood. Transparency is a powerful product feature.
03
Don't ship features you can't do well yet.
Scoping virtual try-on to only the categories where it worked flawlessly was a tough call, but the right one. Shipping a half-baked experience to prove a feature exists does more brand damage than waiting.
Exploration
Capabilities & Archive
Running a marketplace and losing users somewhere between landing and buying? We've mapped that gap before. We know where it usually lives.
Let's Work Together
Great products can't save you if no one can find them.
We're not here to pitch you AI. We're here to understand what's broken in your discovery funnel and tell you honestly how to fix it.
"No buzzwords. Just a real conversation about your catalog and your customers."