StyleIQ
Online fashion does not suffer from a lack of choice. It suffers from a lack of confidence.
Customers can browse thousands of products, but still struggle to understand what works together, what suits the occasion, what fits their budget, and why one outfit might be a better choice than another.
I designed and built StyleIQ to explore whether AI could behave less like a search engine and more like an experienced department-store stylist.
The product combines live retail inventory, structured product intelligence, AI stylist reasoning, editorial image generation, and a customer-facing recommendation experience in one working system.
The challenge
Section titled “The challenge”The first version appeared simple:
- understand the customer
- retrieve suitable products
- compose an outfit
- explain why it works
- generate an editorial image
The difficulty was that useful styling depends on much more than product titles and categories.
A stylist needs to understand silhouette, fabric, formality, colour, layering, season, fit, occasion, price, and how garments behave together. Retail catalogue data is rarely structured at that level.
The central product challenge became:
How do you create an aspirational AI experience without allowing the system to invent, substitute, or misunderstand the products it recommends?
Designing a grounded styling system
Section titled “Designing a grounded styling system”StyleIQ uses a live catalogue from a Christchurch department store rather than a fictional product set.
I built a Python ETL pipeline to discover, retrieve, clean, validate, and structure thousands of womenswear products from the retailer catalogue. That information is then passed through a product-intelligence layer that adds the attributes required for styling and retrieval.
The customer experience follows a clear sequence:
- capture style preferences, occasion, budget, season, and constraints
- retrieve suitable products from the live catalogue
- reason across the selected products as a complete outfit
- generate a stylist explanation
- create editorial imagery from the chosen items
- allow the customer to save and revisit the result
The objective was not simply to recommend attractive products.
It was to create complete looks that were:
- commercially available
- appropriate for the stated occasion
- coherent as an outfit
- explainable to the customer
- faithful to the selected garments
Three AI systems, three different responsibilities
Section titled “Three AI systems, three different responsibilities”One of the most important architecture decisions was not to rely on one model to do everything.
StyleIQ separates AI responsibilities across three distinct systems.
Product intelligence
Section titled “Product intelligence”The first system interprets catalogue products and produces structured attributes for retrieval and outfit composition.
This includes information such as:
- garment type
- silhouette
- fabric and weight
- colour
- formality
- season
- styling behaviour
- layering role
- product evidence and confidence
Stylist reasoning
Section titled “Stylist reasoning”The second system acts as the stylist.
It receives the customer brief and a controlled candidate set, then selects a complete outfit, checks compatibility, explains the recommendation, and returns structured product references.
Editorial image generation
Section titled “Editorial image generation”The third system creates the aspirational visual after the outfit has been selected.
This separation matters because the image is not allowed to decide the outfit. It must visualise an existing recommendation rather than inventing one.
The hardest design decision
Section titled “The hardest design decision”The hardest decision was balancing AI creativity with customer trust.
Early recommendation outputs could appear convincing while still combining products poorly or misunderstanding the retailer inventory. Improving the prompt alone was not enough.
I redesigned the system around stronger product intelligence, controlled retrieval, structured outputs, compatibility checks, and explicit separation between recommendation and image generation.
That changed the product from a visually impressive AI demo into a more credible retail decision-support system.
The customer can see a complete recommendation with:
- the selected products
- the total price
- a stylist rationale
- an editorial image
- direct access to each retail item
- a wishlist for saved looks
Design System
Section titled “Design System”StyleIQ was designed as a reusable product platform rather than a collection of individual screens. I established a cohesive visual language spanning typography, colour, spacing, components and interaction patterns, allowing new features to be built consistently while maintaining the premium retail aesthetic.
Building the product end-to-end
Section titled “Building the product end-to-end”StyleIQ was designed and shipped as a working product rather than a static prototype.
The system includes:
- React and TypeScript front end
- Python catalogue discovery and ETL
- structured product contracts
- multimodal AI enrichment
- retrieval and outfit-composition logic
- Supabase persistence
- recommendation and event history
- asynchronous editorial image generation
- model, prompt, retrieval, and catalogue versioning
- evaluation and QA workflows
- production deployment
The internal interface also exposes the system behind the customer experience.
Each recommendation records operational information such as:
- model version
- prompt version
- retrieval version
- product set
- token usage
- latency
- image-generation state
- recommendation status
This made it possible to evaluate the product as a system rather than judging individual outputs by appearance alone.
Outcomes & reflection
Section titled “Outcomes & reflection”StyleIQ is now a working AI-native retail prototype connected to a real department-store catalogue.
It demonstrates how product design, data architecture, AI reasoning, and front-end implementation can work together to create a customer experience that feels polished without hiding the complexity underneath.
What the product now enables
Section titled “What the product now enables”- personalised outfit recommendations from live retail inventory
- structured product intelligence across thousands of catalogue items
- grounded outfit composition using controlled product retrieval
- stylist explanations linked to real products and prices
- editorial imagery generated from the selected outfit
- saved recommendations and wishlist behaviour
- persistent recommendation history and evaluation metadata
- a complete customer and operational interface
Reflection
Section titled “Reflection”The most important lesson was that the AI model is only one part of an AI product.
Recommendation quality depended just as much on catalogue quality, product understanding, retrieval, system boundaries, evaluation, and interface design.
Building StyleIQ also reinforced the value of working across disciplines. I began with a customer problem and learned the front-end, backend, data, and AI capabilities required to ship it independently.
The result is not simply an AI stylist interface.
It is a complete product system designed to turn real retail inventory into recommendations customers can understand, trust, and act on.