Inspiration
Online stores usually try to sell a product before they understand the customer.
A skincare merchant may advertise the same serum, moisturizer, or treatment to thousands of people with completely different concerns. After clicking the ad, the shopper lands on a product page and is expected to decide whether that product is actually right for them.
I wanted to reverse that relationship.
Instead of:
Ad → Fixed Product → Customer Guesses
Quorlyx Loop enables:
Guided Experience → Customer Understanding → Relevant Products
The idea behind the YouCam integration was therefore not simply to add a skin-analysis button. I wanted visual AI to become part of a larger commerce decision system.
What it does
Quorlyx Loop is a self-hosted customer-experience and visual-commerce platform that connects:
- guided assessments
- Perfect Corp. YouCam Skin AI
- real store catalogs
- availability-aware product filtering
- explainable product recommendations
- generated landing pages
- customer assistant experiences
- advertising workflows
- analytics and behavioral signals
For the YouCam hackathon, I extended Quorlyx Loop with a visual skincare workflow where customer answers and optional photo analysis can influence which real products from a merchant's catalog are recommended.
A typical customer journey looks like this:
- A shopper enters a guided skincare experience.
- They answer a short assessment about their concerns and goals.
- They can optionally consent to photo analysis.
- Quorlyx sends the visual request to Perfect Corp. YouCam through the backend.
- YouCam Skin AI analyzes supported cosmetic skin concerns.
- Quorlyx normalizes that visual evidence into its internal recommendation model.
- The recommendation engine evaluates products from the merchant's actual catalog.
- Ineligible or unavailable products are filtered out.
- The shopper receives relevant products with explanations for why they were selected.
- They can continue toward the appropriate product or shopping action.
The photo is therefore not the destination of the experience. It becomes another source of evidence used to make a better retail decision.
How I used YouCam
The primary hackathon integration uses Perfect Corp. YouCam Skin AI.
Quorlyx communicates with YouCam from the backend so the merchant's provider credentials are never exposed to the customer-facing browser.
The implementation handles the asynchronous YouCam workflow, including:
- provider file/task operations
- task creation
- polling for completion
- supported analysis actions
- response normalization
- photo-quality failures
- transient provider failures
- API-unit quota controls
- idempotency to avoid unnecessary duplicate billable requests
- diagnostics and safe error handling
Depending on the enabled YouCam capabilities, the Skin AI integration can work with cosmetic signals such as:
- acne
- wrinkles
- moisture
- dark circles
- age spots
- redness
- radiance
- oiliness
- pores
- texture
Quorlyx also contains support for Facial Color Analysis and a visual capability path for Apparel Virtual Try-On.
For Apparel VTO, I deliberately chose a fail-safe design: a try-on requires both a customer photo and a validated garment asset from the merchant catalog. If that asset mapping is not available, Quorlyx shows a clear fallback instead of pretending a virtual try-on was successfully created.
From visual analysis to product recommendations
This was one of the most important parts of the project.
Displaying a skin-analysis result by itself may be interesting, but it does not automatically create retail value.
Quorlyx takes the next step.
The recommendation system can combine:
- customer assessment answers
- normalized visual evidence
- catalog metadata
- product eligibility
- availability/inventory information
- merchant-defined constraints
before ranking products.
Hard filters are applied before recommendation scoring, which helps prevent an unavailable or invalid product from being recommended simply because it appears to have a high personalization score.
Recommendations can also include reasons and evidence sources, making the result easier for the customer to understand.
The goal is not:
"Your skin score is X."
The goal is:
"Based on what you told us, what we observed, and what this store actually sells, these are the products that make the most sense for you — and here is why."
Merchant experience
Quorlyx is also designed from the merchant's point of view.
Inside Customer Studio, the merchant can build and configure guided customer experiences instead of manually wiring together separate quizzes, landing pages, recommendations, and visual tools.
The visual-commerce landing-page system supports reusable experience presets including:
- Skin Match
- Virtual Try-On
- Complete Look
- Quiz Only
- Standard Page
These presets use the same underlying experience infrastructure rather than functioning as completely separate products.
This allows a merchant to move toward a workflow such as:
Store → Assessment → Visual Experience → Recommendations → Landing Page → Promotion
A different advertising model
One of the ideas I wanted to explore with Quorlyx Loop is using visual personalization before the merchant decides which product to promote.
Traditional e-commerce advertising often looks like:
Ad → Product
But that forces the merchant to choose the product before knowing anything about the person clicking.
Quorlyx makes another model possible:
Ad → Guided Experience → Customer Understanding → Personalized Products
For example, instead of advertising one skincare serum to everyone, a merchant could advertise:
"Find the skincare routine that fits your needs."
The landing experience can then use assessment answers and optional YouCam analysis to determine which products from that store should actually be presented.
This connects customer acquisition, product discovery, personalization, and merchandising in one journey.
Privacy and safety
Because customer photos are sensitive, I designed the visual-commerce flow around explicit consent.
Photo-enabled experiences require the visitor to approve image processing before analysis begins.
The customer can also continue through an assessment-only path if they do not want to provide a photo.
The backend communicates with YouCam, so provider credentials are not exposed to the browser.
The runtime also distinguishes between different types of failures.
For example, a customer's unsuitable photo is treated differently from an actual YouCam service outage. Correctable photo problems can produce retry guidance instead of incorrectly triggering provider-level failure behavior.
The experience uses cosmetic, non-diagnostic language and does not present the visual analysis as medical diagnosis.
How I built it
Quorlyx Loop uses a multi-part architecture:
Backend
Built primarily with Python and FastAPI.
The backend handles:
- YouCam provider communication
- visual-task lifecycle
- customer experiences
- catalog access
- recommendation logic
- privacy controls
- quotas
- analytics
- merchant configuration
Merchant interface
Built with React and Vite.
Customer Studio gives merchants a visual interface for configuring experiences and other parts of the customer journey.
Visitor runtime
Built with TypeScript for the customer-facing experience that can run on generated landing pages and storefront surfaces.
Data layer
The self-hosted build uses SQLite as its local authoritative data store.
Challenges I faced
One of the biggest challenges was discovering that a production visual-AI workflow is much more complicated than:
image → API → result
A real customer workflow has to consider:
- whether the customer consented
- whether the photo can actually be analyzed
- asynchronous task states
- supported provider actions
- API quota
- retries
- temporary assets
- network/provider failures
- product availability
- how visual evidence should affect recommendations
Photo failures vs provider failures
One particularly important problem was failure classification.
A shopper might submit a photo where the face is too small or unsuitable for analysis.
That is a customer-input problem.
It is not the same as the YouCam API being unavailable.
If both types of errors are counted the same way, several bad photos could incorrectly trigger the provider circuit breaker and disable visual analysis for subsequent customers.
I solved this by introducing explicit classification for correctable photo issues.
Known photo problems can return retry guidance to the visitor without being counted as genuine provider outages.
Unexpected provider errors continue through the normal reliability path.
Another challenge: making AI results commercially useful
The second major challenge was deciding how much authority visual analysis should have over recommendations.
I did not want:
YouCam result → AI invents a product
Instead, the recommendation system is grounded in the merchant's real catalog.
Visual evidence contributes to the decision, but product eligibility and catalog constraints remain deterministic.
This separation made the system more predictable and safer for real merchants.
What I learned
The most important thing I learned is that the value of visual AI often comes from what happens after the API response.
The API can produce powerful information, but a complete retail product also needs to answer:
- What should the customer do with this result?
- Which products are actually available?
- Why is one product more relevant than another?
- What happens if the customer refuses a photo?
- What happens when an image cannot be analyzed?
- Can the merchant use the same intelligence across landing pages and advertising?
- Can the experience remain understandable to a first-time customer?
That changed how I thought about the YouCam integration.
Instead of treating Skin AI as a standalone feature, I treated it as an intelligence layer inside a larger commerce workflow.
What changed during the hackathon
Quorlyx Loop existed before this hackathon as a conversion-intelligence and marketing workflow platform.
During the hackathon submission period, I significantly extended it with a new visual-commerce architecture, including:
- live Perfect Corp. YouCam backend integration
- YouCam Skin AI support
- Facial Color Analysis support
- visual capability routing
- consented customer photo workflows
- asynchronous provider task handling
- normalized YouCam evidence
- catalog-grounded visual recommendations
- inventory-aware filtering
- explainable recommendation evidence
- photo-quality error handling
- YouCam quota controls
- idempotency
- provider diagnostics
- failure/circuit-breaker behavior
- Skin Match experiences
- Virtual Try-On experience infrastructure
- Complete Look experiences
- visual-commerce landing-page integration
- improved Customer Studio workflows
These changes transformed the existing system from primarily conversion intelligence into a platform that can connect visual understanding directly to product discovery and customer journeys.
What's next
The next step is expanding validated garment-asset mapping so Apparel Virtual Try-On can be activated reliably across larger real-world merchant catalogs.
I also want to continue connecting the same Customer Studio workflow across the complete commerce lifecycle:
Assessment → Visual Analysis / Try-On → Explainable Recommendation → Product Action → Campaign Feedback
The broader goal is simple:
Understand the shopper first. Recommend the product second.
Built With
- ai
- api
- color
- corp
- facial
- fastapi
- perfect
- python
- react
- skin
- sqlite
- typescript
- vite
- youcam
Log in or sign up for Devpost to join the conversation.