top of page

AI Stylist

AI tool to create personalized outfits using your photos or past orders.
image.png

Background

  • Myntra is the leading fashion e-commerce platform in India which it has maintained by constantly innovating.

  • But in the past few years, a lot of new e-commerce platform has come up which has started to threaten its monopoly. Myntra is losing its market share every year.

  • In order to create a market differentiator, building outfits has been an obvious choice in the fashion e-commerce space.

How it started

In an emergency leadership meeting, it was decided that we would build an outfit creator through AI using our existing AI models.

  • Similar model: This model ​outputs similar products to a particular product. 

  • Cross-sell Model: This model outputs complimentary products to a particular product.

  • Outfit Model: Keeps one style ID as a reference and recommends other article types around it.

image.png

My role and the team

This was a big project that required cross-functional collaboration among team members from different departments.

​

  • Graphic designer and motion designer: Worked on the FTUX of the feature.

  • UX researcher: Worked on the initial research.

  • Product Manager: Worked on gathering initial requirements and managing cross-functional collaboration.

​

My Role

I was the sole product designer for this project. My role included:

  • Gathering initial requirements and briefing the UX researcher about things that we want to know before we start the project. I also helped the UX researcher in analyzing and formulating insights from the interviews

  • The feature ideation, interactions, and the UI.

  • Prototyping and usability testing.

  • Briefing and managing the graphic team about FTUX requirements.

  • Tech walkthroughs and design UATs.

  • Evaluating and helping in correcting the AI model after release.

​

Research

We conducted some user interviews across our user groups to understand general needs around buying outfits.

​​

Users (Genz and late millennials)

  • Genz are more likely to experiment and hence buy clothes and create outfits.

  • People older than 35 now know what they like and are more brand-conscious and hence are less likely to experiment with different outfits.

​

Insights 

  • People are actually looking for looks and outfit inspiration. They do google searches or search on Instagram and Pinterest for outfit inspirations.

  • People do not buy complete looks. They try to buy clothes and then mix and match them with things they already have.

  • Budget depends on the occasion, if it’s a close wedding then they will spend more but if it’s only like a birthday party then they will have less budget and will try to buy only a few products to match it with what they already have.
     

Research conclusion ​

  • Try to create looks from the products they already have.

  • They don’t know what a good look is, so give users validation.

  • When choosing a look, people may not like the particular part of that outfit. Give users the ability to tweak the outfit based on price, brands, etc.

​

Constraints

  1. Since I was constrained by the DS models we could only design around the outputs of those models.

  2. Model Output

    • The outfit is created around one anchor product.

    • The anchor product could only be top-wear and bottomwear or dresses etc.

    • The Number of output article types could only be either 3 or 4 (top wear, bottom wear, footwear, 1 or 2 accessories)

    • All the outfits will be generated once, so the user cannot create outfits on runtime since running a model takes time.

​

Inspiration

Looked at some of the foreign fashion apps to see if they are doing something similar to get some inspiration. Also, looked dribble and Pinterest to get ideas.

​

Iterations

And so the ideation started. I started with an intial proposal to present how the flow will look like.

Iteration 1​

From closet

Landing Page

Closet

Outfits - Iteration 1 

OR

Outfits - Iteration 2

By taking a photo

Landing Page

Camera

Taking photo

Outfits created from the items in the photo

Next Step

We showed the flow and some rough screens to the senior leadership. They liked the flow and we got the go-ahead to move ahead and refine the flow further.

​

Identifying issues with current screens:

  • Landing Page: 

    • The explainer is too long which is pushing the CTAs outside of the first fold.​

    • The landing page is not ideal for new users as they will not have anything in their closet.

  • Customization:

    • Currently, users can click on the product in the bottom tray to customize the outfit. However, users are more likely to buy individual products rather than the entire outfit.

    • So, on clicking the product, we need to give users the ability to look at each product in more detail without going off the screen.

    • For customization, we will need to find another way. 

Iteration 2

Outfit Page

Customisation page

Landing Page

Select Size

View product

Prototype

Next Step

Everyone was aligned with this and required little tweaks.​

 

Scrolling the outfits:

  • Instead of a horizontal swipe scroll, we decided to convert it to a vertical swipe scroll because we were planning to produce infinite outfits.

Iteration 3 (Swiping vertically)

I started experimenting with the vertical card swipe to browse outfits.

The goal:

  • To keep the user focused on the outfits rather than all the details.

  • Keep the height of the entire page under one fold (no scrolling).

Iteration 1

Problem:

There is too much clutter in the card. It's hard to focus on the outfit.

Iteration 2

Problem:

Crucial information is not visible. Users will need to click every time they want to see more information.

Iteration 3

Problem:

A good balance but users were still trying to scroll the entire page. 

Final Iteration

  • After the above three iterations, we decided to go ahead with the third one as it was a good balance between one and two.

  • However, after we did a quick usability test around the office, users were trying to scroll the entire page.

  • To solve this, I created the onboarding as shown in the final iteration. After which users had no problem scrolling the outfits.

Final Flow

Launch, problems and enhancements

After a lot of little tweaks and covered corner cases, we launched the feature.

​

Problems post-launch:

  • The outfits were not good enough. The Pairings were not correct. For e.g. Shirts were being paired with slippers which is wrong.

​

Enhancements:

  • To solve the above problems we manually excluded many article types when paired with a particular article type. E.g. we excluded slippers and pyjamas for formal shirts. 

  • We started saving the outfits that users were customizing and buying for input data to improve our outfit model.

Impact matrix and post launch numbers

  • We were looking at the following numbers to measure the success of the feature.

    • Engagement: How many people are coming and engaging with the feature​

    • Cart size: Increase in the number of products being added to the cart.

    • ARPU: Increase in the revenue per user.

  • Impact​

    • There was larger-than-expected interest in the feature.​

    • There was a huge uptick in the items being added to the cart.

    • The ARPU increase was not as much as we expected and did not increase proportionally to the number of items added to the cart. The reason was that people were adding more items to the cart than their budget allowed.

Conclusion and Next Steps

  • Overall it was a successful feature that showed promising results.

  • It was showcased in Walmart's global town hall meeting as one of the innovation projects.

​

Next Steps:​

  • Making the outfit model is the top priority for the feature.

  • Feature to save the outfits.

  • Showcasing outfits in different places. e.g. In the product display page.

Thanks for reading

bottom of page