more dresses for computer scientists

Code your passion! Learn to code with fun projects about fashion, pets, beauty. Empowering girls to create their own tech world! the world that does not shy from girls fully embracing what they are

Comment

How it started

I was watching a video on YouTube, where the author was presenting a recommender system as an example for system design interview. He was explaining how recommender system work, based on tracking user preferences. As an example of a system that understands the user well enough to highlight products he or she will like he had a system “that will not recommend a dress to a computer science person”. I could not keep silent. I wrote a comment. My Comment

But one comment is not enough to change a deeply rooted world view where dresses and computer science belong to two different worlds. This is something that young girls get very early, that computer science is not for girls just because they are girly, too girly. And they are heavily penalized as a consequence: technology is power, and women are denied access to this power from early on. I want to challenge this prejudice against «girly things » in my project, so that young women also could get access to the power that the tech world provides.

Proposal

I would like to propose creating an educational platform where young girls can learn coding while being fully themselves as girly girls. On this platform, they should be able to code projects that are relatable for young girls, however girly their interests are: fashion, beauty, pets, music, languages, travel… The purpose of this platform is to empower young girls to enter computer science while embracing fully their female identity and interests.

What the platform will include ;

  • Selection of projects (see below)
  • Coding assistant (LLM, e.g Gemini) + human in the loop (tutor)
  • Support with the girls’ own project ideas: they can start with asking the LLM (Gemini) for help, and get more support from a human tutor
  • Social network for students, with the possibility to follow people, showcase their work, share repos and get kudos (stars, flowers, jewels, crowns, puppies .. whatever !)

Possible project ideas

  • swap the hairstyle in your photo with a hairstyle from another photo (or lipstick color)
    • try on clothes from another photo on your own (virtual try on)
    • animate images: e.g . fashion runway images, unicorns
    • a chatbot to answer questions about fashion, beauty, pets
    • database of your closet (or of cooking recipes)
    • recommender system – not just books and movies, but also tv series and Tik Tok/Instagram users
  • analyze popularity of a cause and a related sentiment – e.g. climate change or fashion item/style
    • search Instagram/Tik Tok users to follow based on their visual content while making search requests in natural language This list is not comprehensive, but just gives some starting point.

Difference from Scratch

Scratch is an example of a similar platform, however, I would like to highlight several important differences from Scratch in my proposal : - it is targeting topics that are considered «girly », specifically - it should teach the concepts of real world applications : chatbots, LLM agents, image processing, databases, recommender systems - it should also teach a real programming language, e.g . python - it should go beyond just teaching the girls to use the apps, it should empower students to create apps Of course, creating age appropriate content and delivering it successfully requires pedagogical education or just vocation, which unfortunately I cannot claim I have. What I want to propose is how it could look like, with project ideas and AI based assistance. I hope that these projects will find their way in education for young girls, because someone with the right background can make it into some real age appropriate educational content.

Submission

As an example of what the projects for girls could be like, I am providing code for some use cases that employ AI (Geminin in Google AI Studio) for fashion recommendation. In the example, I used the photos of some of my favourite fashion infuencers : @jeanne_andreaa, @pauluschkaa, @stephaniebroek and @lara_bsmnn.

Use cases

The code covers several use cases.

Current trends from photos

  • you can download a bunch of images from Instagram and ask Gemini to provide a summary of the current trends;

An outfit in the style of your favourite influencer

  • you can download some photos of an instagram user and ask Gemini to suggest an outfit recreating her style. Then, you can compose this outfir out of the items found on Vinted.

Recreate an outfit with items from Vinted

  • you can upload a photo and recreate the outfit on Vinted (the photo will be described by Gemini in such a way that it will be possible to search for single elements on Vinted).

Get an outfit suggestion based on your request

  • you can download a bunch of images from instagram and create a dataset where you can search for outfit ideas based on natural langage requests, such as outfit in a certain style (e.g. urban chic), outfit containing certain elements (e.g. ankle boots) or colors (e.g. olive green).

For this purpose, a selection of images was described by Gemini; it identified the outfit style and its key elements. These descriptions are then embedded and out in a dataframe that contains image identifiers along with textual descriptions. When a user asks for a suggestion, her request is also embedded, and an image is retrieved whose style and description are closest to the request.

Educational purpose

In this project, girls can learn several things that constitute modern AI : multimodal models, natural language embeddings, retrieval. At the same time, they are encouraged to pursue their girly interest as well, by downloading images from Instagram and searching for fashion item on Vinted (the young generation is often very conscious about their consumption, so many would prefer Vinted to Amazon).

Learning and challenges

I am very passionate about both AI and fashion myself, and I am constantly looking for ways to marry my both passions and to share it with others. So, I was very inspired by the possibility to participate in this Hackathon. Unfortunately, it fell at the time when I was changing jobs, so it turned out, I had less time and energy that I needed. I would have loved to include much more in this submission, but still, what I managed tp do fr this submission, was a great learning experience for me. Here are my key takeaways:

  • Gemini models (both Gemini-1.5-pro and Gemini-1.5-flash) are surprisingly good at understanding fashion. They could name styles correctly and identified outfit components very accurately.
  • However, as all LLMs, they might be inconsistent in their outputs. So, to obtain quality textual descriptions for fashion photos, multiple runs of an LLM might be needed, so that you can aggragate the final description over several LLM outputs.
  • Same sort of aggregation might be needed because, for instance, several style labels can be applied to an outfit, and the LLM might not capture them all in a single run.
  • To make a better application for real life use cases, more work will be needed to correctly filter Vinted search results (for country, size, price range, item condition; for instance). Same refers to handling Instagram content: more and better tooling needs to be implemented.
  • My next step will be creating a Fashion Agent that will reason and use tools to respond to more sophisticated search and recommendation requests.
  • It was great fun! Expect more of such projects from me in the next months!

Built With

Share this project:

Updates