Introduction

My name is Chris, I'm the co-founder and CEO of PopSize, an AI-powered platform that helps you find your perfect clothing size using real consumer data.

The Problem with Traditional Sizing

As you all know, finding your right size is not just a UX problem, it's a data problem. And the thing is that current solutions don't really work because they are not using real consumer data. So I walk you through why it doesn't work and what we are building with AI.

Why traditional solutions don't work

So first of all, why traditional sizing approaches fail? Current solutions use two main approaches. The first one is the BMI model, so body mass index, so the weight over height squared.

The problem with that is that it exaggerates the thinness of tall people and fatness in tall people. So that leads to an accuracy of around 60, 65%.

The other methodology that is used and quite in vogue nowadays is the body scan. The problem of the body scan is mainly technical because for the body scan to work your phone needs a lidar, needs a wide angle camera, the software needs to run very efficiently sometimes with no networks and that leads to very not not really a better accuracy in sizing.

The AI-Powered Solution

So with that, we thought about a solution on how we can find the perfect clothing size topping around 97% in the next couple months. And the approach is using shoppers' data and brands' data

Leveraging Shoppers' Data

To start with shoppers, thanks to our mobile app and also the widget that is plugged onto the brands and marketplaces websites, shoppers can input their closet information, including the brand, the type of item, the product, the size label, and how it fits. And with this, we compare it to all the brands, like we have billions of data points in our database now. We compare it to the size charts, the material, the fabrics, and also the item description with our model to understand it.

Vectorization and Prediction

1And all this data, we normalize it, and also we create object vector for shoppers and for brands' data points. And with this vectorization, we have created a modular architecture that combines embedding and prediction. So basically, we use the data that we collect that we put in a transformer-based model that compare all the vector that is used to then doing the size prediction. It's like a very lightweight predictor layer.

And so that in the end, you have a size prediction. The advantage of this approach is that you have good results even with few data. Also, our model can easily scale for a billion or 10 billion products.

Current Accuracy and Results

So currently, we are at 88% of sizing accuracy, and we target 97 by a couple months from now, compared with 75% to standard options, like the best one currently. And our model not only calculate your body measurement, but also understand how you fit, what you like, what your preferences, what your taste, and how you wear and buy clothes. And that's how we can increase the conversion rate online by 25%. We reduce the return rate by 35%.

Future Vision

And by 2027, we aim at reducing the CO2 emission by 20%. And PopSize in the future, we include also LLMs.

Dedicated Fashion LLM and AI Chatbot

1We are currently developing a dedicated fashion LLM so that when recommending you a size, the LLM can naturally explain to you why we recommend this size to you. And also in the future, starting next year, we'll create our AI chatbot so that for any need, you can ask the chatbot what clothes you should buy or what clothes you should pack for a for travel for an event or etc so thank you everyone for for listening and if you have any questions feel free