Product Labs🔬: Decoding “Maya - Myntra’s AI Fashion Assistant”

Love weddings, but hate the shopping? Maya is here to save the day 💃🏼

What’s exciting here is that Myntra is one of the early adopters of AI in the Indian landscape. It’s been a year since Myntra launched Fashion GPT. This is when other players in this space are yet in the RnD phase and yet to roll out some cool stuff.

What’s in it for you?

Before diving deep into what (or who 😋) is Maya, for the uninitiated lets tell you about Myntra. Established in 2007, Myntra (part of the Flipkart group) is an Indian fashion e-commerce company headquartered in Bengaluru, Karnataka, India.

Well, we are puzzle pieces - what do we know about dressing up for a wedding or looking fab for vacation? But the word around the block is sifting through a never-ending sea of options can be extremely frustrating aka “The Paradox of Choice”.

Enter Maya, your ultimate shopping bestie! She's got 2.3 million styles in fashion, beauty, footwear, and home décor at her fingertips, and can read your mind (almost) to find exactly what you're looking for. Stuck on a gift for your friend? Need a killer outfit for a night out? Maya can chat with you, ask questions, and curate a selection of products that are totally for you. ✨ Plus, she remembers your conversations, so the more you chat, the better she gets at recommending things you'll love.

The one letdown about Maya is how the product discovery is shrouded inside the Xplore option. Unless a user is explicitly aware of the existence of Maya, they could be regular users of Myntra app (like Trouble 😆) and be blissfully unaware.

Maya is a step up from Myntra’s previous AI offering Fashion GPT - while Fashion GPT could take your request and generate collections of results (and honestly, it isn’t uncommon to find the recommendations underwhelming or biased) Maya being a chatbot can converse with you in a multi-turn manner to understand exactly what you’re looking for before presenting her results.

What’s the tech under the hood?

Maya uses Azure’s OpenAI services and ChatGPT for new product discovery (version unknown 😕) Given that Fashion GPT was powered by ChatGPT-3.5, safe to assume Maya uses ChatGPT-3.5 or ChatGPT-4.0. It personalizes your shopping by translating your desired look into prompts using ChatGPT. Maya then asks clarifying questions to refine your intent, enabling Myntra's search engine to curate matching items. The model's internet-trained datasets ensure tailored recommendations. Pretty neat, uh?

What is intriguing?

Surprise, surprise!✨

Maya wasn't initially intended to be a chatbot at all! Initially, Maya served as a virtual influencer for Myntra, showcasing new launches in her unique style on platforms like Instagram. Despite efforts to engage with Gen Z, Maya didn't quite gain the traction expected, with her last post dating back to July 2023. However, there's hope for a comeback!

Interestingly, Myntra swiftly pivoted by integrating Maya as a chatbot into their app just before their annual Big Fashion Festival in September 2023, following Ross's famous "Pivot, Pivot, Pivot!" mantra. This transition from virtual influencer to chatbot demonstrates Myntra's agility, leveraging existing capabilities like Fashion GPT for strategic integration. (How to pivot effectively - not a guide by Ross 🙃)

Initially, Myntra projected that over 20% of its user base would engage with Maya, expecting a 16% average increase in categories per order with Fashion GPT. However, concrete numbers on Maya's impact on conversion rates are still awaited.

Spark and Trouble eagerly await updates on this intriguing development!

Hot off the Wires 🔥

We're eavesdropping on the smartest minds in research. 🤫 Don't miss out on what they're cooking up! In this section, we dissect some of the juiciest tech research that holds the key to what's next in tech.⚡

Imagine being able to create a detailed 3D model from a single picture in under half a second! This sci-fi dream is inching closer to reality thanks to TripoSR, a groundbreaking new model from a collaboration between Tripo AI and Stability AI.

Tripo AI is a pioneer in 3D generative AI, using deep learning to create 3D objects from scratch. Stability AI, on the other hand, excels in developing powerful diffusion models, to generate realistic images from text. That’s a killer collab!

So, What’s New?

Traditionally, 3D reconstruction relied on capturing an object from multiple viewpoints (stereo reconstruction with 2D images). While prior attempts at single-image 3D generation using 2D diffusion models addressed data scarcity, they were slow and cumbersome.

TripoSR breaks the mould - it builds upon the success of Large Reconstruction Models (LRMs) to churn out 3D creations in less than 0.5s on an NVIDIA A100 GPU. Talk about speed⚡

Wanna try it out? Head over to the playground!
(reconstructions here may take 5-10 seconds, given that it’s a public demo)

Under the Hood…

TripoSR utilizes a triplane-based neural radiance field or NeRF (anyone recall those Nerf blasters we used to play with 🤩 aah…good old days!). Imagine this as a compressed, powerful way to describe 3D objects, allowing TripoSR to capture intricate details and complex shapes.

It uses a pre-trained DINOv1 image encoder to understand the incoming image, followed by a transformer-based decoder that translates this into the triplane NeRF representation. Finally, a feed-forward NeRF model predicts the colour and density of each point in 3D space, building up the final object.

TripoSR outshines traditional LRM (Large Reconstruction Model) architecture with a meticulously curated subset of the Objaverse dataset, diverse rendering techniques for robustness & channel optimizations.

Also, to get rid of floating artifacts (those pesky disconnected bits that appear in 3D reconstructions), TripoSR cleverly incorporates a "mask loss" into its objective while training.

Why does this matter?

Evaluations on benchmark datasets using metrics like Chamfer Distance (CD) & F-Score (FS) demonstrate that TripoSR not only captures the overall 3D structure accurately but also excels at reconstructing fine details.

This breakthrough has massive implications for various industries. Entertainment, gaming, industrial design, and architecture – all are poised for a revolution.

• TripoSR can drive hyper-personalized gaming experiences - just snap a selfie, transform into a 3D avatar & conquer virtual worlds with your unique look🎮

• Shopping sprees could get a whole lot more immersive! TripoSR could democratize AR-powered online shopping, by enabling stores to offer 3D models for any product on-the-fly 🛋️👗

The future of 3D reconstruction is even wilder when you consider brain-computer interfaces like Neuralink. Imagine a world where you can simply think of an object and models like TripoSR conjure its 3D version in real-time – a revolutionary way to interact with the world.

Spooky as it may sound, the possibilities are truly mind-blowing!

Curious about the nitty-gritty? The paper and code are a click away!

10x Your Workflow with AI 📈

Work smarter, not harder! In this section, you’ll find prompt templates 📜 & bleeding-edge AI tools ⚙️ to free up your time.

Fresh Prompt Alert!🚨

Feeling like your to-do list is a runaway train? Tame the beast with the GTD method in this prompt, and transform your workflow from chaotic to chill!

3 AI Tools You JUST Can't Miss 🤩

• Bash AI - Transform data into actionable insights and shareable strategies with AI in a flash!

• Brancher AI - Build applications powered by AI, no coding required!

• Fix My Code - Optimize your code to be digitally accessible & inclusive with this AI coding companion

Spark 'n' Trouble Shenanigans 😜

Did you see the new iPhone ad? To highlight its insane storage space, Apple used lip-sync techniques (used in deep fake generation) to make photos in the gallery sing “Don’t let me go”, as the user was about to delete them to free up space. How creative! 💡

Check it out! 👇