Superstyle

superstyle

Inspiration

We spent hours shopping and visualizing outfits in our heads while memorizing the best style for us. Websites online don't curate style just for you, they just have one model trying on the clothes and you have to guess if it looks good on you. We often found that clothes we bought online didn't actually fit our style and ended up sitting in our closet, wasting our time and money.

Our goal was to create an integrated platform that helps visualize clothes on your body, assists in discovering your unique style, stores your style in dynamic memory, and facilitates a streamlined shopping experience. We also wanted to create a seamless integration into third-party clothing websites, so customers shopping on other sites can still access the benefits of our platform.

We aim to pioneer and simplify the styling process for anyone.

What it does

SuperStyle revolutionizes styling through a consumer platform and developer platform. Our consumer platform is intended for the daily user looking to discover their style and our developer platform is the access point for the API and integrations.

Consumer Platform:

Robust Clothing Discovery: Allows users to discover inspiration for style through a robust search for clothing items, characteristics, and aesthetics aggregated across the whole internet.
AI-Powered Virtual Try-On: Pioneers accurate AI-powered technology that generates a photorealistic image of the user wearing any outfit of their choice based on a selfie.
Conversational AI Personal Stylist: A voice and text AI stylist that conversationally inquires about the outfit you're looking for while generating shopping recommendations linked to real products.

Developer Platform (API): Inspired by Stripe

Virtual Try-On Endpoint: A REST API that accepts a selfie and clothing description/image, returning a generated try-on image, enabling any e-commerce site to offer virtual try-on.
Outfit Analysis Endpoint: Analyzes any clothing image to identify individual items with detailed metadata (description, color, style tags, category, price range) and returns shopping links for each item.
AI Stylist Endpoint: Generates personalized outfit recommendations based on user preferences, occasion, and body type, complete with try-on images for each suggestion.
Embeddable Widget: A single <script> tag that drops a virtual try-on experience onto any e-commerce site, scanning product images and adding an interactive "Try On" overlay.
Self-Service Developer Portal: A dashboard with API key management, real-time usage analytics, and request history.

How we built it

We used various technologies throughout the development process including Google's Gemini API, Supermemory, OpenAI API, ElevenLabs, Meshy, and more. We started by building a strong backend API foundation and integrated it with a clean and easy-to-use frontend.

API Platform

We built a fully serverless REST API on AWS Lambda (Node.js 20, ARM64) using the Hono web framework, deployed with AWS SAM and exposed through a Lambda Function URL with CORS.
We implemented a production-grade middleware chain: authentication (API key validation against DynamoDB with static key fallback) → idempotency (24-hour response caching for expensive AI operations to prevent duplicate charges) → rate limiting (20 requests/day + 1 concurrent request per user with automatic stale lock cleanup) → usage tracking (non-blocking request logging for the analytics dashboard).
API keys are stored in DynamoDB with a GSI on user ID, supporting self-service generation, one-time reveal, and instant revocation through the developer portal.
Generated images are stored in AWS S3 with a dual lifecycle strategy: ephemeral images auto-delete after 24 hours via S3 lifecycle rules and DynamoDB TTL, while persistent images are stored indefinitely.
We built comprehensive API documentation using Astro Starlight and an embeddable JavaScript widget compiled to a single IIFE bundle with esbuild, requiring only a <script> tag and API key to integrate virtual try-on into any website.

Data Management

Supermemory's user profiles feature dynamically stores a user's style preferences over time and uses that data in our RAG pipelines as well as other data retrieval for future use.
Supermemory's graph memory feature uniquely stores important connections between clothing types, aesthetics, and preferences a user interacts with on our platform.
Our API integration feature leverages Supermemory and AWS DynamoDB as cross-platform data storage to import user style preferences into third-party websites.

Image Generation

We use Google Gemini 2.5 Flash for vision analysis to extract relevant clothing items from uploaded inspiration images, improving accuracy for downstream generation.
We use Gemini's image generation to produce high-quality, realistic virtual try-on images, supporting both reference environment placement and clean studio backgrounds.
We used Meshy.ai to generate personalized 3D avatars from user inputs, exporting optimized .glb files which were dynamically rendered on our website using a Three.js pipeline.
We extracted structured avatar metadata (like proportions and visual attributes) and fed it into our memory layer to enrich the user profile, enabling our AI stylist to make more context-aware, visually aligned recommendations, improving personalization over time.

Voice and LLM Agent Integration

Our conversational voice AI stylist is powered by the ElevenLabs agent using the real-time WebSocket for speech-to-text, processing, and text-to-speech, with OpenAI API powering the LLM layer.
We leveraged detailed prompt engineering as well as parameter tuning to tailor the agent to our specific use case. The agent uses the context of the user's preferences and asks directed questions to extract important information about the type of outfit the user is trying to plan.
The chat dynamically loads the transcript from the voice agent, and the text agent is separately powered by OpenAI API for text-based queries to serve recommendations and power the web search. OpenAI's API is uniquely powerful for text-based processing and recommendations.
We then use Serper to retrieve the recommendations and display specific links for the user to explore.

Privacy

We heavily value user privacy with sensitive AI applications and ensure careful data handling.
We allow users the option to automatically delete their data after 1 day via S3 lifecycle rules and DynamoDB TTL.
All data, especially images, are encrypted and hosted in the cloud for the highest security.
We do not use any user data to train our models, it is purely for user use.

UI/UX

Our frontend is built with Next.js 16 and React 19, styled using Tailwind CSS 4 with an OKLCH color system for perceptually uniform colors across light and dark themes.
We use shadcn/ui (built on Radix primitives) for accessible, consistent UI components and Framer Motion for smooth animations and transitions.
The consumer platform features a sidebar navigation layout for seamless movement between search, try-on, and profile, while the developer portal uses a horizontal top nav with a real-time analytics dashboard built with Recharts.
We prioritized intuitive interactions: one-click try-on from search results, swipe-based outfit discovery, and inline product links from AI recommendations.

Challenges we ran into

We took an extensive amount of time exploring different image generation models and most did not live up to our standards. However, Gemini's image generation through Google's Gemini API was able to produce highly accurate try-on images with realistic body placement and clothing rendering.
Ensuring the generated .glb files from Meshy.ai were properly optimized and compressed so they rendered smoothly in-browser without latency or performance drops required significant iteration on export settings and Three.js rendering configuration.
Building the API widget to be truly plug-and-play was challenging as we needed to handle authentication across origins (via postMessage), scope all CSS to avoid conflicts with host sites, use MutationObserver for lazy-loaded images, and compile everything into a single dependency-free IIFE bundle.

Accomplishments that we're proud of

The API backend is a full-fledged API platform mirroring a production system — complete with authentication, rate limiting, idempotency, usage tracking, and a self-service developer portal.
The accuracy of the image generation using Google Gemini API for users to try on clothes produces photorealistic results.
The conversational agent and text agent work in real time, offering a human-like experience while generating accurate recommendations linked to real products online.
The web search is robust and aggregates data from the whole internet, making our discovery engine powerful.
Our users' highly correlated style preferences are uniquely stored in Supermemory's graph memory instead of standard databases, enabling richer personalization over time.
The embeddable widget allows any e-commerce site to add virtual try-on with a single <script> tag.

Overall, we're proud that we were able to ship a production-grade product hosted on AWS that handles scaling, along with an extensive backend API with real-world integration capabilities.

What we learned

Building a production-grade API taught us various real-world concepts including API frameworks, API key management, rate limiting, idempotency, cloud hosting, spike volume handling, stateful APIs, and most importantly building a reliable API.
We also learned how to seamlessly integrate APIs from companies like Google, ElevenLabs, OpenAI, Supermemory, Serper, and Meshy.
Since we set high goals, a skill we learned was effective project and product management throughout these 36 hours.

What's next for SuperStyle

We see a bright future for SuperStyle through simple third-party integrations and widespread consumer use. Our monetization plans are two-fold.

We anticipate an increase in clothing purchases if customers can virtually try on items. We plan to charge a simple subscription fee for clothing websites to integrate our widget on their platform, enabling a better user experience for their customers.

We also plan to market this to an audience our age who use social media frequently for fashion advice but never receive customized recommendations. We plan to keep the platform completely free for consumers by charging a small commission fee to clothing websites they buy from through our recommendations, similar to an affiliate marketing structure.

SuperStyle's use cases are impactful on a personal and business level, and we hope to revolutionize the styling industry.