Inspiration
We were frustrated by the gap between online shopping and the fitting room experience. Every time we bought clothes online, it was a gamble — would it actually look good on us? Returns are expensive, wasteful, and time-consuming. We wanted to bring the fitting room to the browser, powered by AI, and make it feel as natural as asking a friend "how does this look on me?" That's when we imagined Stella — a voice-powered AI stylist that doesn't just show you clothes, but actually dresses you in them, right on your screen.
What it does
SuperNova TryOnMe is a Chrome extension that transforms online shopping into a virtual fitting room experience across Amazon, Shein, Temu, and Google Shopping. Users upload 5 photos of themselves, and our AI generates professional model poses. From there, they can:
- Try on any garment directly from product pages with one click — our 5-step AI pipeline handles garment extraction, outfit conflict resolution, and photorealistic try-on
- Talk to Stella, our real-time voice AI stylist powered by Amazon Nova Sonic — she searches products, builds outfits, triggers try-ons, and gives personalized style advice, all through natural conversation
- Build complete outfits with 6 categories (top, bottom, shoes, necklace, earrings, bracelets) using the Outfit Builder wardrobe — Stella orchestrates Nova Act to search Amazon and display numbered items the user selects by voice
- Animate try-on results into 6-second fashion videos using Grok Imagine Video
- Add items to cart through AI-powered browser automation via Nova Act
- Save favorites to the cloud with retailer tracking, outfit grouping, and price totals
How we built it
The backend runs on Node.js/Express deployed on EC2 with PM2:
- Amazon Nova 2 Sonic — Bidirectional HTTP/2 streaming for Stella's real-time voice with tool-calling mid-conversation
- Amazon Nova Canvas — Background removal.
- Amazon Nova 2 Lite — Fast product classification and outfit detection
- Amazon Nova Act — Browser automation agent for smart search and add-to-cart
- Grok Imagine Video — Try-on result animation
AWS services include Bedrock, S3, DynamoDB, Cognito, and ElastiCache (Redis). The Chrome extension (Manifest V3) injects try-on buttons into 4 retailers, manages auth tokens, routes voice commands, and renders the outfit builder wardrobe and smart search UIs. The full codebase is 14,500+ lines across the extension and backend.
Challenges we ran into
The hardest challenge was making Stella reliable in the outfit builder flow. Amazon Nova Sonic eagerly calls tools the moment it has enough context.
Identity preservation in try-on was another battle.
Accomplishments that we're proud of
- Stella actually works — a real-time voice stylist that searches, builds outfits, tries on clothes, animates results, and adds to cart, all through natural conversation in multiple languages
- Multi-retailer support — the same try-on experience works across Amazon, Shein, Temu, and Google Shopping
- The outfit builder — a visual wardrobe where Stella searches 6 categories simultaneously via Nova Act, displays numbered items, and the user selects by voice, with the try-on triggering automatically when all items are chosen
- Production-grade architecture — Redis caching, S3 storage, Cognito auth, DynamoDB persistence, with 14,500+ lines of shipped code
What we learned
- Nova Sonic is incredibly powerful for voice AI, but behavioral control MUST be enforced at the code level — prompt instructions are unreliable for controlling tool-calling timing
- Multi-model orchestration is the future — no single model does everything well, but combining specialized models (Nova for voice + classification, Gemini for image generation, Grok for video) creates a product that feels magical
- Identity preservation in AI try-on requires deliberate engineering — image ordering, face anchoring, and output validation all matter significantly
- Browser automation with Nova Act unlocks capabilities that traditional APIs can't touch — like searching Amazon with natural language and adding items to a real shopping cart
What's next for SuperNova TryOnMe
- AR mirror mode — real-time try-on using the webcam feed, so users can see themselves in clothes as they move
- Social sharing — share try-on results and animated videos directly to Instagram, TikTok, and WhatsApp
- More retailers — expanding to Zara, H&M, ASOS, and Nordstrom
Built With
- amazon-ec2
- amazon-nova-act
- amazon-nova-canva
- amazon-nova-lite
- amazon-nova-sonic
- s3
Log in or sign up for Devpost to join the conversation.