TWINFIT-AI

fashion e-commerce loses billions to returns — 25-35% of orders come back, mostly due to fit. Shoppers can't tell if an M at Zara equals an M at Myntra, and they can't see garments on their own body. TwinFit fixes both as a B2B SaaS that retailers integrate with one line of code. The shopper flow: a "Try It On" button appears under any product image via our embeddable widget. It opens TwinFit in a modal with the garment pre-loaded. The shopper enters measurements and instantly gets a size recommendation scored against brand-specific size charts (H&M, Zara, Myntra + generic), with confidence, fit notes tuned for Indian ethnic wear, and a return-risk flag. Meanwhile a vision LLM (via Fireworks AI) analyzes the product image automatically — category, sleeve, fit, fabric, color — no catalog work for the retailer. Finally, the shopper uploads a photo and IDM-VTON, an SDXL-based diffusion try-on model, renders them wearing the exact garment, preserving face, pose, and body. AMD is at the core of our inference story: we deployed and ran IDM-VTON on an AMD GPU via ROCm during the hackathon — PyTorch's ROCm backend runs the CUDA codebase unmodified — and our repo ships the full AMD deployment kit (setup scripts, notebook, vLLM serving script for Gemma 3). Our backend uses a resilient multi-engine cascade (own AMD GPU → hosted IDM-VTON → FLUX fallback) so a shopper never sees a broken screen, which this hackathon's shared GPU pool stress-tested in real time. Built with FastAPI, Next.js, and a demo retailer storefront showing the one-line integration. Roadmap: Shopify app for one-click distribution to every clothing store, then a fit-data flywheel training India-specific size models. India fashion e-comm is $21.6B growing 24% CAGR — a 20% return reduction pays for TwinFit many times over.