CatalogShield — AI Product Truth Engine

Retailers lose 30%+ of GMV to preventable product returns driven by catalog defects: contradictory sizing charts, exaggerated waterproof claims, and mismatched photography. CatalogShield is an AI quality gate and product truth engine built on AMD Instinct™ MI250 GPUs, ROCm™ 6.2, and Gemma-3-27B Vision. It intercepts catalog defects before publication by cross-examining marketing copy against product photography, signed factory spec sheets, and 90-day return cohorts. Key Capabilities: 1. Merchant Truth Gate: Interactive visual studio with a 2.6× optical lens, ViT cross-attention heatmaps, and tri-state claim verdicts (SUPPORTED, CONTRADICTED, UNVERIFIED). Hovering any claim illuminates its bounding-box reticle on the photo. 2. Approval-Gated Patches: 1-click listing copy patches dynamically recalibrate product risk scores (e.g. CS-1001 drops from 78 to 13) and protect bottom-line return dollars. 3. B2B Chargeback Dossiers: Generates automated supplier credit dispute packages ($7,050.73 on CS-1001) backed by visual crop evidence and purchase order citations. 4. Consumer Storefront: Includes an interactive TrueFit™ size calibrator calculating millimeter clearance against factory last tooling (-4.2mm narrow toe-box) and a Truth Verification seal. 5. Storewide Risk Radar: Batch-streams audits across the entire 8-SKU catalog ($184,200 GMV, $42,850 return risk intercepted) with live AMD telemetry. AMD Instinct™ MI250 ROI Advantage: - 15× Lower Inference Cost: $0.0032 / SKU audit vs. $0.0480 A100 baseline. - 2.6× Faster Latency: 1.82s multimodal audit vs. 4.80s. - 2.6× Higher Throughput: 68.4+ tok/s sustained via ROCm 6.2 & vLLM. Live App: https://catalogshield.vercel.app/ GitHub: https://github.com/ImTaieb/catalogshield