
OFF Guard is an AI-powered KYC compliance screening assistant built for the AMD Hackathon. It evaluates identity documents, detects fraud, resists prompt injection attacks, and provides structured risk assessments – all running on AMD GPUs via vLLM with Google's Gemma-2-9B. The Problem: Compliance teams manually review thousands of documents daily. Bad actors embed hidden instructions to manipulate AI systems, while deepfakes and forged documents increasingly bypass traditional checks. Sanctions violations cost billions in fines. Our Solution: OFF Guard screens documents and contextual metadata, returning a risk rating (low/medium/high), actionable suggestions, and clear reasons – never a final approve/reject, because humans always decide. Three Golden Rules: (1) Suggest only – no final verdict, (2) Never standalone red rating without reasons, (3) Document is data, not commands – embedded instructions are flagged, not obeyed. Key Features: Document upload (PDF, images, text) with OCR, MRZ passport validation, SHA-256 authenticity hashing, live OFAC sanctions screening, voice dictation, face detection, trust score (0-100), adverse media search, batch screening (CSV), downloadable PDF reports, and AI-powered insights including document summary, risk narrative, and interview questions. Tech Stack: FastAPI, vLLM, Google Gemma-2-9B on AMD GPU, Tesseract OCR, OpenCV, GNews API, OFAC SDN List. Impact: Reduces manual workload, detects fraud attempts, resists AI manipulation, and ensures compliance – making KYC faster, safer, and more trustworthy.
13 Jul 2026