SafeO

SafeO — Multilingual AI Security Engine (AMD Unicorn Track) SafeO is a real-time cybersecurity decision engine that protects enterprise apps at the moment malicious input tries to enter — before it's saved, executed, or forwarded. The problem: ERP fields, CRM notes, and forms are full of attack surface — SQL injection, prompt injection, ERP fraud, and critically, Arabic/Arabizi obfuscation that English-first tools miss entirely. Perimeter tools (WAF, SIEM) only react after the damage is done. What SafeO does: Every scan runs a 3-tier ML pipeline, returning ALLOW, WARN, or BLOCK in milliseconds, before data persistence: Tier 1: deterministic heuristics (SQLi, XSS, prompt injection, homograph, entropy) Tier 2: DistilBERT/TF-IDF classifier, ROCm-ready with feedback retraining Tier 3: optional LLM via Fireworks AI on AMD hardware For WARN/BLOCK, a 5-agent LangGraph investigation runs: Multilingual (normalizes Arabic/Arabizi/mixed-script), Policy (compliance rules), Forensics (MITRE ATT&CK mapping, powered by Gemma 3 4B via Fireworks on AMD), Verifier (cuts false positives), Remediation (playbook + Jira/WhatsApp output). Every step is SHA-256 hashed into a tamper-evident audit chain. Key features: Role-adaptive SafeO Assistant (same engine, explanation depth adapts to user); Visual Evidence Capture (Playwright screenshots + annotated threats); drag-and-drop Workflow Builder; one-click Jira escalation from logs; and our core differentiator — Arabic/Arabizi security normalization catching mixed-script evasion English-only tools miss. AMD infrastructure: ForensicsAgent runs Gemma 3 4B via Fireworks AI, which serves entirely on AMD Instinct GPUs — every forensics reasoning step is AMD-powered. Codebase is ROCm-ready (bf16 LoRA, PyTorch ROCm, AraBERT paths); amd_setup/ scripts are built and ready to activate on AMD Developer Cloud once credits provision. Self-improving: Bayesian threshold engine adapts block sensitivity from analyst feedback within safety bounds