I work at the intersection of capable AI and secure AI — production-grade agentic/RAG systems, and the red-teaming frameworks that stress-test them. Currently a GenAI Security Testing Developer at Cosmic Info Solution, evaluating LLM systems against prompt injection, jailbreaks, and system-prompt extraction.
- 🧩 fullstack-crm — Production CRM with deal pipeline, REST API, and an AI agent that requires explicit approval before writing data. Live demo →
- 📊 trading-erp — Full-stack ERP unifying sales, purchasing, inventory, and double-entry accounting, with a Claude-powered copilot.
- 🤝 multi-agent-system — Cooperative/competitive/communicative multi-agent RL (PPO), validated across seeds. Live demo →
- 🛡️ AI-Security-Gateway — Middleware that detects prompt injection, jailbreaks, PII, and malicious tool calls in real time. 100% recall / 0% FPR on eval, cuts LLM compromise rate 90.8% → 19.7%. Live demo →
- 🔍 Enterprise-Agentic-RAG-Platform — LangGraph agentic loop, hybrid BM25+FAISS retrieval, Corrective RAG with web fallback. RAGAs Faithfulness 0.91 / Relevancy 0.88, gated in CI/CD.
- 🕵️ Capstone-Agent (CodeSentinel) — Five-agent system that debates, reaches consensus, and patches security vulnerabilities in code, with auto-generated reports.
- 📡 TrustMesh — Agentic SIM-swap fraud detection using telecom CAMARA network APIs.
- 🧠 mind-mri-agent-second-brain — Zero-dependency, offline-first Graph-RAG memory core for autonomous agents; cuts LLM context input 95–99% vs. flat-RAG.
"Detecting Human-Written and AI-Generated Phishing Emails Using DistilBERT and Explainable AI" — IEEE, 2025. Tri-class classifier (~98% accuracy) with SHAP/LIME explainability. Read it →
LangGraph LangChain FAISS BM25 RAGAs Claude API Transformers DistilBERT PPO Stable-Baselines3 PettingZoo
Open to AI Engineer / GenAI Engineer / AI Security roles — UAE, hybrid, or remote.