SentinelRAG

THE PROBLEM: Enterprise RAG pipelines fail silently. When an AI model hallucinates a critical data point or falls victim to a prompt injection, engineering and security teams lack real-time visibility and control. THE SOLUTION: SentinelRAG acts as a deterministic trust layer for enterprise AI. We built a multi-model evaluation engine behind a hardware-level proxy to ensure secure, auditable, and grounded AI behavior. CORE ARCHITECTURE & FEATURES: - Edge-Level Security: All LLM traffic routes through Veea Lobster Trap. This firewall executes sub-millisecond Deep Prompt Inspection (DPI) to block injections and obfuscation before they reach the LLM. - Dual-Stage Evaluation: We use a two-step hallucination detection pipeline. Heuristics handle instant hard-data mismatch checks, followed by Gemini 2.5 Flash-Lite acting as a strict semantic judge to evaluate Faithfulness (60%) and Relevance (40%). - Reasoning Engine: Core RAG generation is powered by Gemini 2.5 Flash retrieving context from ChromaDB. - Explainable Audit Trail: Every query, reasoning metric, and blocked attack is logged relationally to SQLite for full XAI compliance. SentinelRAG bridges the gap between raw LLM capabilities and secure, production-ready enterprise infrastructure.