Bridging the Safety Gap: A Guardrail Pipeline for Trustworthy LLM Inferences