VotalAI announced production deployments of its advanced AI Runtime Guardrail Stack, leveraging NVIDIA NeMo Data Designer, the NVIDIA NeMo Guardrails toolkit, and the NVIDIA Nemotron 3.5 Content Safety model stack on NVIDIA AI infrastructure. The platform, featuring the vai35-4B-v2 model, secures agentic AI workloads with real-time policy enforcement from prompt to tool call, protecting against jailbreaks, prompt injections, encoding attacks, and unauthorized tool use. Key highlights include a policy-driven runtime with clean separation between the Policy Decision Point (PDP) and Policy Enforcement Point (PEP), enabling per-tenant, per-workload custom policies without redeploying the runtime. The stack delivers 151+ detection techniques across 19 adversarial attack categories, with end-to-end runtime latency of approximately 190 ms at 16k to 32k token prompts and roughly 10% overhead relative to the underlying LLM response time at contexts up to 256k tokens. VotalAI uses NVIDIA NeMo Data Designer to synthesize diverse, region-aware, tenant-isolated adversarial datasets, feeding a Supervised Fine-Tuning (SFT) and preference-tuning pipeline that produces the vai35-4B-v2 guardrail model.