NVIDIA NCP-AAI : Agentic AI

A healthcare AI company is deploying diagnostic agents that process medical imaging and patient data. The system must deliver consistent sub-100ms inference times for critical diagnoses while supporting deployment across multiple hospital sites with different NVIDIA GPU configurations (from RTX 6000 workstations to DGX systems). The agents need to maintain high accuracy while being portable across different hardware environments and capable of running efficiently on various GPU memory configurations.
Which optimization strategy would deliver the BEST performance improvements while maintaining deployment flexibility across diverse NVIDIA hardware configurations?

• A. Deploy models using NVIDIA TensorRT optimization in their original FP32 precision format without any quantization or memory optimization, requiring 32GB+ GPU memory across all deployment sites.
• B. Deploy agents using NVIDIA NIM containers with CPU-optimized inference to avoid GPU memory constraints and ensure consistent performance across different hospital infrastructure configurations.
• C. Deploy agents with NVIDIA CUDA-optimized Docker containers using a sequential inference architecture that processes each layer individually with GPU-to-CPU memory transfers between operations to avoid memory issues.
• D. Deploy agents using model optimizations with post-training quantization with Nvidia NIM deployment for portable performance across different GPU platforms and memory configurations.

A company is deploying a multi-agent AI system to handle large-scale customer interactions. They want to ensure the system is highly available, cost-effective, and scalable across multiple NVIDIA GPUs using container orchestration tools.
Which practice is most crucial for successfully deploying and scaling an agentic AI system in production?

• A. Use a static assignment of requests across agents to maintain consistent agent operation and simplify coordination while scaling infrastructure resources as needed.
• B. Deploy agents on a single machine to obtain a dimensioning baseline and thereby reduce setup complexity before expanding system scope.
• C. Implementing automated workload management and resource scheduling frameworks to optimize GPU utilization and maintain service availability.
• D. Optimize GPU utilization frameworks with workload optimization separate from cost analysis, prioritizing resource performance for peak load scenarios in deployment.

An AI engineer at an oil and gas company is designing a multi-agent AI system to support drilling operations.
Different agents are responsible for subsurface modeling, risk analysis, and resource allocation. These agents must share operational context, reason through interdependent planning steps, and justify their collaborative decisions using structured, transparent logic. The architecture must support memory persistence, sequential decision-making and chain-of-thought prompting across agents.
Which implementation best supports this design?

• A. Use stateless LLM endpoints behind an API gateway and pass shared prompts across agents to simulate context and reasoning.
• B. Use LangChain to coordinate third-party agent APIs and store shared information in external memory, with logic encoded in static prompt chains.
• C. Orchestrate NeMo agents via Triton, use vector memory for shared context, ReAct planning, and NeMo Guardrails for reasoning.
• D. Fine-tune separate NeMo models for each agent role using LoRA, with pre-scripted action flows deployed via TensorRT for latency reduction.

When evaluating GPU utilization inefficiencies in deploying Llama Nemotron models across A100 and H100 clusters, which approaches help identify optimal resource allocation strategies? (Choose two.)

• A. Allow Nemotron variants to profile actual workload characteristics and allocate resources based on observed demands.
• B. Allocate all agents to Hl00 GPUs, allowing resource profiles to automatically adjust for model size and computational requirements.
• C. Assess concurrent execution capabilities by employing multi-instance GPU partitioning for varying workload types.
• D. Profile resource utilization for each Nemotron variant and match models to appropriate GPU tiers.

In a global financial firm, an AI Architect is building a multi-agent compliance assistant using an agentic AI framework. The system must manage short-term memory for multi-turn interactions and long-term memory for persistent user and policy context. It should enable contextual recall and adaptation across sessions using NVIDIA's tool stack.
Which architectural approach best supports these requirements?

• A. Leverage NVIDIA Triton Inference Server with dynamic batching to cache session-level inputs between inference calls, and use an external Redis store for long-term memory.
• B. Leverage RAPIDS cuDF for memory tracking by streaming multi-turn conversation logs as GPU- resident data frames, assuming transactional history can be recalled and reasoned over using dataframe operations.
• C. Rely exclusively on TensorRT to encode all prior knowledge into compiled model weights, allowing inference-only execution with no external memory dependencies across sessions.
• D. Leverage NVIDIA NeMo Framework with modular memory management, integrating conversational state tracking, knowledge graphs, and vector store retrieval, while using LoRA-tuned models to adapt responses overtime.