NVIDIA Certified Professional – Agentic AI (NCP-AAI)

Learning plan

About 80 hours of focused study, split into 6 phases. Pick your weekly pace and the timeline adjusts.

Estimated finish: 8 weeks

What the exam covers

Source: official exam guide. Weights can change, so check before you start.

Before you start

• NVIDIA recommends 1-2 years of AI/ML experience and hands-on work on production agentic AI projects
• Strong Python, plus familiarity with Docker and an agent framework such as LangGraph
• A free NVIDIA developer account for build.nvidia.com endpoints. A cloud GPU is optional for self-hosting NIM

1. Weeks 1–2 · 16 h

   Agent architecture, reasoning and memory

   • Compare single-agent, supervisor/worker and peer multi-agent designs, and choose one for a given task and failure tolerance
   • Explain ReAct, plan-and-execute and reflection loops, and when tool calling beats free-form reasoning
   • Design short-term (conversation buffer, summaries) and long-term (vector store, structured state) memory, and say how each fails

   Use: NVIDIA DLI Building Agentic AI Applications With LLMs · Hugging Face Agents Course

2. Weeks 2–4 · 18 h

   Agent development and knowledge integration

   • Build a tool-calling agent with LangGraph using NIM-hosted models, with typed tool schemas and error handling
   • Implement RAG with chunking, embedding and re-ranking NIMs, and handle PDFs, tables and structured data sources
   • Package agents with NVIDIA NeMo Agent Toolkit and know where AI Blueprints and Nemotron models fit

   Hands-on: Build a LangGraph research agent that calls a web-search tool and a RAG tool backed by NIM embedding and re-ranking endpoints. Give it a 3-step plan-and-execute loop and persist state between runs.

   Use: NVIDIA DLI Building RAG Agents With LLMs · NVIDIA NeMo Agent Toolkit on GitHub · LangGraph on GitHub · NVIDIA API catalog (build.nvidia.com)

3. Weeks 4–5 · 12 h

   Evaluation and tuning

   • Measure task success, tool-call accuracy, retrieval recall@k and faithfulness, and compare agent versions on a fixed test set
   • Tune prompts, retrieval parameters and model choice from eval results rather than intuition
   • Use LLM-as-judge carefully, calibrating it against a human-labelled sample

   Hands-on: Write 30 test questions with reference answers for your research agent. Score retrieval recall@5 and answer faithfulness, change one variable (chunk size or re-ranker on/off) and report the delta.

   Use: NVIDIA DLI Evaluating RAG and Semantic Search Systems

4. Weeks 5–7 · 18 h

   Deployment, scaling and operations on NVIDIA

   • Self-host a NIM container, size GPU memory for a given model, and scale replicas behind a load balancer or on Kubernetes
   • Know NVIDIA AI Enterprise, NGC, Triton and Dynamo, and when to use hosted endpoints vs self-hosting
   • Instrument agents with traces, latency and token metrics, and plan model and prompt version rollouts

   Hands-on: On a cloud GPU VM, pull a small LLM NIM from NGC with Docker, point your agent at the local endpoint instead of build.nvidia.com, and load-test it at 1, 4 and 16 concurrent users while recording p95 latency.

   Use: NVIDIA DLI Deploying RAG Pipelines for Production at Scale · NVIDIA NIM docs · NVIDIA nim-deploy (Helm and Kubernetes examples)

5. Week 7 · 6 h

   Safety, compliance and human oversight

   • Add input/output rails and topic restrictions with NeMo Guardrails, and defend tool-using agents against prompt injection
   • Design human-in-the-loop approval for high-impact actions, plus audit logging for compliance

   Use: NeMo Guardrails docs

6. Week 8 · 10 h

   Practice and review

   • Re-quiz the four biggest domains (architecture, development, evaluation, deployment), which make up 56% of the blueprint
   • Practise a 65-question timed set in under 120 minutes

   Use: NCP-AAI exam page and blueprint

You're ready to book when…

• You have built, evaluated and self-hosted a multi-tool agent on NVIDIA NIM
• You can pick an agent architecture and memory design for a scenario and defend it
• You can explain how you would measure and improve a faltering agent using metrics, not guesswork
• You finish timed practice sets of 60-70 questions within 120 minutes at 80% or higher

Exam tips

• The exam is 60-70 questions in 120 minutes, remotely proctored, and needs a Certiverse account.
• NVIDIA's exam page showed "Coming soon" next to the registration link, so confirm you can book before setting a date.
• The published weights add up to 98%, not 100%, so treat them as approximate.