NCP-AAI — NVIDIA-Certified Professional

Prep for NCP-AAI

390 scenario questions on building and running agentic AI systems, with an explanation for every one of the four options — 1,560 in total. The professional level, so the questions are longer and the wrong answers are better.

Coming soon: App Store Coming soon: Google Play

• Exam NCP-AAI, Agentic AI (NVIDIA-Certified Professional)
• Questions 390, split into 6 practice sets of 65
• Explanations 1,560 — one per answer option
• Price USD 5.99, one-time. No in-app purchases
• Network None. The app never connects
• Platforms iOS and Android

The exam this app is for

NCP-AAI is the professional-level certification for agentic AI: systems where a model plans, calls tools, keeps state across many steps and acts on the world. It assumes you have already built one. The questions are not definitions — they are situations from a running platform, where two or three of the four options are things a competent team might actually do, and only one holds up under the constraint the scenario gives you.

A tool registry grows from twelve tools to two hundred and the model starts choosing plausibly wrong ones. A booking workflow reserves a flight and a hotel, then the car step fails permanently. A run drifts over sixty steps until its own summaries have replaced the user's original request. A vendor's web page contains hidden text telling the agent to email the user's account details somewhere. Reading the constraint carefully is half the exam, and this bank is written to train that.

What the 390 questions cover

Ten domains, with this app's own distribution below. The shares are an editorial choice for practice, not an official blueprint.

• Agent Architecture and Design60 questions · 15%

  Single agent against a crew, explicit state machines against free-form loops, multimodal inputs into an agent workflow, self-critique on the same endpoint and why it flatters itself, and how far tenants can share one deployment.

• Agent Development60 questions · 15%

  Tool design and selection at scale, structured tool calls instead of parsed prose, arithmetic that belongs in code rather than in the reasoning text, concurrency limits on a shared endpoint, retries with backoff, and compensating actions when step three of a booking fails for good.

• Evaluation and Tuning54 questions · 14%

  Why a pass rate stalls when every failure is logged as "task failed", what a slow six-week decline points to when nothing was deployed, the latency and variance a single success rate hides, and reading a temperature change that improves the mean and wrecks consistency.

• Deployment and Scaling48 questions · 12%

  Token cost per run worked out from call counts and prices, fitting a multi-step workflow inside a latency budget, per-tenant rate and spend caps, canary rollouts of a new planning strategy, and where autoscaling helps and where it does not.

• Cognition, Planning and Memory42 questions · 11%

  Goal drift over long runs, constraints from the first message getting lost, context that overflows partway through a task, and separating what an assistant should remember forever from what belongs to this session only.

• Knowledge Integration and Data Handling42 questions · 11%

  Cross-language retrieval that returns nothing useful, tables flattened into prose before embedding, deciding when retrieval beats putting a small stable glossary in the prompt, and routing cheap chatter away from an expensive pipeline.

• NVIDIA Platform Implementation30 questions · 8%

  Serving choices on NIM, routing between a small fast model and a large reasoning one, tuning for time-to-first-token in an interactive product, retrieval with NeMo Retriever, and where a guardrail layer sits relative to input, output and tool use.

• Run, Monitor and Maintain18 questions · 5%

  Tracing one failed run across retrievals, tool calls and generations; the telemetry that makes a drop from 92% to 78% explainable; and recording the exact model version and sampling parameters that served each run.

• Safety, Ethics and Compliance18 questions · 5%

  Indirect prompt injection from content the agent reads, least-privilege access to a customer database, sensitive data ending up in long-term memory, and what leaves the building when transcripts go to an externally hosted model.

• Human-AI Interaction and Oversight18 questions · 5%

  Approval gates on irreversible actions only, thresholds that send large refunds to a human, clean handoff to a specialist mid-conversation, and rubber-stamping in a review queue — oversight that exists on paper but not in practice.

How you study with it

• Six sets of 65, longer than the associate apps because the questions are longer. A full set is a genuine endurance run, which is the part people underestimate at professional level.
• Exam mode gives you a set against a 120-minute clock with no feedback until the end. Practice mode explains each answer as you go, which is how to work through a scenario domain the first time.
• Per-domain scoring across all ten domains after every set. With six domains carrying most of the weight, the breakdown usually shows one that is quietly costing you ten points.
• Missed questions and bookmarks collect into their own review runs. On this bank that matters: the questions you get wrong are usually wrong for a reason you can name once and then stop repeating.
• All four options explained, and this is the whole value at professional level. The distractors are good — they are what a reasonable engineer would try — so the explanation of why each one fails under the stated constraint is the lesson.

Questions people ask

How much harder is this than the associate exams?

Different, more than harder. The associate exams ask whether you know a thing; this one gives you a situation with a constraint attached and asks what you would do. Expect several lines of setup per question and options that are all defensible until you re-read the constraint. That is why the sets here are 65 questions and 120 minutes.

Should I have built an agent before?

It helps a great deal. If you have never run a multi-step agent with tools, start with the practice mode and read the explanations as material rather than as feedback — that is a slower but workable path. The exam itself is aimed at people with production experience.

Is there arithmetic in it?

Some, and it is the kind you do on paper: token cost per run from call counts and prices, a latency budget added up across retrieval and generation steps, throughput against a concurrency limit. The explanations show the working.

Is it tied to one agent framework?

No. The questions are about patterns — planning, tool calling, memory, guardrails, observability — not about one library's API. NVIDIA's own platform components appear in their own domain, since the exam expects you to know what each one is for.

Do I need the associate apps first?

Not at all; they are separate exams and separate banks. If you are coming from the language-model side, Prep for NCA-GENL covers the retrieval and evaluation groundwork this exam builds on. The four apps share no questions.

Where do the questions come from?

They are written with AI and organised against the published NCP-AAI domains. They are original practice material, not real exam questions, and no exam question is reproduced. If a scenario or an explanation looks wrong, write to us — corrections ship to everyone in the next update.

← All four apps

Disclaimer. Prep for NCP-AAI is an independent study aid. It is not affiliated with, authorized by, endorsed by, or sponsored by NVIDIA Corporation. NVIDIA, NCP-AAI and the NVIDIA-Certified Professional program name are trademarks or certification program names of NVIDIA Corporation, used here only to identify the exam this app helps you prepare for. No certification result is guaranteed.