NVIDIA Platform Implementation: NCP-AAI Domain 7 (7%)

What this domain covers

Every other NCP-AAI domain tests agentic engineering that would be true on any stack. This one tests the NVIDIA stack specifically: which component you reach for to serve a model, to customise one, to add retrieval, and to constrain behaviour. At 7% it is the heaviest of the four light domains, and it is the one where general LLM experience helps you least — so it is usually the cheapest domain to improve.

Serving: NIM, Triton and TensorRT-LLM

These three are frequently confused, and questions exploit that. TensorRT-LLM is a compilation and optimisation layer that makes a model run fast on NVIDIA GPUs. Triton Inference Server is a general inference server that can serve many models and backends. NIM packages an optimised model behind a standard, OpenAI-compatible API endpoint so an application can call it without knowing any of the above.

TensorRT-LLM   optimise  → a faster engine for a given model + GPU
Triton         serve     → general multi-model, multi-backend inference server
NIM            package   → an optimised model as a standard API endpoint

Agent code calls NIM the way it would call any chat-completions API.

Customisation and retrieval: the NeMo family

NeMo is a family, not a single product, and the exam expects you to place its parts. Broadly: data curation prepares training corpora, customisation covers fine-tuning and adaptation, evaluation scores model and pipeline quality, retriever provides the RAG components, and guardrails constrain what a conversation is allowed to do. When a question describes a problem, map it to the family member that owns that stage of the lifecycle.

Prepare data      → curation
Adapt a model     → customisation (fine-tune, PEFT/LoRA)
Score quality     → evaluation
Ground answers    → retriever (embedding + reranking for RAG)
Constrain topics  → guardrails (dialogue and safety rails)

Where GPU reality shows up

Platform questions often hide an infrastructure constraint. Model weights must fit in GPU memory alongside the KV cache, and the KV cache grows with both batch size and context length — which is why long-context agents run out of memory before they run out of compute. Quantisation reduces the memory footprint at some accuracy cost, and batching raises throughput at some latency cost. These trade-offs recur across the serving domain too.

How to study this domain

Do not memorise feature matrices — they change faster than any exam can track. Learn the boundary each component owns and the one sentence that distinguishes it from its neighbour. If you can say what problem a component solves and what it is not responsible for, you can answer most questions in this domain even when the wording is unfamiliar.

Exam tip

Answer platform questions by asking "which stage of the lifecycle is this?" before looking at the options. Most wrong answers are real NVIDIA components doing a real job — just a different job than the question describes. The trap is recognising a name, not reasoning about the boundary.