Free NVIDIA-Certified Associate: AI Models and Application Development (NCA-AMAD) Practice Questions

Sample NCA-AMAD Practice Questions

Browse all 10 free NVIDIA-Certified Associate: AI Models and Application Development practice questions below.

1. A support agent's evaluation reports that 91% of final answers match the reference answers. Yet operations finds that many customer records the agent said it had updated were never changed. Trajectory logs show the agent often wrote a confirmation without calling the update API. Which metric should the team make primary?

   • Task success rate per scenario, counting a task only when the record really changes as required
   • Mean end-to-end latency per request, broken down by which tools the agent called in each session
   • Answer correctness from an LLM judge that compares each final reply with the reference reply
   • ROUGE overlap between each final reply and its reference, tracked separately for each scenario

2. A long-document assistant serves a 70B model. Profiling shows that prompt processing and token generation load the GPUs in different ways. The team wants prompt processing to run with low tensor parallelism to cut communication, and token generation with a higher degree of tensor parallelism. Which deployment allows this?

   • Pipeline parallelism across two nodes, with the early layers on one node and the later layers on the other
   • More identical replicas, each running both phases with one shared parallelism setting for the model
   • Separate prefill and decode workers on different GPUs, each set up on its own, handing off the KV cache
   • A small draft model that proposes output tokens while the large model is still reading the incoming prompt

3. A retailer's assistant must answer questions about prices and stock for 200,000 products, which change several times a day. The team is debating whether to fine-tune a model on the catalog every night. Which approach fits best?

   • Fine-tune the model each night, so that current prices are stored in its weights
   • Look up current catalog data at query time and pass it to the model as context
   • Add few-shot examples with sample prices, so that the model learns their format
   • Write all current prices into the system prompt, and refresh the prompt each hour

4. A RAG assistant often answers "I can't find that" for questions whose answers exist in the knowledge base. When the team inspects failures, the right passage is usually absent from the five chunks passed to the model, while the chunks that are retrieved are mostly on topic. Which retrieval metric best tracks this failure?

   • BLEU, the n-gram overlap between each answer and a reference answer written by experts
   • Precision@K, the share of the K retrieved chunks that are relevant to each test question
   • Faithfulness, the factual consistency of each answer with the context it was given
   • Recall@K, the share of each question's relevant chunks found among the top K results

5. A platform team must stand up matching GPU-enabled Kubernetes clusters on Amazon EKS, Azure AKS and Google GKE, each with NVIDIA's GPU and NIM operators, and rebuild them reliably after changes. Clicking through each cloud console has produced clusters that drift apart. What fits best?

   • NVIDIA's Terraform modules, which define the clusters and NVIDIA operators as code
   • A runbook of console steps, with a different engineer following it for each cloud
   • Manual installs of GPU drivers on each node after the clusters have been created
   • One Kubernetes cluster stretched across all three clouds, so there is one copy to keep

6. An insurer automates claims intake. Regulators require the same documented sequence for every claim: verify the policy, extract damages, check fraud rules, then route for payment. Claim documents arrive as photos, scanned forms and free-text emails in many layouts. A rules-only RPA pipeline breaks on the layout variety. Which design fits best?

   • A swarm of peer agents that negotiate each claim among themselves until they reach consensus
   • An RPA pipeline with more layout templates and parsing rules added for each new document type
   • A fixed workflow of the required steps, with LLMs handling the unstructured inputs in each step
   • A fully autonomous agent that decides for each claim which steps to run and in what order to run them

7. A benchmark logs a request with an end-to-end latency of 3.0 seconds, a time to first token of 1.0 second and 401 output tokens. A teammate computes inter-token latency by dividing the end-to-end latency by the number of output tokens. Which value matches the definition used by NVIDIA's benchmarking tools?

   • 10 ms, from the full request time plus the first-token time over 400 tokens
   • 5.0 ms, from the time after the first token spread over the 400 that follow it
   • 2.5 ms, from the time to first token spread over the 400 later tokens
   • 7.5 ms, from the full request time spread over all 401 output tokens

8. A SaaS vendor fine-tuned a separate LoRA adapter for each of 300 business customers, all on the same 8B base model. Most customers send only a few requests per hour, and the vendor wants to avoid running 300 separate model deployments. How should the adapters be served?

   • Retrain a single shared adapter on all customers' data, so that one adapter serves them all
   • Merge all 300 adapters into a single base model, so that one model covers each customer
   • Run one base deployment that loads each customer's adapter on demand in mixed batches
   • Merge each adapter into its own copy of the base and run 300 small deployments in parallel

9. An HR assistant keeps quoting last year's parental leave policy. The new policy PDF was added to the source folder in March, but the vector index was built once in January, and both policy versions now sit in the folder. What should the team change?

   • Raise the number of retrieved chunks, so that both versions reach the model each time
   • Re-ingest changed documents on a schedule and remove superseded versions from the index
   • Fine-tune the LLM on the new policy, so that the model knows the current rule directly
   • Add a system prompt rule that tells the model to prefer the most recent policy it sees

10. An electronics company wants to boost its chip designers' productivity with AI that helps write and debug design scripts and answers questions from internal design documentation. Which approach reflects how NVIDIA and its partners describe such deployments?

    • A general chatbot trained on public data, since design knowledge is widely published
    • A rules-based script generator, since AI models cannot work with design languages
    • An autonomous agent that replaces the designers and signs off on designs itself
    • Domain agents built on open Nemotron models and NeMo tools, tuned on design data