NVIDIA AI Model and Application Development Certification Exam

The test taker will be assessed on their understanding of agentic AI concepts, including reasoning patterns, safety concepts and failure modes, multi-agent components, agent orchestration, human oversight mechanisms in agentic AI, measuring and improving agent performance, and integrating agents with tools and external systems.

• Agentic AI concepts: how agents differ from generative AI and rule-based systems; core building blocks; harnesses and runtimes.
• Single LLM calls versus multi-step agentic workflows.
• Integration with tools and external systems, including APIs, tool use, and Model Context Protocol (MCP).
• Performance improvement through feedback, model selection, and prompt and workflow refinement.
• Safety controls, including guardrails, output filters, prompt-injection defenses, and hallucination reduction.
• Performance measures such as task-completion rate, accuracy, token usage, latency, cost, and the resource impact of multi-step workflows.
• Orchestration and execution order, including tool and API integration.
• NVIDIA tools for building agents, including NVIDIA NeMo™, NVIDIA Agent Toolkit, and NVIDIA NemoClaw™.
• Human oversight mechanisms, including human-in-the-loop workflows.
• State and context management: stateless and stateful interactions and context engineering.
• Multi-agent components, collaboration patterns, and when to choose a single-agent versus multi-agent design.
• Safety concepts and failure modes, including tracing and NeMo Guardrails.
• Reasoning patterns such as ReAct, reflection, and planner-executor.
• Short-term and long-term agent memory, and when each is appropriate.

17%
Exam Weight

Collecting and converting raw data into meaningful, structured information by removing noise, personally identifiable information (PII), and toxic or harmful data to curate high-quality data and create a self-improving loop, or data flywheel, to continuously refine AI models and generate better outcomes.

• Feature engineering, feature stores, and exploratory data analysis.
• Data governance standards, lineage and provenance for auditability, and model risk management.
• Data ingestion and ETL components, including distributed processing with Spark, Databricks, Kafka, and NVIDIA RAPIDS.
• Data preprocessing and cleaning, including data labeling and text preprocessing.
• Privacy requirements: PII handling, data retention, data sovereignty, bias, and misuse risks.
• Structured, semi-structured, and unstructured data.
• AI evaluation metrics and datasets, including why application-specific evaluations can be more useful than general benchmarks.
• Data-quality issues such as missing values, duplication, and staleness.
• Data flywheels: feedback collection, using production data for improvement, and associated challenges.
• Training, inference, and retrieval data flows; batch versus streaming processing.
• How data readiness, quality, and reliability affect model performance.
• Production quality monitoring, including accuracy, groundedness, hallucination rate, and task completion.

17%
Exam Weight

Foundational knowledge of AI Inference, including inference-serving infrastructure, inference pipeline phases and models, optimization techniques, improving throughput, latency and performance, and understanding inference costs.

• Inference-serving infrastructure: model servers, gateways and routing, provider abstraction, Dynamo, autoscaling, request queuing, cold starts, and scaling strategies.
• Inference pipeline phases and modes, including prefill and decode and real-time versus batch inference.
• Optimization techniques: quantization and precision tradeoffs, KV and prefix caching, speculative decoding, tensor parallelism, disaggregated serving, and context-length effects on GPU memory.
• Serving best practices for batch, real-time, and streaming services.
• Improving throughput and latency through batching, token streaming, TensorRT-LLM, quantization settings, and caching.
• Latency versus throughput tradeoffs.
• Performance measures: latency, throughput, time to first token, and load testing.
• Inference cost drivers, including model size, tokens, traffic, hardware, and latency targets.

15%
Exam Weight

Key concepts for adapting a general pre-trained AI or large language model (LLM) for specific tasks or use cases, including fine-tuning types and approaches, data requirements, risks, pre-deployment testing and validation, and deploying fine-tuned models.

• Deployment of fine-tuned models through adapter serving or model merging.
• A/B testing and offline experiments for validating AI model impact before deployment.
• Fine-tuning methods: SFT, PEFT, LoRA, QLoRA, RLHF, preference optimization (DPO, KTO, GRPO), distillation, and instruction tuning.
• Fine-tuning risks, including catastrophic forgetting and overfitting.
• Selecting a fine-tuning approach and preparing labeled training data.
• What fine-tuning is, when to use it, and its data requirements.
• RAG and retrieval best practices: document ingestion, chunking, embeddings, vector indexing, query and retrieval workflows, hybrid search, and reranking.
• RAG pipeline components for text, code, and multimodal data; semantic versus hybrid search.
• Tradeoffs among prompt engineering, RAG, and fine-tuning, including data needs and cost/effort.
• Prompt engineering and optimization: system prompts, zero-shot and few-shot prompting, chain-of-thought prompting, structured output, and prompt optimization.

14%
Exam Weight

Knowledge of varied models and types across computer vision, non-language and language-models, understanding model selection criteria and constraints, and evaluating model efficacy and interpretability. 

• Neural networks versus classical machine-learning models.
• Computer-vision tasks and model types: image classification, object detection, segmentation, and OCR.
• Why parallel computing is suited to neural-network training and inference.
• Non-language model types: embedding, reranking, multimodal, omni, vision-language, and world foundation models; generative versus discriminative models.
• NVIDIA model families and their use cases, including Nemotron.
• Evaluating model efficacy, robustness, and interpretability.
• Language-model types: LLMs, small language models, reasoning models, base and instruction-tuned models, open and proprietary models, dense and mixture-of-experts models, and foundation versus task-specific models.
• Model-selection criteria, including workload fit and size, cost, and performance constraints.
• Core LLM mechanics: tokens, context windows, autoregressive generation, and latent/embedding space.

11%
Exam Weight

How to build enterprise AI workflows using a supported suite of NVIDIA libraries, microservices, and frameworks, moving from prototype to production, streamlining AI application development, and optimizing hardware utilization.

• AI service deployment and orchestration on infrastructure, including containers, Kubernetes, model and microservice deployment, infrastructure as code, and service meshes.
• NVIDIA AI Enterprise: its purpose, enterprise problems addressed, and deployment environments.
• NVIDIA Blueprints as reference architectures; matching blueprints to business scenarios and building on NVIDIA NIM™.
• Consistent AI deployment across on-premises, cloud, edge, and hybrid environments, including self-hosting for data control.
• Productivity use cases for LLM-powered applications.
• NVIDIA NIM microservices: model packaging, accelerated inference, compute-aware selection, and differences from traditional model serving.
• NVIDIA AI Factory architecture: stack layers, their functions, and relationships among infrastructure, platform, models, applications, and workload profiles.

10%
Exam Weight

Physical AI concepts, understanding simulation and digital twins for robotics, and identifying NVIDIA Physical AI model families and use cases.

• Physical AI versus digital AI, including the three-computer pattern, robotics and autonomous-machine use cases, and edge inference.
• Simulation and digital twins for robotics: NVIDIA Omniverse™, NVIDIA Isaac Sim™, synthetic data, domain randomization, the sim-to-real gap, and the sim-to-real flywheel.
• NVIDIA Physical AI model families and use cases, including NVIDIA Cosmos and NVIDIA Isaac™ GR00T.

6%
Exam Weight

Model lifecycle management, AI monitoring concepts and metrics, AI Factory operations, compute parameters and sizing for AI workloads, and observability components for deployed systems.

• Model lifecycle management: lifecycle and release management, retraining triggers, and versioning.
• AI monitoring concepts and metrics: observability, post-training drift, latency, throughput, health indicators, SLA breach detection, MTTD, and MTTR.
• Compute sizing for AI workloads: GPU memory, replicas, horizontal versus vertical scaling, concurrent users, throughput, and latency SLAs.
• AI Factory operations, including toolchain integration, runbooks, and playbooks.
• Compute parameters that affect workload performance: batch size, maximum sequence length, precision settings, hardware utilization and cost tradeoffs, and Run:AI compute management.
• Observability components for deployed systems: dashboards, prompt/agent/retrieval tracing, logging and alerting, telemetry, drift detection, and anomaly detection.

10%
Exam Weight