Concept guide

Certification concept guides

Exam concepts explained with 3D visuals and worked examples, focused on what actually separates the right answer from the wrong ones.

• Prompt engineering techniquesHow zero-shot, few-shot, chain-of-thought, negative prompts and templates shape model output, and the attacks that target prompts.
• Retrieval-augmented generation (RAG)Answering from documents the model has never seen: chunking, embeddings, vector search, grounding, keeping the knowledge base current, evaluation, and when to choose RAG over fine-tuning.
• How large language models workHow an LLM produces text, step by step: next-token prediction, tokens, embeddings, transformers and attention, training, temperature and top-p, and the context window.
• Fine-tuningFurther training a pre-trained model on your own data: preparing data, epochs, learning rate and overfitting, PEFT and LoRA, continued pre-training, instruction tuning, preference tuning, distillation, catastrophic forgetting, and when to choose fine-tuning.
• Evaluation metrics for AI modelsHow to tell whether a model is good enough: the accuracy trap, confusion matrix, precision, recall and F1, ROC and AUC, MAE and RMSE, BLEU, ROUGE and semantic similarity, LLM-as-a-judge with faithfulness and relevance, and A/B tests with business metrics.
• Machine learning basicsHow AI, machine learning and deep learning relate, features and labels, train/validation/test splits, supervised, unsupervised and reinforcement learning, decision trees and common algorithms, overfitting and underfitting, gradient descent and hyperparameters, and data preparation.
• AI agentsAI agents that plan and use tools to finish tasks: function calling, ReAct and planning, short- and long-term memory, multi-agent orchestration, MCP and A2A, and safeguards such as least privilege and human approval.
• Hallucinations and guardrailsWhy generative AI hallucinates and how to reduce it, plus guardrails that stop harmful requests and answers: input and output checks, content filters and severity, PII masking, and red teaming.
• Responsible AIThe principles of responsible AI and how to put them into practice: bias from data, fairness metrics and the four-fifths rule, explainability with SHAP and LIME, model cards, AI governance and risk-based regulation.
• MLOpsMLOps: the ML lifecycle, pipelines and reproducibility, model registry, batch and real-time inference with canary, shadow and blue/green rollouts, drift monitoring, retraining strategies, training-serving skew and feature stores, and running LLM services.