AWS Certified AI Practitioner (AIF-C01) - Practice Test 2

Batch transform is the optimal choice for scheduled, bulk processing tasks. It reads data from S3, processes it all at once without requiring a persistent endpoint, and writes results back to S3 -- making it cost-effective for end-of-day scheduled workloads. Real-time inference (A) keeps an always-on endpoint for immediate, low-latency responses, which is wasteful for scheduled jobs. Streaming inference (B) handles continuous data streams, not scheduled batches. Asynchronous inference (D) is designed for large single-request payloads with near-real-time responses, not bulk scheduled processing. See more: Amazon SageMaker

Amazon Comprehend Medical is a HIPAA-eligible NLP service purpose-built to extract medical entities such as medications, dosages, diagnoses, tests, and procedures from unstructured clinical text. It also identifies Protected Health Information (PHI). Amazon Kendra (A) is an intelligent enterprise search service, not an NLP extraction tool. Amazon Transcribe Medical (C) converts medical speech to text -- it doesn't analyze existing text documents. Amazon Lex (D) builds conversational chatbots and cannot perform medical entity extraction. See more: AWS Managed AI Services

AWS Batch is a fully managed service designed specifically for batch computing workloads. It automatically provisions the right amount of compute resources, manages job queues, and scales dynamically -- all without manual infrastructure management. This makes it ideal for cost-optimized batch image classification. Amazon EC2 (A) requires manual instance management, scaling, and job scheduling. AWS Lambda (B) has a 15-minute execution limit and limited compute, making it unsuitable for large batch jobs. Amazon ECS (D) requires managing container clusters and adds operational overhead. See more: AWS Cloud Computing

RAG dynamically retrieves relevant content from the firm's actual documents before generating a response, ensuring answers are grounded in real data. This also handles updates easily -- when policies change, only the document store needs updating, not the model. Zero-shot prompting (A) uses the model's built-in knowledge with no examples, but cannot pull from the firm's private documents. Model pre-training (B) requires training from scratch on massive data and is extremely expensive. Fine-tuning (C) embeds knowledge into model weights, which becomes stale when documents change and can still hallucinate details. See more: Bedrock & Generative AI

The context window (measured in tokens) defines the total amount of text an LLM can hold and process at one time, including the system prompt, conversation history, and generated output. For example, a model with a 200K token context window can process roughly 150,000 words. Temperature (A) controls randomness in output generation. Top P (B) is nucleus sampling, which restricts token selection to the most probable candidates. Embeddings (D) are dense numerical vector representations of text used for semantic search and similarity -- not a limit on processing capacity. See more: Bedrock & Generative AI

Amazon Comprehend analyzes text to detect sentiment (positive, negative, neutral, or mixed), meeting the sentiment detection requirement. Amazon Bedrock provides access to foundation models capable of generating accurate, contextually relevant summaries based on the detected sentiment. Amazon Transcribe (A) converts speech to text and cannot perform sentiment analysis. Amazon Rekognition (C) analyzes images and videos, not articles. Amazon Polly (E) converts text to speech -- neither can perform NLP analysis or text generation. See more: Bedrock & Generative AI

Temperature controls the randomness of token selection during generation. Decreasing temperature toward 0 makes the model select the highest-probability tokens more consistently, producing focused, deterministic, and repeatable outputs. Increasing the context window (A) allows more text to be processed but doesn't affect output randomness. Number of epochs (C) is a training hyperparameter, irrelevant during inference. Increasing Top K (D) actually expands the pool of candidate tokens, which increases output variation rather than reducing it. See more: Prompt Engineering

AWS Glue is a fully managed, serverless ETL service specifically designed for extracting, transforming, and loading data. It uses crawlers to automatically discover schemas, a centralized data catalog for metadata management, and Apache Spark-based transformation jobs -- all without managing infrastructure. Amazon QuickSight (A) is a BI and visualization tool for creating dashboards, not ETL. Amazon Athena (B) queries data in place using SQL but does not move or transform it. Amazon Redshift (D) is a data warehouse optimized for analytics, not a general-purpose ETL tool. See more: AWS Cloud Computing

Inference latency measures the time from when a user submits a query to when the chatbot delivers its response. This is the critical performance metric for real-time, user-facing applications -- directly impacting user experience and satisfaction. Loss (A) measures how well the model learned during training and is irrelevant in production. F1 score (B) evaluates classification accuracy, not response speed. Training accuracy (D) measures correctness during model training, which is separate from production performance. See more: Bedrock & Generative AI

Amazon Bedrock Agents orchestrate multi-step workflows by breaking complex tasks into sub-steps, invoking external APIs through action groups, accessing knowledge bases, and maintaining context across interactions. They enable autonomous task execution without human intervention at each step. Guardrails (A) apply safety filters to inputs and outputs but don't execute tasks. Knowledge Bases (C) provide RAG-based data retrieval but cannot orchestrate multi-step actions. Model Evaluation (D) assesses model quality using test prompts and does not run workflows. See more: Bedrock & Generative AI

Prompt engineering -- crafting precise system prompts, user instructions, role definitions, and format specifications -- is the primary mechanism for controlling LLM output style, tone, and format at inference time without any retraining. Learning rate (A) is a training hyperparameter that controls how fast model weights update during training. Number of epochs (B) determines how many times the model trains over the dataset -- also a training concept. Context window (D) is a fixed model property defining max input size, not output style. See more: Prompt Engineering

Inference is the production phase where a trained model receives new, unseen input data and generates predictions, classifications, or generated content. It occurs after training is complete. Pre-training (A) is the initial large-scale training of a foundation model on broad datasets. Fine-tuning (B) adapts an already-trained model to a specialized task using additional task-specific data. Validation (D) is a process during training that evaluates model performance on a held-out dataset to prevent overfitting -- not a production prediction phase. See more: AI/ML Fundamentals

Text classification using NLP assigns predefined category labels to text inputs. In this case, a classifier trained on historical tickets learns to map message content to departments (billing, technical, returns). AWS services like Amazon Comprehend support custom classification. Anomaly detection (A) identifies unusual patterns rather than categorizing into predefined classes. Clustering (B) groups similar items without predefined labels -- it's unsupervised and wouldn't use known department categories. Linear regression (D) predicts continuous numerical values, not discrete categories. See more: AI/ML Fundamentals

Object detection models identify and locate specific objects within images by drawing bounding boxes around them and classifying each detected object. For defect detection on an assembly line, object detection can locate and classify defects (scratch, dent, crack) with precise coordinates. Amazon Rekognition provides pre-built object detection capabilities. Image segmentation (C) classifies each pixel but is typically more complex than needed for locating defects. Speech recognition (A) converts audio to text and Text generation (B) produces text output -- neither applies to image defect detection. See more: AWS Managed AI Services

Auditing training data for demographic imbalances (using tools like Amazon SageMaker Clarify) identifies skewed representation before the model learns from biased patterns. Evaluating outputs across diverse demographic groups tests whether the model treats different groups equitably in actual recommendations. Reducing training time (A) might cause underfitting but doesn't remove bias. Lowering temperature (C) reduces randomness but doesn't change the underlying biased patterns the model learned. Adding layers (E) increases model capacity but doesn't address data-level bias. See more: AI Challenges & Responsibilities

Prompt chaining connects multiple LLM calls sequentially, passing the output of one prompt as the input to the next. Each step in the chain builds on previous results: translate -> summarize -> generate action items. This enables complex multi-step reasoning that a single prompt cannot achieve. Zero-shot prompting (A) uses a single prompt with no examples. Few-shot prompting (B) provides examples within a single prompt. Negative prompting (C) instructs the model what to avoid in a single prompt -- none of these chain outputs between separate LLM calls. See more: Prompt Engineering

Domain adaptation fine-tuning trains the model on domain-specific pharmaceutical text, teaching it the specialized vocabulary, abbreviations, and concepts used in drug research. Since prompt engineering has already been exhausted, the model fundamentally lacks the domain knowledge -- fine-tuning addresses the root cause. Switching to a larger general model (A) may help marginally but won't solve specialized terminology gaps. Increasing temperature (C) adds randomness, which could make the model less reliable with technical content. Removing technical terms (D) would eliminate the critical information needed to answer pharmaceutical questions accurately. See more: Bedrock & Generative AI

A confusion matrix captures true positives (fraud correctly flagged), false positives (legitimate transactions wrongly flagged), true negatives, and false negatives (missed fraud). From it, precision and recall are derived, which are critical for fraud detection -- especially given class imbalance where fraud cases are rare. High recall minimizes missed fraud; high precision minimizes false alarms. MSE (A) and R2 (B) are regression metrics for continuous predictions, not binary classification. A correlation matrix (D) shows linear relationships between numerical variables, not classification outcomes. See more: AI/ML Fundamentals

Amazon Bedrock Knowledge Bases implement RAG by ingesting private documents, chunking and embedding the content into a vector store, and dynamically retrieving relevant passages at inference time to ground responses in actual company data. This keeps the model's answers accurate and up-to-date with internal content. Adjusting inference parameters (A) affects output style but doesn't connect the model to private documents. Invocation logging (B) records API calls for monitoring, not knowledge integration. Switching models (D) changes the foundation model but doesn't solve the knowledge access problem. See more: Bedrock & Generative AI

Amazon S3 is the only AWS storage service that Amazon Bedrock integrates with for training and validation datasets. Bedrock reads data directly from S3 buckets in supported formats such as JSONL. Amazon EFS (A) is a managed file system for EC2 and Lambda -- Bedrock cannot directly read from it. Amazon DynamoDB (C) is a NoSQL database for low-latency key-value lookups, not a dataset repository for Bedrock. Amazon EBS (D) provides block-level storage attached to EC2 instances and is not accessible to Bedrock. See more: Bedrock & Generative AI

Data protection and encryption are core requirements of HIPAA (protecting PHI) and SOC 2 (security principle). This includes KMS encryption, access controls, and data handling policies. Threat detection and monitoring using services like Amazon GuardDuty and AWS CloudTrail fulfills audit and security monitoring requirements mandated by both HIPAA and SOC 2. Auto scaling (A) and cost tags (B) are operational/financial practices, not compliance evidence. Microservices architecture (E) is a design pattern, not a compliance control. See more: AI Challenges & Responsibilities

AWS PrivateLink creates a VPC interface endpoint that enables private, direct communication between the VPC and Amazon Bedrock entirely over the AWS private network -- with zero internet traffic. This is the standard solution for VPCs with internet restrictions. A NAT Gateway (A) routes traffic through the public internet via an Elastic IP, violating the no-internet policy. Amazon CloudFront (C) is an internet-facing CDN that would introduce internet traffic. VPC peering (D) connects two VPCs together but does not provide access to AWS services like Bedrock. See more: AWS Security Services

Amazon Comprehend is an NLP service that provides key phrase extraction, entity recognition, topic modeling, and sentiment analysis across large volumes of text. Its topic modeling capability clusters similar posts to reveal recurring themes and trends in social media data. Amazon Transcribe (A) converts speech audio to text -- it doesn't process existing text. Amazon Polly (B) is a text-to-speech service. Amazon Rekognition (D) analyzes images and videos, not text content. See more: AWS Managed AI Services

Evaluating model accuracy requires comparing its predictions against ground truth labels in a held-out benchmark dataset. Metrics like accuracy, precision, recall, and F1 score quantify the classification quality. Using a standard benchmark also enables fair comparison across models. Parameter count (A) reflects model size and complexity, not predictive performance. Processing time (C) measures computational efficiency, not classification accuracy. Storage cost (D) is a financial consideration, unrelated to model quality. See more: AI/ML Fundamentals

Creative writing style alignment is a subjective, brand-specific quality that cannot be measured by generic benchmarks. Amazon Bedrock's human evaluation workflow allows internal reviewers to rate model outputs based on brand voice, tone, and quality using prompts drawn from real marketing use cases. Public benchmarks (A) rank models on general language tasks, not brand-specific style. Lowest latency (B) optimizes speed, not creative quality. Most parameters (D) indicates model capacity but doesn't predict which model best fits a specific brand's voice. See more: Bedrock & Generative AI

Average time to resolve customer inquiries directly measures how much faster agents answer questions with AI assistance -- the core productivity metric. If the AI assistant reduces resolution time, it demonstrates a concrete productivity gain. Employee satisfaction (A) is a useful HR metric but doesn't directly quantify the AI's operational impact on productivity. Number of model parameters (C) is a technical attribute unrelated to business outcomes. Website session duration (D) measures online engagement, not agent efficiency. See more: AI/ML Fundamentals

When sensitive data is present in training data, it becomes encoded in the model's weights and may be reproduced in responses. The only reliable fix is to delete the compromised model, remove all sensitive data from the training dataset, and retrain from scratch on the clean dataset. Post-processing filters (A) might catch some outputs but cannot guarantee the model won't expose sensitive data in novel ways. Encrypting weights (C) secures the model file but doesn't prevent the model from outputting sensitive information during inference. Guardrails (D) provide a partial defense but cannot comprehensively block all sensitive numeric patterns. See more: AI Challenges & Responsibilities

A hardened system prompt that instructs the model to recognize and reject prompt injection patterns, ignore instructions to override guidelines, and refuse requests that attempt to change its behavior is the most direct defense against adversarial prompting. Decreasing temperature (A) makes outputs more deterministic but doesn't remove the model's susceptibility to well-crafted adversarial inputs. Limiting output tokens (C) restricts response length but doesn't prevent the model from being manipulated. Restricting to SageMaker JumpStart models (D) limits model choice but doesn't address prompt-level security. See more: AI Challenges & Responsibilities

Transfer learning takes a model that has already learned general features (edges, textures, shapes) from broad training and adapts it to a new related task -- in this case, industrial equipment recognition -- using a much smaller labeled dataset. This avoids training from scratch, saving significant time and compute. Unsupervised clustering (A) groups similar items without labels and is not a model adaptation strategy. Increasing the learning rate indiscriminately (B) can destabilize training and destroy previously learned features. Reducing parameters (D) simplifies the model architecture but doesn't leverage pre-trained knowledge. See more: AI/ML Fundamentals

Using a fully managed SaaS product with built-in AI features means the vendor controls the model, infrastructure, data handling, and security layers -- leaving the company with minimal security responsibilities. This represents the least ownership. Building from scratch (A) gives maximum ownership; the company is responsible for everything from data to infrastructure. Fine-tuning (B) shares responsibility with the FM provider. Building on an FM API (C) offloads model security but retains application-level responsibilities. Least to most ownership: SaaS -> API -> Fine-tuning -> From scratch. See more: AI Challenges & Responsibilities

Amazon SageMaker JumpStart offers one-click deployment of pre-trained foundation models, and SageMaker endpoints host those models within your VPC for secure, managed, low-latency inference without managing any container orchestration infrastructure. Amazon Personalize (A) is a recommendation service for personalization use cases, not general FM deployment. Amazon EKS (B) requires managing Kubernetes clusters -- adding significant operational overhead. PartyRock (D) is a public, browser-based playground for generative AI exploration, not suitable for VPC-based production deployment. See more: Amazon SageMaker

Amazon Bedrock charges per input and output token with On-Demand pricing. Reducing the number of few-shot examples (e.g., from 12 to 4) directly shrinks the input token count per request, lowering the cost of each invocation. Since the model already performs well, fewer examples may still yield acceptable quality. Provisioned Throughput (A) guarantees dedicated capacity with hourly commitments -- far more expensive for a weekly invocation. Fine-tuning (B) incurs training costs and requires Provisioned Throughput, which is cost-prohibitive for a once-weekly call. Invocation logging (D) records API calls but doesn't affect cost. See more: Bedrock & Generative AI

Amazon Transcribe converts recorded call audio into text transcripts (with support for speaker diarization and custom vocabularies). Amazon Comprehend then analyzes the transcripts to extract key phrases, detect named entities (product names), and identify sentiment -- providing all the requested insights. Amazon Lex (A) builds chatbots for interactive conversations, not call recording analysis. Amazon Rekognition (C) analyzes images and video, not audio. Amazon Polly (D) generates speech from text -- the opposite of what's needed. See more: AWS Managed AI Services

AWS Artifact is the central self-service portal for accessing AWS compliance reports, certifications (SOC, PCI, ISO), and security documents from third-party auditors and ISVs. It provides on-demand document access and notification capabilities when new reports become available. AWS Trusted Advisor (A) provides best-practice recommendations for cost, security, and performance -- not compliance documentation. Amazon Inspector (B) scans workloads for software vulnerabilities and network exposures, not compliance report distribution. AWS Security Hub (D) aggregates security findings across services but doesn't distribute third-party compliance reports. See more: AWS Security Services

The F1 score is the harmonic mean of precision (percentage of predicted entities that are correct) and recall (percentage of actual entities correctly identified). It balances both concerns -- important for NER where you want to minimize both false positives and missed entities. A model with high precision but low recall (finds few entities) or high recall but low precision (generates many false positives) would have a low F1 score. Inference latency (A) measures speed. Training loss (B) measures learning progress during training. R2 score (D) evaluates regression models, not classification/extraction tasks. See more: AI/ML Fundamentals

BERT was pre-trained using Masked Language Modeling (MLM), where random tokens are masked and the model learns to predict them from bidirectional context. This capability makes BERT-based models ideal for predicting corrupted or missing words in historical documents. Time series models (A) predict numerical values over time, not text words. GANs (C) generate new content (images, audio) but are not designed for in-context word prediction. Clustering models (D) group similar items without predicting specific values. See more: Bedrock & Generative AI

Amazon Rekognition Content Moderation uses automated ML to detect explicit nudity, graphic violence, suggestive content, and other unsafe categories in images in real time, enabling filtering before images are delivered to users. Retraining on cleaner data (B) reduces unsafe output probability but cannot guarantee safety in production. Lower temperature (C) makes outputs more predictable but doesn't specifically prevent inappropriate content. Relying on user reports (D) is entirely reactive -- users have already been exposed to harmful content before any action is taken. See more: AWS Managed AI Services

A model that performs well on training data but poorly in production is a classic sign of overfitting -- it memorized training patterns rather than learning generalizable relationships. The recommended fix is to increase training data volume and diversity so the model encounters more varied examples and learns broader patterns. Underfitting (A) causes poor performance on both training and production data -- not just production. Data leakage (C) causes artificially inflated training metrics from test data contamination, which is a different problem. Hyperparameter tuning (D) is a useful optimization step but doesn't address the fundamental cause of overfitting. See more: AI/ML Fundamentals

Two controls form the foundation of secure LLM use: clear, specific prompts with system-level safety instructions reduce the attack surface for prompt injection and unintended outputs. IAM roles configured with least privilege ensure only authorized principals can invoke Bedrock APIs, limiting exposure in the event of credential compromise. SageMaker Model Monitor (B) tracks model drift, not Bedrock security. Automated model evaluation (C) assesses quality, not security posture. Provisioned Throughput (D) manages capacity for cost and performance, not access security. See more: AI Challenges & Responsibilities

Hallucination occurs when an LLM generates text that sounds confident and coherent but contains fabricated or incorrect information -- invented statistics, false claims, or non-existent sources. The model produces plausible-sounding content because it generates based on statistical language patterns rather than factual grounding. Mitigation strategies include RAG, grounding with external knowledge, and human review. Overfitting (A) causes poor generalization to new data but doesn't explain confident factual errors. Underfitting (C) produces poor-quality outputs broadly. Data leakage (D) involves test data contaminating training, causing inflated performance metrics. See more: Bedrock & Generative AI

Tokens are the atomic units of text that LLMs process. A tokenizer splits input text into tokens, which can be whole words, subwords (e.g., 'playing' -> 'play' + '##ing'), or individual characters depending on the model. English text averages approximately 1 token per 4 characters. Token counts determine context window limits and directly drive per-API-call pricing. Security credentials (A) are API keys or authentication tokens -- a different concept entirely. Generation speed (C) is measured in tokens per second. Computational efficiency (D) is measured in FLOPS or GPU utilization. See more: Bedrock & Generative AI

Algorithmic bias occurs when the ML model itself produces systematically discriminatory outputs -- in this case, the model's learned decision boundaries disproportionately flag vehicles based on geographic proxies for socioeconomic status. This likely stems from biased historical enforcement patterns in the training data. Sampling bias (A) is a data collection problem where certain groups are over- or under-represented in the training sample. Observer bias (B) is a human cognitive bias affecting data collection by researchers. Measurement bias (D) results from flawed instruments or inconsistent measurement methodologies. See more: AI Challenges & Responsibilities

A VPC interface endpoint for Amazon Bedrock (powered by AWS PrivateLink) creates a private connection that routes all Bedrock API traffic through the AWS internal network, never touching the public internet. This satisfies the regulatory requirement. Amazon CloudFront (A) is an internet-facing CDN, which would route traffic over the public internet. An internet gateway (B) explicitly enables internet access for the VPC, violating the requirement even with security group restrictions. AWS Direct Connect (D) provides private connectivity from on-premises data centers to AWS but doesn't isolate intra-AWS service communication from the internet. See more: AWS Security Services

Amazon SageMaker Feature Store is a purpose-built repository for storing, discovering, and sharing ML features across teams. It provides an online store for real-time inference access (millisecond latency) and an offline store for batch training. Teams discover and reuse the same standardized features, ensuring consistency and eliminating redundant computation. SageMaker Clarify (A) detects bias in data and models and provides explainability reports, not feature management. Data Wrangler (B) helps with data preparation and transformation pipelines. Model Registry (C) manages model versions, metadata, and deployment approvals. See more: Amazon SageMaker

Amazon SageMaker Canvas is a no-code ML tool with a visual, point-and-click interface that lets business users without coding or ML knowledge build predictive models. Users simply import data, select target columns, and Canvas automatically handles feature engineering, algorithm selection, and training. SageMaker built-in algorithms (A, B) require Python coding and ML knowledge. Data Wrangler (B, C) requires understanding of data transformations. Amazon Personalize (C) is for recommendation engines, not demand forecasting. See more: Amazon SageMaker

Each Availability Zone (AZ) consists of one or more data centers with independent power, cooling, and networking. Deploying across multiple AZs within a single Region provides high availability and fault tolerance against data center failures. A single instance (A) has no redundancy. Multi-Region (C) is more than the minimum needed for a single data center failure. CloudFront (D) is for content caching, not application availability. See more: AWS Cloud Computing

AWS IAM is a global service -- users, groups, roles, and policies created in IAM are available across all AWS Regions without replication. EC2 instances (A), S3 buckets (B), and RDS databases (D) are all regional services where resources exist only in the region where they were created. See more: AWS Cloud Computing

Amazon Q Developer in the AWS Console can analyze your AWS bill and answer natural language questions about cost drivers, resource usage, and account state. It provides conversational access to account information without requiring knowledge of specific AWS tools or query syntax. Cost Explorer (A), Budgets (C), and CloudWatch (D) require manual navigation and configuration. See more: Amazon Q

Amazon Q Apps is a no-code GenAI application builder within Amazon Q Business. Non-technical employees can describe app requirements in natural language and Q Apps generates working applications connected to internal company data. Bedrock Playground (A) requires technical knowledge. Lambda (C) requires coding. PartyRock (D) doesn't connect to internal company data. See more: Amazon Q

Amazon Macie is a fully managed ML-powered service that scans S3 buckets to discover, classify, and alert on sensitive data -- particularly PII. It's the recommended pre-training step to ensure training data doesn't contain personal information. Comprehend (A) detects PII in text input, not S3 files. Config (C) monitors resource configuration. GuardDuty (D) detects security threats, not PII. See more: AWS Security Services