AI0-001 Practice Test 29 — 15 Questions

Question 1

A data scientist is building a recommendation system using Apache Spark for feature engineering. They need to process streaming user click data in real-time before feeding into the model. Which tool should they use for the streaming data ingestion?

Choose an answer to begin — your selection is scored in the full session.

Accepted Answer

Apache Kafka. Apache Kafka is the correct choice because it is a distributed streaming platform designed for high-throughput, fault-tolerant, real-time data ingestion. It acts as a durable message broker that can ingest streaming click data and make it available for Spark Structured Streaming to process in micro-batches or continuous processing mode, which is essential for real-time feature engineering in a recommendation system.

Answer

Amazon S3

Answer

Airflow

Answer

Snowflake

Question 2

A machine learning engineer is troubleshooting a recurrent neural network that fails to learn long-range dependencies in sequential data. The gradients are computed using backpropagation through time. Which phenomenon is most likely occurring, and what architectural change would best address it?

Accepted Answer

Vanishing gradients; use LSTM or GRU units. In standard RNNs, backpropagation through time (BPTT) multiplies gradients across many time steps, causing them to shrink exponentially (vanishing gradients). This prevents the network from learning long-range dependencies. LSTM or GRU units introduce gating mechanisms that preserve gradient flow over many time steps, directly solving this problem.

Answer

Underfitting; increase the number of time steps

Answer

Exploding gradients; apply gradient clipping

Answer

Overfitting; reduce the number of layers

Question 3

A bank wants to ensure its credit scoring model is fair across demographic groups. The model currently uses features like zip code, income, and credit history. To mitigate potential bias, which TWO actions should the data science team prioritize?

Accepted Answer

Analyze the model for disparate impact using statistical tests. Analyzing the model for disparate impact using statistical tests (e.g., the 80% rule or chi-square test) directly measures whether the model produces systematically different outcomes for protected groups. This is a foundational step in fairness auditing, as it quantifies bias before any mitigation is applied, aligning with regulatory expectations like the Equal Credit Opportunity Act (ECOA).

Answer

Remove all features that could be correlated with protected attributes

Answer

Implement a fairness metric like demographic parity or equalized odds

Answer

Apply differential privacy to the training data

Question 4

A platform team is preparing a feature store for a recommendation system. They need point-in-time correct feature retrieval so that training datasets do not leak future information, and they need the same features served online with low latency. Which architecture best satisfies both requirements?

Accepted Answer

Maintain an offline store with time-stamped feature values for training and a synchronized online store keyed by entity for low-latency serving.. A feature store that pairs a time-stamped offline store with a synchronized online store supports point-in-time correct training joins and fast online retrieval from the same feature definitions. This dual-store pattern prevents label leakage and reduces training-serving skew. Single-store, lake-query, and duplicated-code approaches each fail at least one of correctness or latency.

Answer

Use a single relational database table that overwrites feature values in place with no timestamps.

Answer

Compute features on the fly in the training pipeline and again independently in the serving path using separate code.

Answer

Store features only in a data lake and have the online service query the lake on each request.

Question 5

A financial services firm has deployed an AI-powered document summarization service that processes internal memos. To reduce the risk of prompt injection attacks that could manipulate the model's output, the security team wants to implement a defense that inspects and filters the input text before it reaches the model. Which of the following is the MOST appropriate technique to achieve this?

Accepted Answer

Implement input sanitization by removing or escaping special characters and known prompt injection patterns.. Prompt injection attacks rely on malicious text entering the model's context. Input sanitization removes or escapes dangerous characters and known injection patterns before the model processes the input, directly mitigating the risk. Differential privacy, adversarial training, and output encoding address different concerns and do not filter input at inference time, so they fail to meet the scenario's specific requirement.

Answer

Apply differential privacy during model training to limit the influence of any single input.

Answer

Use adversarial training by generating adversarial examples and retraining the model to be robust.

Answer

Enforce strict output encoding to prevent cross-site scripting in the summarization results.

Question 6

Which similarity metric is MOST appropriate for comparing dense vector embeddings in a vector store used for document retrieval, when the embeddings are normalized to unit length?

Accepted Answer

Cosine similarity. Cosine similarity measures the angle between two vectors and is the standard metric for comparing dense embeddings in vector stores. When embeddings are normalized to unit length, cosine similarity is mathematically equivalent to the dot product, making it both efficient and semantically meaningful for document retrieval. It focuses on orientation rather than magnitude, which aligns with how embedding models encode semantic similarity.

Answer

Jaccard similarity

Answer

Manhattan distance

Answer

Euclidean distance

Question 7

A company is evaluating fairness metrics for a hiring model. They want to ensure that the model has similar true positive rates (TPR) across demographic groups. Which fairness metric should they use?

Accepted Answer

Equalized odds. Equalized odds requires that both true positive rates (TPR) and false positive rates (FPR) are equal across demographic groups, which directly matches the requirement for similar TPR across groups. It is the standard fairness metric when the goal is parity in error rates rather than parity in outcomes.

Answer

Calibration

Answer

Individual fairness

Answer

Demographic parity

Question 8

An organization is evaluating a third-party large language model to integrate into their customer-facing application. As part of supply chain security, which THREE steps should they take to vet the model before deployment?

Accepted Answer

Conduct security testing, including red teaming, to identify vulnerabilities in the model. Option A is correct because security testing such as red teaming is a core supply chain vetting step that probes the third-party LLM for prompt injection, jailbreaks, data leakage, and other adversarial vulnerabilities before it is exposed to customers. Option C is correct because reviewing the model card and documentation reveals the model's intended use, limitations, training provenance, and known biases, allowing the organization to assess whether the model is suitable and safe for its customer-facing scenario. Option E is correct because an SBOM for AI components enumerates the model's dependencies, libraries, and versions, enabling the organization to identify known vulnerabilities and manage supply chain risk. Option B does not belong because federated learning is a training technique for building or adapting models on distributed internal data, not a vetting step for evaluating a third-party model. Option D does not belong because running a model inversion attack is an offensive research technique that could itself compromise privacy, rather than a standard supply chain security review step.

Answer

Use federated learning to retrain the model on internal data

Answer

Run a model inversion attack on the model to verify training data privacy

Question 9

A team is deploying a model that must comply with GDPR. Users can request deletion of their data. Which TWO practices should be implemented to support this compliance? (Select TWO.)

Accepted Answer

Maintain the ability to delete a user's data from training sets and derived features. GDPR's 'right to erasure' requires that upon user request, the organization must delete not only the user's raw data but also any derived features or embeddings that were generated from that data. Without this capability, the model could still indirectly retain user information through trained parameters or feature stores, violating compliance.

Answer

Enable output caching for frequently requested predictions

Answer

Validate inputs to prevent prompt injection attacks

Answer

Use a vector database to store user embeddings

Question 10

A hospital's radiology department is deploying an AI system that analyzes chest X-rays to flag potential pneumonia. Because patient data cannot leave the hospital's on-premises network, the model must run locally. The IT team wants to ensure the model's inference results can be explained to radiologists and auditors. Which approach best satisfies the explainability requirement while keeping the model on-premises?

Accepted Answer

Apply LIME or SHAP to generate local explanations for each prediction without modifying the model.. Post-hoc explainability techniques such as LIME and SHAP allow clinicians to understand individual predictions from complex models without sacrificing accuracy. They operate locally and do not require moving data off-premises, satisfying both the explainability and data residency requirements. Retraining a simpler model risks accuracy, while confidence scores and cloud dashboards fail to provide the needed transparency under the given constraints.

Answer

Use a black-box deep learning model and provide a confidence score for each prediction.

Answer

Deploy the model to a cloud service that offers built-in explainability dashboards.

Answer

Replace the deep learning model with a logistic regression classifier trained on the same data.

Question 11

A retail analytics team is preparing a recommendation model for production. They need to serve many concurrent requests with predictable latency and also reduce the cost of running the model on GPU nodes. Which TWO practices best support these goals? (Choose two.)

Accepted Answer

Batch multiple incoming requests into a single forward pass within a short time window.. Throughput and cost on GPU nodes improve when each forward pass does more useful work and each operation is cheaper. Dynamic batching lets one pass serve several concurrent requests, while lower-precision execution reduces the cost of every operation. Together they raise requests served per GPU-second, which is the lever that lowers cost per prediction while keeping latency bounded.

Answer

Retrain the recommendation model with a larger embedding dimension to improve ranking quality.

Answer

Store the model on a network file share so all replicas read the same artifact at startup.

Answer

Increase the number of replicas so every request is handled by a dedicated pod with no queuing.

Question 12

A retail analytics team is building a retrieval-augmented generation assistant over product manuals. They need a vector index that supports fast approximate nearest neighbor search and can be updated as new manuals are published without rebuilding the entire index. Which TWO components should they use to meet these requirements? (Choose two.)

Accepted Answer

A dedicated vector database such as Milvus or Pinecone that supports incremental upserts and ANN indexes like HNSW.. A working retrieval pipeline needs embeddings to represent manual chunks semantically and a vector store that indexes those embeddings for fast approximate similarity search while allowing incremental additions. Combining an embedding model with a vector database that supports upserts and HNSW-style indexes satisfies both the speed and the update-without-rebuild requirements.

Answer

A columnar analytics warehouse that stores embeddings as arrays for SQL aggregation.

Answer

A message queue that streams manual PDFs directly into the index without transformation.

Answer

A relational database with B-tree indexes on the manual text column.

Question 13

A company is building an AI-powered document processing system that extracts information from scanned PDFs. The system must handle varying document layouts and languages. The team wants to use a pre-trained model and fine-tune it on their own data. Which TWO techniques are most appropriate to improve the model's ability to generalize to new document layouts? (Choose two.)

Accepted Answer

Data augmentation with random rotations, scaling, and cropping of document images.. Data augmentation with geometric transformations exposes the model to layout variations, while layout-aware pre-training objectives help the model learn structural relationships. Together, they enhance generalization to unseen document formats. Other options focus on training efficiency or model compression, which do not directly address layout variability.

Answer

Reducing the model size by pruning 50% of the weights before fine-tuning.

Answer

Fine-tuning all layers of the model with a very low learning rate.

Answer

Using a larger batch size during training to stabilize gradients.

Question 14

An AI platform team is deploying a large language model for internal document summarization. Legal requires that no prompt or document content leaves the company's virtual private cloud, and the security team wants to control the exact model weights and runtime version. The team already has GPU capacity reserved in their own VPC. Which deployment approach best satisfies these constraints?

Accepted Answer

Self-host the model weights on the reserved GPUs inside the VPC and expose an internal inference endpoint. The constraints combine data residency, model weight control, and runtime version control, and the team already owns reserved GPU capacity. Running the weights on that capacity inside the VPC keeps all payloads private and lets the team pin both the weight revision and the serving runtime. Managed APIs and serverless services, even with private endpoints or encryption, still process content on provider infrastructure.

Answer

Use a provider's managed private endpoint with a business associate agreement

Answer

Deploy the model to a serverless inference service in a different cloud region

Answer

Call a public foundation model API using customer-managed encryption keys

Question 15

A healthcare analytics team is preparing to fine-tune a 7-billion-parameter open-weight language model on a single server with four NVIDIA A100 40 GB GPUs. Full fine-tuning runs out of memory, and the team wants to train on their clinical notes dataset while keeping GPU memory within the available budget. Which TWO techniques should they apply to reduce memory consumption during fine-tuning? (Choose two.)

Accepted Answer

Gradient checkpointing. Parameter-efficient tuning with LoRA removes the need to store optimizer state for all base parameters, and gradient checkpointing cuts activation memory by recomputing intermediates during the backward pass. Together they target the two dominant memory consumers in fine-tuning, making a 7B model trainable on four 40 GB GPUs without changing the base weights.

Answer

Converting the model to ONNX format before training

Answer

Quantization-aware training of the base weights at 4-bit

Answer

Increasing the global batch size