AI0-001 Implementing AI Solutions Practice Test 3 — 15 Questions

Question 1

A team is designing an AI microservice for image classification. Which THREE practices should they implement for effective integration and testing?

Accepted Answer

Use streaming responses for real-time feedback. Option A is correct because streaming responses (e.g., chunked transfer encoding or server-sent events) let the microservice return partial inference results as they are produced, which is essential for real-time feedback in interactive image-classification scenarios. Option B is correct because unit tests on the data preprocessing pipeline verify deterministic transformations such as resizing, normalization, and tensor conversion before data reaches the model, catching input-shape or scaling bugs early. Option E is correct because integration tests exercise the API endpoint that fronts the model, validating request/response contracts, serialization, error handling, and end-to-end wiring between the service and the model. Option C does not belong because training on the full dataset before deployment is a modeling workflow choice, not an integration or testing practice, and it risks data leakage if the full set includes test data. Option D does not belong because a monolithic deployment contradicts the microservice design goal and makes isolated, scalable testing harder rather than simpler.

Answer

Train the model on the full dataset before deployment

Answer

Deploy the model as a monolithic application to simplify testing

Question 2

A company is building a code generation assistant for internal developers. They want the assistant to generate code snippets consistent with the company's coding style and use private libraries. They have a few thousand examples of internal code. Which THREE considerations are critical when deciding between fine-tuning a base LLM and using RAG?

Accepted Answer

RAG eliminates the need for any model updates when private libraries change, because it retrieves the latest documentation at inference time.. Option C is correct because RAG decouples knowledge from model parameters: when private libraries change, you simply update the external vector store/index and the retriever fetches the latest documentation at inference time, so no retraining or weight updates are needed. Option D is correct because with RAG the sensitive internal code stays in the retrieval corpus rather than being baked into the model's weights, which reduces the risk of proprietary code leaking through the model and simplifies access control and data-governance compliance. Option E is correct because fine-tuning adjusts the model's weights on the company's examples, allowing it to internalize coding conventions, naming patterns, and idioms so that style guidance in the prompt becomes largely unnecessary. Option A is not correct because a few thousand high-quality examples can meaningfully adapt a model via fine-tuning (especially with parameter-efficient methods like LoRA), so millions are not strictly required. Option B is not correct because while a larger context window helps fit more retrieved snippets, RAG works with modest context windows by retrieving only the top-k most relevant chunks, so a high context window is not a critical deciding factor.

Answer

Fine-tuning a few thousand examples is insufficient; millions are required for any meaningful adaptation.

Answer

RAG requires the model to have a high context window size to accommodate retrieved code snippets.

Question 3

A financial services firm is implementing an AI system to detect fraudulent transactions. The model must be explainable to regulators, who require understanding of why a transaction was flagged. The team is choosing between a gradient-boosted tree model and a deep neural network. Both have similar accuracy. Which factor is MOST important in this decision?

Accepted Answer

The availability of SHAP (SHapley Additive exPlanations) values for the gradient-boosted tree model.. In a regulated environment, explainability is paramount. SHAP values for gradient-boosted trees offer consistent, locally accurate explanations that regulators can review, directly meeting the requirement to understand why a transaction was flagged. While both models can use SHAP, tree-based models often produce more stable explanations, making them the better choice when accuracy is comparable.

Answer

The deep neural network's ability to be trained on GPUs for faster experimentation.

Answer

The lower latency of the gradient-boosted tree model during inference.

Answer

The ability of the deep neural network to automatically learn feature interactions without manual engineering.

Question 4

A hospital's AI triage assistant must never expose protected health information to a third-party large language model provider. The solution runs on Azure and the team wants to use an LLM for summarizing clinical notes without sending data outside the hospital's compliance boundary. Which Azure service configuration BEST meets this requirement?

Accepted Answer

Deploy an open-weight model to Azure Machine Learning managed online endpoints inside a virtual network with private endpoints and no public egress.. Keeping protected health information inside the compliance boundary requires that inference happen on infrastructure the hospital controls. Deploying an open-weight model on Azure Machine Learning managed online endpoints within a virtual network with private endpoints ensures prompts and clinical notes never traverse the public internet or a third-party service. Provisioned throughput and customer-managed keys change cost and encryption posture but not the data boundary, and the public API plainly violates it.

Answer

Use Azure AI Language's summarization API and enable customer-managed keys on the storage account that holds the notes.

Answer

Call the public OpenAI API and rely on the provider's zero data retention policy for enterprise customers.

Answer

Deploy the model through Azure OpenAI Service with a provisioned throughput deployment in the hospital's own Azure subscription.

Question 5

A data science team is preparing a dataset for a binary classification model. The dataset has 95% negative class and 5% positive class. Which technique should they apply to avoid biased model predictions?

Accepted Answer

Apply resampling techniques such as SMOTE or random undersampling. The dataset is severely imbalanced (95% negative vs. 5% positive), which causes classifiers to favor the majority class and produce biased predictions. Resampling techniques such as SMOTE (Synthetic Minority Over-sampling Technique) generate synthetic minority-class samples, while random undersampling reduces majority-class samples, rebalancing the class distribution so the model learns both classes effectively.

Answer

Normalise all numerical features to a [0,1] range

Answer

Shuffle the dataset randomly before splitting into train and test sets

Answer

Remove all rows with missing values

Question 6

In the AI project lifecycle, which phase involves partitioning the dataset into training, validation, and test sets?

Accepted Answer

Data preparation. Data preparation includes splitting the data to evaluate model performance and prevent leakage.

Answer

Data acquisition

Answer

Model selection

Answer

Problem definition

Question 7

An AI system uses a pre-trained image classification model to detect defects in manufacturing. The team wants to deploy the model in an edge device with limited GPU memory. Which technique should they consider first?

Accepted Answer

Apply quantization to reduce model size. Quantization reduces the precision of the model's weights and activations (e.g., from 32-bit floating point to 8-bit integer), which significantly shrinks the model size and memory footprint while often maintaining acceptable accuracy. This is the most direct and effective first step for deploying a pre-trained model on an edge device with limited GPU memory, as it requires no retraining and immediately addresses the memory constraint.

Answer

Train the model from scratch using a smaller dataset

Answer

Use a larger model with more parameters for higher accuracy

Answer

Increase the batch size to improve throughput

Question 8

In prompt engineering, which technique involves providing a few correct input-output examples in the prompt to guide the model's response?

Accepted Answer

Few-shot prompting. Few-shot prompting provides a small number of input-output examples directly in the prompt so the model can infer the desired task format and pattern without any weight updates. This is distinct from zero-shot (no examples) and chain-of-thought (which elicits step-by-step reasoning rather than demonstrating examples).

Answer

System prompt engineering

Answer

Chain-of-thought prompting

Answer

Zero-shot prompting

Question 9

A recommendation system for an e-commerce site is producing stale suggestions that do not reflect recent user behavior. The system is updated offline every 24 hours. Which change would MOST directly address this issue?

Accepted Answer

Implement online learning to update the model incrementally in real time. Implementing online learning to update the model incrementally in real time directly addresses the staleness issue by allowing the model to incorporate recent user behavior as it happens. Online learning updates model parameters continuously or at short intervals, so recommendations reflect the latest interactions. This is the most direct solution to the problem of a 24-hour offline update cycle.

Answer

Increase the number of features used in the model

Answer

Add more training data from the past year

Answer

Use a deeper neural network architecture

Question 10

During testing a chatbot, the QA team observes that the bot sometimes responds with harmful content when given adversarial prompts. Which type of testing should be prioritised to catch these edge cases?

Accepted Answer

Red-teaming and adversarial testing. Red-teaming and adversarial testing are specifically designed to probe an AI system for vulnerabilities, including generating harmful or unsafe outputs from adversarial prompts. This approach simulates real-world attacks to uncover edge cases that standard functional tests miss, making it the correct priority for catching harmful content in a chatbot.

Answer

Unit tests for data pipeline functions

Answer

Regression testing on previously fixed bugs

Answer

Integration tests for API connectivity

Question 11

A company is fine-tuning an LLM for a domain-specific task using LoRA. They have limited GPU memory and need to reduce memory footprint without sacrificing fine-tuning quality. Which approach should they consider?

Accepted Answer

Use QLoRA with 4-bit quantized base model. QLoRA combines 4-bit NormalFloat quantization of the base model with LoRA adapters, drastically reducing GPU memory usage while preserving fine-tuning quality through techniques like double quantization and paged optimizers. This directly addresses the constraint of limited GPU memory without sacrificing the model's ability to learn domain-specific tasks effectively.

Answer

Use a larger batch size to speed up training

Answer

Fine-tune all layers of the base model

Answer

Increase the rank of LoRA adapters

Question 12

Which similarity measure is commonly used in vector search to find the angle between vectors, making it well-suited for high-dimensional embeddings?

Accepted Answer

Cosine similarity. Cosine similarity measures the cosine of the angle between two vectors, making it ideal for high-dimensional embeddings because it focuses on orientation rather than magnitude. It is widely used in vector search and NLP tasks where vector direction encodes semantic meaning.

Answer

Manhattan distance

Answer

Euclidean distance

Answer

Dot product

Question 13

A team is designing an AI agent that needs to interact with external APIs, search the web, and perform multi-step reasoning. Which TWO architectural components are essential for this agentic workflow? (Choose TWO.)

Accepted Answer

ReAct pattern (Reasoning + Acting). The ReAct pattern (Reasoning + Acting) is essential because it interleaves chain-of-thought reasoning steps with actions, allowing the agent to plan, invoke tools, observe results, and revise its plan across multiple steps—exactly what multi-step reasoning with external APIs and web search requires. Tool use / function calling is equally essential because it provides the mechanism for the model to invoke external APIs and web search functions with structured arguments and receive structured results back into the reasoning loop. Together, ReAct supplies the iterative reasoning-and-action control flow while function calling supplies the concrete interface to external systems. Fine-tuning the base model is not required for this workflow, since tool use and reasoning patterns can be implemented via prompting and orchestration without retraining weights. Single-turn response generation is insufficient because the scenario demands multi-step iteration rather than one-shot answers. A static prompt with no iterations also fails, as the agent must dynamically observe tool outputs and loop through reasoning cycles.

Answer

Fine-tuning the base model

Answer

Single-turn response generation

Answer

Static prompt with no iterations

Question 14

An organisation is developing a document intelligence system that extracts information from scanned invoices. Which THREE data preparation steps are critical to ensure high extraction accuracy? (Choose THREE.)

Accepted Answer

Cleaning and correcting OCR output. Option A (Cleaning and correcting OCR output) is correct because OCR on scanned invoices inevitably introduces character-level errors, and fixing those errors before feeding text to the extraction model directly improves field-level accuracy. Option D (Annotating bounding boxes and field labels) is correct because supervised document intelligence models need labelled ground truth that ties specific fields (e.g., invoice number, total) to their spatial locations to learn accurate extraction. Option E (Image preprocessing such as deskewing and binarisation) is correct because scanned invoices often suffer from rotation, noise, and uneven lighting, and correcting these at the image level raises OCR quality and downstream extraction accuracy. Option B (Removing punctuations and stopwords) is not appropriate because invoice fields such as dates, currency amounts, and vendor names rely on punctuation and specific tokens, so removing them would destroy critical information. Option C (Normalising all text to lowercase) is not appropriate because case can carry meaning in invoice data (e.g., currency codes, product identifiers, proper names), and lowercasing everything can reduce extraction fidelity.

Answer

Removing punctuations and stopwords

Answer

Normalising all text to lowercase

Question 15

A company is building an AI-powered document intelligence system to extract key fields from scanned invoices. The data contains 95% of invoices from one vendor and 5% from others. During model training, the F1 score is 0.95 on the overall test set, but the performance on the minority vendor invoices is very poor. What is the MOST likely cause?

Accepted Answer

The dataset is imbalanced, and the model is biased toward the majority class. The F1 score of 0.95 on the overall test set is misleading because 95% of invoices come from one vendor, so the model can achieve high overall accuracy by performing well on the majority class while failing on the minority class. This is a classic class imbalance problem where the model is biased toward the majority class. The poor performance on minority vendor invoices confirms that the model has not learned to generalize across vendors.

Answer

The model is overfitting on the minority class

Answer

The data has a train/test leakage problem

Answer

The feature extraction is incorrect for the minority vendor invoices