Mixed Deck — All CAIC Topics Flashcards

When should an AI consultant recommend using serverless inference (e.g., AWS Lambda, Google Cloud Run) instead of a dedicated GPU instance?

Answer: For lightweight, infrequent inference tasks where cold starts are acceptable

Serverless is cost-effective for sporadic, lightweight inference (e.g., small classifiers) where paying for idle GPU capacity would be wasteful.

Which AI observability tool is specifically designed to trace, evaluate, and debug LLM applications end-to-end?

Answer: LangSmith

LangSmith, built by the LangChain team, provides tracing, evaluation, and debugging specifically for LLM-powered applications.

Which cloud security principle dictates that an AI service should only have the minimum permissions required to perform its function?

Answer: Principle of least privilege

The principle of least privilege limits the blast radius of a compromised service by restricting its IAM permissions to only what it needs.

What is the principle of least privilege in CAIC technology?

Answer: Users get only the minimum access needed for their role

Least privilege limits access to only what is needed, reducing potential damage from errors or compromised accounts.

Which business area typically sees the HIGHEST and most rapid initial ROI from AI implementation?

Answer: Repetitive process automation and customer service operations

Repetitive, high-volume processes like customer service and back-office operations offer clear labor savings and fast measurable returns, making them ideal first AI use cases.

In Kanban applied to an AI team, Work In Progress (WIP) limits PRIMARILY help by:

Answer: Reducing bottlenecks and improving flow of tasks through the pipeline

WIP limits force the team to finish in-progress work before starting new items, surfacing bottlenecks and improving overall throughput.

An AI system must maintain a consistent snapshot of training data while a bulk import job is running. Which PostgreSQL transaction isolation level provides this guarantee without blocking the import?

Answer: REPEATABLE READ

REPEATABLE READ ensures the transaction sees a consistent snapshot taken at the start of the transaction, preventing non-repeatable reads without blocking concurrent writes.

A client's analytics dashboard shows a sudden spike in daily active users. Before concluding this is real growth, what should an AI consultant check first?

Answer: Data pipeline logs for tracking anomalies or bugs

Unusual spikes in metrics often result from tracking code bugs, duplicate event firing, or bot traffic rather than genuine user growth.

What is the primary purpose of a model registry in an MLOps architecture?

Answer: Centralizing model artifact versioning, metadata, and stage promotion across the model lifecycle

A model registry tracks model versions, associated metadata (metrics, lineage, parameters), and manages promotion stages (staging → production) to enable governed model lifecycle management.

Which document formalizes the business case, objectives, and initial scope of an AI consulting project at inception?

Answer: Project charter

The project charter is the founding document that authorizes the project and defines high-level objectives and constraints.

Which visualization type is best suited for showing the distribution and spread of a continuous numeric variable?

Answer: Box plot

A box plot displays the median, quartiles, and outliers of a continuous variable, making distribution and spread immediately visible.

When managing a cross-functional AI team, which communication artifact best keeps data scientists, engineers, and business stakeholders aligned?

Answer: A model card documenting model capabilities and limitations

Model cards provide a standardized, accessible summary that bridges technical and business audiences.

A financial services firm wants to use AI for credit scoring. Which ethical consideration is MOST critical?

Answer: Ensuring the model does not perpetuate or amplify discriminatory bias against protected groups

In high-stakes decisions like credit scoring, algorithmic fairness is paramount because biased models can cause material harm and violate anti-discrimination laws.

An AI consultant is helping a retailer measure the success of a newly deployed AI-powered product recommendation engine. Which business metric is MOST relevant?

Answer: Increase in average order value and improvement in conversion rate

A recommendation engine's business purpose is to increase purchase value and conversion — average order value and conversion rate directly measure whether it is achieving its commercial objective.

An AI content moderation system over-removes posts from minority language communities. The best corrective measure is to:

Answer: Collect representative training data for underserved languages and retrain

Collecting representative data and retraining addresses the root cause — insufficient coverage of minority languages in the original training set.

What distinguishes a generative model from a discriminative model in machine learning?

Answer: Generative models learn the joint distribution P(X,Y); discriminative models learn P(Y|X)

Generative models model how data is generated (joint distribution), while discriminative models learn decision boundaries directly from input to label.

A fintech AI system must ensure that two database operations — debiting a source account and crediting a destination account — always succeed or fail together. Which database property guarantees this?

Answer: Atomicity

Atomicity ensures that all operations within a transaction are treated as a single unit — either all are committed or all are rolled back if any step fails.

Which technique helps AI project teams estimate story points for data-dependent tasks with high uncertainty?

Answer: Planning poker with explicit uncertainty ranges

Planning poker surfaces team disagreements and allows uncertainty ranges to be encoded in estimates for AI's inherent unknowns.

A generative AI model produces realistic fake images of real individuals. The primary ethical concern is:

Answer: Non-consensual synthetic media enabling defamation and identity harm (deepfakes)

Deepfakes can be weaponized for defamation, harassment, and fraud without the depicted person's consent, causing serious reputational and personal harm.

Which of the following best describes 'differential privacy' as used in AI systems?

Answer: Adding calibrated statistical noise to data or outputs to prevent individual privacy leakage

Differential privacy adds mathematically calibrated noise so that individual records cannot be identified from model outputs or aggregate statistics.