System Architecture & Design Flashcards

An enterprise wants to implement AI governance across dozens of models. Which architectural component acts as the central control plane for policy enforcement?

Answer: An AI governance layer with policy engines, audit logging, and model metadata registry

A centralized governance layer can enforce consistent policies (bias thresholds, explainability requirements, data lineage) across all models rather than relying on per-model implementations.

What is the primary architectural benefit of separating the feature engineering pipeline from the model training pipeline?

Answer: It enables features to be computed once and reused across multiple models and training runs

Decoupling feature engineering means expensive transformations are computed once and stored, enabling multiple models to consume the same features without redundant computation.

In a federated learning architecture, what problem does aggregation at the parameter server solve?

Answer: It combines locally trained model updates from edge devices without requiring raw data to leave each device

Federated aggregation averages or combines gradient updates from clients, improving the global model while keeping raw data on-device, addressing privacy and data sovereignty constraints.

When architecting a real-time fraud detection AI system, why is asynchronous processing often unsuitable for the prediction step?

Answer: Fraud predictions must block the transaction decision synchronously before the transaction completes

Fraud detection must make a block/allow decision before the transaction commits, requiring a synchronous request-response pattern with guaranteed sub-second latency at the decision point.

Which pattern best handles the cold-start problem in an AI recommendation system for new users?

Answer: Hybrid architecture combining collaborative filtering for existing users with content-based or rule-based fallbacks for new users

A hybrid system gracefully falls back to content-based or demographic-based recommendations when user interaction history is insufficient, ensuring new users still receive relevant suggestions.

In a Kappa architecture for AI, what is the key difference from Lambda architecture?

Answer: Kappa architecture eliminates the batch layer, using a single replayable streaming layer for both real-time and historical processing

Kappa architecture simplifies Lambda by using a single replayable event stream (e.g., Kafka) as the source of truth, eliminating the complexity of maintaining parallel batch and streaming code paths.

What is the role of a model serving framework (such as TorchServe, TensorFlow Serving, or Triton) in the AI system architecture?

Answer: It provides optimized runtime environments for model inference with batching, versioning, and hardware acceleration management

Model serving frameworks handle the operational concerns of inference—dynamic batching, model versioning, GPU/CPU scheduling, and health checks—so application teams don't have to build these capabilities themselves.