MLOps and ML Engineering

Derive the prediction and training objective of linear regression with regularization.

Explain logistic-regression probabilities, decision boundaries, and cross-entropy loss.

Explain how decision trees choose splits and how random forests and gradient boosting reduce error.

Diagnose underfitting and overfitting using bias–variance reasoning and learning curves.

Design a train-validation-test strategy that prevents temporal, group, and feature leakage.

Select evaluation metrics that match class imbalance, ranking needs, and business error costs.

Assess and improve the calibration of probabilistic predictions.

Conduct slice-based error analysis and turn observed failure patterns into testable model improvements.

Trace forward propagation and backpropagation through a small neural network.

Select activation functions, weight initialization, and normalization methods for stable training.

Distinguish the effects of batch size, learning rate, momentum, and adaptive optimizers on convergence.

Diagnose vanishing gradients, exploding gradients, unstable loss, and stalled training from observed signals.

Select regularization methods for a specified deep-learning overfitting problem.

Estimate the parameter count, activation memory, and computational cost of a neural-network architecture.

Choose a distributed-training strategy for a model that does not fit or train efficiently on one accelerator.

Design an ablation study that isolates the contribution of a model component or training choice.

Calculate convolution output dimensions, parameter counts, and receptive fields for a CNN.

Explain how residual connections and feature pyramids affect visual representation learning.

Compute IoU and apply non-maximum suppression to object-detection predictions.

Interpret precision-recall curves and mean average precision for an object detector.

Select architectures and metrics for semantic, instance, and panoptic segmentation tasks.

Diagnose the effects of class imbalance, occlusion, weather, and domain shift in automotive vision data.

Design an evaluation protocol for a safety-relevant perception model using scenario-based slices.

Explain trade-offs among camera-only perception, multi-sensor fusion, and latency-constrained deployment.

Compare word-, subword-, and byte-level tokenization for vocabulary size, robustness, and sequence length.

Compute scaled dot-product attention for a small example and explain its computational complexity.

Trace the flow of information through a transformer block, including residual paths and normalization.

Distinguish causal language modeling, masked language modeling, and sequence-to-sequence training objectives.

Select a pretrained language-model architecture for classification, extraction, generation, or embedding tasks.

Choose between full fine-tuning and parameter-efficient fine-tuning for a given resource constraint.

Evaluate classification, extraction, summarization, and generation systems with task-appropriate metrics.

Diagnose NLP model errors caused by tokenization, context limits, label ambiguity, and distribution shift.

Choose among prompting, retrieval-augmented generation, fine-tuning, and tool use for a specified LLM requirement.

Design a document chunking and metadata strategy for a retrieval-augmented generation system.

Select embedding, indexing, and reranking approaches for a given recall, latency, and cost target.

Construct prompts and output schemas that improve instruction following and machine-readable responses.

Design an evaluation set that separately measures retrieval quality, answer quality, groundedness, and latency.

Diagnose hallucination and retrieval failures from LLM application traces.

Mitigate prompt injection, sensitive-data exposure, and unsafe output with layered controls.

Select batching, caching, quantization, and model-routing techniques for an LLM inference workload.

Define the interfaces and artifacts connecting data, training, validation, registry, deployment, and monitoring stages.

Design versioning and lineage controls that make an ML experiment reproducible.

Define model-registry stages and promotion criteria for moving a model toward production.

Design CI, CD, and continuous-training checks for an ML repository.

Choose among batch, asynchronous, streaming, and online inference for a specified workload.

Select shadow, canary, blue-green, and rollback strategies for releasing a model safely.

Define monitoring signals for service health, data quality, drift, model quality, and business impact.

Explain the Terraform workflow and interpret how declarative infrastructure supports reproducible ML environments.

Translate an ambiguous ML product prompt into functional requirements, success metrics, constraints, and service-level objectives.

Estimate storage, throughput, accelerator, latency, and cost requirements for an ML system.

Design data ingestion, labeling, validation, and feature-management components for an ML platform.

Design training, evaluation, retraining, and model-promotion workflows for a large-scale ML system.

Design a scalable serving architecture using containers, orchestration, autoscaling, and accelerator-aware scheduling.

Design observability, fault-tolerance, fallback, and incident-recovery mechanisms for an ML service.

Map a cloud-agnostic ML architecture to suitable Azure and Databricks components.

Present an ML system design by making explicit trade-offs among quality, latency, reliability, complexity, and cost.

Deliver a concise, technically rigorous explanation of an ML concept using assumptions, equations, and practical implications.

Solve a timed model-debugging case by forming hypotheses and selecting discriminating experiments.

Solve a timed production-incident case using metrics, logs, traces, rollback criteria, and follow-up actions.

Explain an autonomous-driving CV project in terms of personal ownership, design choices, evaluation, failures, and measurable impact.

Answer behavioral questions with evidence-based stories about ambiguity, conflict, failure, leadership, and cross-functional work.

Defend an ML architecture against follow-up questions about scale, reliability, data quality, safety, and cost.

Critique a sample ML system design and prioritize its highest-risk weaknesses.

Conduct a timed mock interview and convert feedback into a targeted final-review plan.