Curated AI interview questions covering models, datasets, evaluation, prompt workflows, AI product design, risk, and production deployment.
Artificial Intelligence is the field of building computer systems that can perform tasks that normally require human intelligence, such as understanding language, recognizing images, making decisions, planning, learning from data, and solving problems.
AI is the broad field of creating intelligent systems. Machine Learning is a subset of AI where systems learn patterns from data. Deep Learning is a subset of Machine Learning that uses multi-layer neural networks for complex pattern learning.
Narrow AI is designed to perform one specific task or a limited group of tasks, such as spam detection, face recognition, translation, or fraud detection. Most real-world AI systems today are narrow AI systems.
General AI, also called AGI, refers to an AI system that can understand, learn, and adapt across many unrelated tasks at a human-like level. It is not a proven production technology today.
The main types are supervised learning, unsupervised learning, semi-supervised learning, self-supervised learning, and reinforcement learning. Each type differs in how training signals are provided and what kind of problem it solves.
Supervised learning trains a model using examples that contain both input features and known output labels. Examples include spam classification, loan approval prediction, medical diagnosis, and price prediction.
Unsupervised learning finds patterns in data without predefined labels. It is commonly used for clustering, anomaly detection, dimensionality reduction, and customer segmentation.
Reinforcement learning trains an agent to make decisions by interacting with an environment and receiving rewards or penalties. It is used in robotics, games, resource allocation, and control systems.
Self-supervised learning creates training signals from the data itself. For example, a language model can learn by predicting missing words or the next token without needing every example manually labeled.
Training data is the dataset used to teach a model patterns. It should be representative, clean, relevant, and aligned with the real-world prediction task because poor training data usually leads to poor AI behavior.
Feature engineering is the process of creating, selecting, transforming, or combining input variables so a model can learn useful patterns. Examples include encoding categories, normalizing values, creating ratios, and extracting date features.
Model training is the process of adjusting model parameters so predictions become closer to expected outputs. It uses data, a loss function, and an optimization method to learn patterns.
Inference is the stage where a trained model is used to make predictions on new data. For example, a fraud model performs inference when it scores a new transaction.
Training is the learning phase where the model updates parameters using historical data. Inference is the prediction phase where the trained model processes new inputs. Training is usually compute-heavy; inference must often be fast and reliable.
A loss function measures how wrong a model prediction is compared with the expected answer. During training, the model tries to minimize this loss. Examples include mean squared error and cross-entropy loss.
Overfitting happens when a model learns the training data too closely, including noise, and performs poorly on unseen data. It can be reduced with more data, regularization, cross-validation, dropout, early stopping, or a simpler model.
Underfitting happens when a model is too simple or poorly trained to capture the real pattern. It performs badly on both training and test data. It can be improved with better features, a stronger model, or more training.
Bias is error from overly simple assumptions, while variance is error from being too sensitive to training data. A high-bias model underfits, and a high-variance model overfits. Good models balance both.
Data leakage happens when information that should not be available during prediction enters training or evaluation. It makes metrics look better than real-world performance. Examples include using future data or preprocessing train and test data together.
A confusion matrix shows true positives, true negatives, false positives, and false negatives. It helps explain the types of mistakes a classification model makes instead of reporting only one score.
Accuracy is the percentage of predictions that are correct. It is useful for balanced datasets but can be misleading when classes are imbalanced, such as fraud detection or rare disease prediction.
Precision measures how many predicted positive cases are actually positive. It matters when false positives are expensive, such as incorrectly blocking a legitimate payment.
Recall measures how many actual positive cases the model finds. It matters when false negatives are expensive, such as missing fraud, disease, or safety violations.
The F1 score is the harmonic mean of precision and recall. It is useful when both false positives and false negatives matter and the dataset is imbalanced.
ROC-AUC measures how well a model separates positive and negative classes across different thresholds. A higher AUC means the model ranks positive examples above negative examples more reliably.
Cross-validation evaluates a model by splitting data into multiple folds, training on some folds, and validating on the remaining fold. It gives a more stable estimate than a single train-test split.
Regularization discourages overly complex models so they generalize better. Common methods include L1 regularization, L2 regularization, dropout, and early stopping.
A neural network is a model made of connected layers of nodes that transform inputs into outputs. Neural networks are useful for learning complex non-linear patterns in images, text, audio, and tabular data.
Deep learning uses neural networks with many layers to learn representations from data. It is powerful for computer vision, speech, NLP, recommendations, and generative AI, but often needs large data and compute.
Natural Language Processing, or NLP, is the AI field focused on understanding, generating, and processing human language. Examples include translation, summarization, chatbots, sentiment analysis, and search.
Computer Vision is the AI field focused on interpreting images and videos. Examples include object detection, face recognition, OCR, medical image analysis, and quality inspection.
A recommendation system suggests items, content, or actions based on user behavior, similarity, preferences, or context. Examples include movie suggestions, ecommerce recommendations, and personalized feeds.
Knowledge representation is the way an AI system stores facts, rules, relationships, or concepts so it can reason about them. Examples include knowledge graphs, ontologies, semantic networks, and rule-based systems.
AI search and planning are methods for finding actions or paths that lead to a goal. They are used in pathfinding, scheduling, game agents, robotics, and optimization problems.
Model drift happens when model performance changes over time because real-world data changes. It can be managed with monitoring, alerts, retraining, evaluation on fresh data, and rollback plans.
Concept drift happens when the relationship between input features and target output changes. For example, fraud behavior can evolve, making older fraud rules or models less effective.
Explainable AI makes model decisions understandable to humans. It is important in domains like healthcare, finance, hiring, and insurance. Techniques include feature importance, SHAP, LIME, simpler models, and decision logs.
Responsible AI means building AI systems with fairness, transparency, privacy, accountability, safety, and human oversight. It includes bias checks, documentation, monitoring, and clear ownership.
Bias in AI is systematic unfairness or skewed behavior caused by data, labels, model design, or deployment context. It can be reduced by better data collection, fairness evaluation, audits, and human review.
Fairness means an AI system should not create unjustified harm or discrimination across users or groups. The exact fairness definition depends on the domain, legal requirements, and business risk.
Human-in-the-loop AI keeps humans involved in reviewing, correcting, approving, or overriding model decisions. It is useful when errors are costly or when ethical judgment is required.
An AI guardrail is a control that reduces unsafe or incorrect model behavior. Examples include input validation, output filters, confidence thresholds, human review, refusal rules, access control, and monitoring.
A confidence threshold is a cutoff used to decide whether a prediction is accepted, rejected, or sent for human review. Changing the threshold changes the balance between precision and recall.
Batch inference runs predictions on many records at scheduled intervals, such as nightly customer scoring or weekly demand forecasting. It is useful when immediate response is not required.
Real-time inference produces predictions immediately when a request arrives, such as fraud scoring during checkout. It requires low latency, reliability, monitoring, and scaling.
A production AI deployment includes packaging the model, validating inputs, exposing it through an API or batch job, tracking model versions, monitoring quality and latency, logging predictions, and preparing rollback.
Monitor latency, error rate, throughput, resource usage, prediction distribution, confidence scores, drift, accuracy on labeled feedback, false positives, false negatives, and business impact.
Handle imbalanced data with better metrics, resampling, class weights, threshold tuning, anomaly detection, or collecting more minority-class examples. Accuracy alone should not be trusted.
Choose a model based on problem type, data size, data quality, interpretability, latency, cost, risk, maintenance effort, and evaluation results. Start simple when a simple model meets the requirements.
Common mistakes include starting with a model before defining the business problem, using poor data, leaking data into evaluation, optimizing the wrong metric, skipping monitoring, and ignoring human review or rollback.
Explore 500+ free tutorials across 20+ languages and frameworks.