What is the difference between AI, Machine Learning, and Deep Learning?

Artificial Intelligence is the broad concept of machines performing intelligent tasks. Machine Learning is a subset of AI that includes statistical methods enabling machines to improve tasks with experience. Deep Learning is a subset of machine learning that uses neural networks with three or more layers. Think of it as nested circles - all deep learning is machine learning, but not all machine learning is deep learning. I use traditional ML for structured data with clear features, and deep learning when I have unstructured data like images or text.

Explain the difference between supervised and unsupervised learning

Supervised learning is a machine learning approach where models are trained using labeled data - the model learns to map inputs to correct outputs. Think spam detection - we show the model thousands of emails already marked spam or not spam. Unsupervised learning identifies patterns in unlabeled data and is commonly used for clustering and dimensionality reduction. Customer segmentation is a classic example - the model groups similar customers without us telling it what the groups should be.

What is overfitting and how do you prevent it?

Overfitting occurs when a model learns the training data too well and fails to generalize to new data. This usually results in poor performance on unseen data. The telltale sign is great training metrics but poor validation performance. I prevent it through regularization like L1 or L2, early stopping, dropout for neural networks, and ensuring I have enough training data. Cross-validation helps me detect it early.

What is the bias-variance tradeoff?

Bias refers to errors caused by overly simple models, while variance refers to errors from overly complex models. High bias can lead a model to miss relevant relations between features and target outputs - that is underfitting. High variance can cause the model to fit too closely to the training data, including noise and errors - that is overfitting. The goal is to find a good balance between these two to minimize total error. In practice, I start simple and add complexity only when validation metrics improve.

What is reinforcement learning?

Reinforcement learning is a learning method where an agent learns by interacting with an environment using rewards and penalties. Unlike supervised learning where we provide correct answers, the agent discovers optimal actions through trial and error. It is widely used in robotics, gaming, and autonomous systems. The agent maximizes cumulative reward over time by learning which actions lead to the best outcomes.

What evaluation metrics are commonly used in AI models?

Common metrics include accuracy, precision, recall, F1-score, and ROC-AUC. The choice of metric depends on the problem and business requirements. Accuracy works for balanced datasets but is misleading for imbalanced ones. Precision matters when false positives are costly - like spam detection. Recall matters when false negatives are costly - like disease detection. F1-score balances precision and recall. ROC-AUC measures the model's ability to distinguish between classes across all thresholds.

How would you handle class imbalance in a classification problem?

Class imbalance occurs when some classes have significantly more samples than others. I handle it using resampling or weighted loss functions. For a fraud detection project with 0.1% positive rate, I used a combination of SMOTE for oversampling and adjusted class weights in XGBoost. I also changed my success metric from accuracy to precision-recall AUC since accuracy was meaningless at that imbalance.

Explain how transformers work and why attention matters

A Transformer is a neural network architecture that uses self-attention instead of recurrence. It enables parallel processing and better scalability. The attention mechanism allows models to focus on important parts of the input sequence, improving understanding of context and relationships. The word bank means something different in river bank versus bank account - attention captures this by looking at surrounding words. This parallelization made training much faster and enabled the massive models we have today.

What is a neural network and how does backpropagation work?

A neural network is a computational model inspired by the human brain. It consists of interconnected layers of neurons that learn patterns from data. Backpropagation is an algorithm used to update neural network weights. It minimizes the loss by propagating errors backward through the network. During the forward pass, input flows through the network to produce output. During the backward pass, we compute gradients of the loss with respect to each weight and update weights in the direction that reduces loss.

What is dropout and batch normalization?

Dropout is a regularization technique that randomly disables neurons during training - it helps prevent overfitting by forcing the network to learn redundant representations. Batch normalization normalizes the inputs of each layer to stabilize and speed up training. It also improves model performance by reducing internal covariate shift. I typically use both together - batch norm after each layer and dropout before the final layers.

How do you decide between using a pre-trained model versus training from scratch?

Pre-trained models when you have limited data, similar domain to the pre-training data, or tight timelines. Training from scratch when you have massive domain-specific data, compute budget, and the pre-trained models do not fit your domain well. For most projects, I start with pre-trained and fine-tune. Transfer learning is a machine learning method where a model developed for a task is reused as the starting point for a model on a second task.

What is tokenization in NLP?

Tokenization is the process of breaking text into smaller units called tokens. These tokens are used as inputs for NLP models. There are different approaches - word-level tokenization splits on spaces, character-level uses individual characters, and subword tokenization like BPE or WordPiece splits into meaningful subunits. Modern LLMs use subword tokenization because it handles rare words and different languages well while keeping vocabulary size manageable.

Design a recommendation system for an e-commerce platform

I would start by clarifying requirements - real-time personalization or batch? What scale? Then propose a two-stage system: candidate generation using collaborative filtering to get 1000 candidates quickly, then a ranking model for top 20. Collaborative filtering handles the cold start by falling back to popularity for new users. The ranking model can incorporate more features - browsing history, time of day, seasonality. Serve from a feature store with sub-100ms latency requirement. Batch retrain daily, but update user embeddings in near-real-time from recent interactions.

How would you design an ML pipeline for a production fraud detection system?

Fraud detection needs real-time inference under 50ms typically. I would design a streaming pipeline with Kafka for event ingestion, a feature store for real-time feature computation, and a lightweight model served via a REST API. The model itself might be an ensemble - a fast rules engine for obvious cases plus a gradient boosting model for nuanced decisions. Critical addition is monitoring for concept drift since fraud patterns change constantly. Also need a feedback loop - when fraud analysts label cases, that flows back to training data.

What is MLOps and how do you monitor AI models in production?

MLOps combines machine learning with DevOps practices. It focuses on automating model training, deployment, and monitoring. Models are monitored by tracking accuracy, latency, data drift, and user feedback. Alerts and dashboards are commonly used. Model drift occurs when a model's performance degrades due to changes in data patterns - continuous monitoring is required to detect it. I track prediction distributions daily and trigger retraining when drift exceeds thresholds.

Career Resources · 47 Questions

AI Engineer Interview Questions & Career Resources (2026)

LLM integration, RAG systems, and prompt engineering questions from interviews at Anthropic, OpenAI, and Google. Based on 47 hiring manager conversations and real candidate experiences.

Build your resume View career guide

Resume Score

ATS Optimization

85/ 100

Keywords92%

Formatting88%

Impact76%

“3 callbacks in 5 days. Wild.”

Sarah K. - PM

🎨

Portfolio Examples

Projects that solve real problems stand out. Following a tutorial and uploading it to GitHub does not count. Hiring managers want to see that you can identify a problem, design a solution, implement it, and ideally deploy it where people can actually use it.

Project Ideas by Experience Level

Entry Level

entry

Sentiment Analysis API

Build a REST API that classifies text sentiment using a fine-tuned transformer model. Deploy it somewhere accessible with basic rate limiting.

Why it works: Shows you can deploy, not just train. Most entry candidates stop at the notebook.

PythonFastAPIHuggingFace TransformersDockerAWS or Heroku

entry

Document Q&A System

Build a system that answers questions about uploaded documents using RAG. Simple UI where users upload a PDF and ask questions.

Why it works: RAG is everywhere right now. Shows you understand modern AI architectures beyond basic ML.

LangChainOpenAI APIPinecone or ChromaDBStreamlit

Mid Level

mid

ML Pipeline with Monitoring

End-to-end pipeline with data validation, automated training, deployment, and drift detection. Dashboard showing model health.

Why it works: Demonstrates production mindset. Shows you think about what happens after the model is trained.

Airflow or PrefectMLflowGrafanaGreat ExpectationsAWS or GCP

mid

Real-time Feature Store

Build a feature store that computes features in real-time from streaming data. Serve features with sub-100ms latency.

Why it works: Feature engineering is where most ML work happens. This shows you understand the infrastructure side.

KafkaRedisFeastPython

Senior Level

senior

Multi-model Serving Platform

Platform that serves multiple models with A/B testing, canary deployments, and automatic rollback on degradation.

Why it works: Shows you can think at platform level, not just individual models.

KubernetesSeldon or KServeIstioPrometheus

senior

Custom Training Framework

Framework for distributed training with mixed precision, gradient accumulation, and checkpoint management. Demonstrate on a non-trivial model.

Why it works: Deep understanding of training at scale. Few candidates can do this.

PyTorchDeepSpeed or FSDPWeights & Biases

Real Examples That Landed Jobs

"My fake review detector project led to 3 interview requests in one week. It was not complex - just a fine-tuned BERT model with a simple API. But it solved a real problem and I could demo it live."
- Marcus Chen, AI Engineer at Stripe

"The project that got me hired was a RAG system for legal documents. Not because it was technically impressive but because I built it for my lawyer friend and could talk about real user feedback."
- Sarah Kim, Senior AI Engineer at Notion

Ready to Put This Into Action?

Your resume is the first impression. Make it count with our AI-powered resume builder.

Build your resume Try interview prep