AI/ML Question Bank & Practice Problems

๐ŸŽฏ Career Resources ยท Hope AI โ€” ML & DS Course

๐ŸŽฏ TL;DR โ€” 100 real AI/ML interview questions organized by topic, with concise model answers, difficulty ratings, and which companies ask each type. Use this as your daily drill bank โ€” 10 questions per day for 10 days = full coverage.

๐Ÿ“‹ Question Bank Overview

Topic# QuestionsDifficultyWho Asks
ML Fundamentals25Easyโ€“MediumAll companies
Deep Learning20MediumGoogle, Meta, Anthropic
NLP / LLMs20Mediumโ€“HardAnthropic, OpenAI, any LLM team
AI Agents & RAG15Mediumโ€“HardAI startups, Anthropic, LangChain ecosystem
Python & SQL20Easyโ€“MediumAll companies

๐Ÿ”ข How to Use This Bank

  • Daily drill: Pick one topic, answer 5 questions out loud without looking at answers.
  • Spaced repetition: Mark questions you got wrong. Review wrong answers 3 days later.
  • Depth practice: For starred (โญ) questions, prepare a 2-minute verbal answer.
  • Coding practice: For any question with a "Code it" tag, implement it in Python.
  • Track progress: Check off questions as you master them.

๐Ÿ’ก ML Fundamentals โ€” 25 Questions

Easy (must nail these in seconds)

Q1 โญ What are the three types of machine learning?

A: Supervised (labeled data, predict output), Unsupervised (no labels, find structure), Reinforcement (agent learns from rewards/penalties). Semi-supervised and self-supervised are increasingly common in AI.

Q2 What is overfitting? How do you detect and fix it?

A: Model memorizes training data, fails on new data. Detect: large gap between train and validation accuracy. Fix: regularization (L1/L2), dropout, more data, early stopping, simpler model.

Q3 โญ Explain bias-variance tradeoff.

A: High bias = underfitting (model too simple to capture patterns). High variance = overfitting (model captures noise). Goal: find the sweet spot. Ensemble methods (random forest, boosting) help balance both.

Q4 What is the difference between a parameter and a hyperparameter?

A: Parameters are learned from data (weights, biases). Hyperparameters are set before training (learning rate, number of layers, regularization strength). Hyperparameters are tuned with grid search, random search, or Bayesian optimization.

Q5 What is cross-validation and why do you use it?

A: Split data into K folds. Train on K-1 folds, test on the remaining fold. Repeat K times, average results. More reliable than a single train/test split, especially on small datasets.

Medium (practice verbal explanation)

Q6 โญ Explain gradient descent. What are its variants?

A: Iteratively updates parameters by moving in the direction that reduces loss. Variants: Batch GD (full dataset per step โ€” stable but slow), Stochastic GD (one sample per step โ€” noisy but fast), Mini-batch GD (standard โ€” best of both). Adam = adaptive learning rates, most commonly used optimizer.

Q7 L1 vs L2 regularization โ€” when do you use each?

A: L1 (Lasso): adds absolute value of weights to loss. Produces sparse models (many weights go to 0) โ€” good for feature selection. L2 (Ridge): adds squared weights. Shrinks all weights but keeps all features. Use L1 when you suspect many irrelevant features.

Q8 โญ How does Random Forest work? Why does it reduce variance?

A: Trains many decision trees, each on a random subset of data (bootstrap sampling) and random subset of features. Averages their predictions. Variance cancels because individual trees are correlated โ€” reducing correlation is key, which is what random feature selection achieves.

Q9 How does XGBoost differ from Random Forest?

A: XGBoost is sequential (each tree corrects errors of previous โ€” boosting). Random Forest is parallel (trees train independently โ€” bagging). XGBoost typically outperforms RF on structured data; RF is more robust to hyperparameters.

Q10 โญ What is data leakage? Give an example.

A: When future or test data information leaks into the training process. Example: scaling features using the entire dataset before split โ€” the scaler "knows" test set statistics. Fix: always fit preprocessing (scalers, encoders) ONLY on training data, then transform test data.

Q11 How do you handle class imbalance?

A: (1) Resample: oversample minority (SMOTE) or undersample majority. (2) Class weights: penalize wrong predictions on minority class more. (3) Threshold tuning: adjust decision threshold. (4) Use right metric: precision-recall instead of accuracy. (5) Try anomaly detection framing.

Q12 What metrics would you use to evaluate a fraud detection model?

A: Precision-Recall AUC (skewed classes make accuracy useless). Focus on Recall (don't miss fraud) and Precision (don't annoy legit users). Also F1 at chosen threshold. Business metric: expected cost = false negative cost ร— FN rate + false positive cost ร— FP rate.

Q13 What is feature importance in tree-based models?

A: Mean decrease in impurity (Gini/entropy) per feature across all trees (MDI). Or permutation importance: shuffle feature values, measure accuracy drop. MDI can be biased toward high-cardinality features; permutation importance is more reliable.

Q14 When would you use a decision tree over logistic regression?

A: Decision tree: when you need interpretability for non-linear decisions, feature interactions are important, mixed data types, no feature scaling needed. Logistic regression: when you need calibrated probabilities, interpretable coefficients, fast inference, limited data.

Q15 What is the curse of dimensionality?

A: As dimensions increase, data points become sparse โ€” distance metrics lose meaning, models need exponentially more data. Fix: dimensionality reduction (PCA, UMAP), feature selection, regularization.

Medium-Hard

Q16 Explain the Naive Bayes "naive" assumption and when it still works.

A: Assumes all features are conditionally independent given the class โ€” almost never true in reality. Still works well for: text classification (word counts), spam filtering, when dataset is small. Feature correlation doesn't affect the ranking of class probabilities as much as you'd expect.

Q17 โญ What is the difference between parametric and non-parametric models?

A: Parametric: fixed number of parameters regardless of data size (linear regression, logistic regression, neural nets). Non-parametric: parameters grow with data (k-NN, decision trees, SVMs with kernel). Parametric: faster, more interpretable. Non-parametric: more flexible.

Q18 How do SVMs find the optimal hyperplane?

A: Maximizes the margin (distance between hyperplane and nearest data points โ€” "support vectors"). Uses convex optimization (QP problem). The kernel trick maps data to higher dimensions implicitly, enabling non-linear boundaries without explicit transformation.

Q19 What is the EM algorithm? Where is it used?

A: Expectation-Maximization. E-step: estimate hidden variables given current parameters. M-step: update parameters to maximize likelihood. Repeat. Used in: Gaussian Mixture Models, topic modeling (LDA), handling missing data.

Q20 โญ How do you select the right evaluation metric for a regression problem?

A: MSE: penalizes large errors heavily (sensitive to outliers). MAE: more robust to outliers. RMSE: same scale as target variable, easier to interpret. Rยฒ: how much variance is explained (useful for comparison). MAPE: percentage error (good when target varies widely).

Q21โ€“Q25 (Quick-fire โ€” know these cold)

  • Q21: What is the difference between bagging and boosting? โ†’ Bagging: parallel, reduces variance. Boosting: sequential, reduces bias.
  • Q22: What is k-fold stratified cross-validation? โ†’ Each fold preserves class distribution โ€” essential for imbalanced datasets.
  • Q23: What is the elbow method in K-Means? โ†’ Plot inertia vs. K, pick K where curve "elbows" โ€” diminishing returns on adding clusters.
  • Q24: What is PCA? โ†’ Linear dimensionality reduction. Projects data onto directions of maximum variance (principal components). Requires feature scaling first.
  • Q25: What is the difference between discriminative and generative models? โ†’ Discriminative: models P(y|x) directly (logistic regression, SVM). Generative: models P(x|y) and P(y) (Naive Bayes, GMM). Generative can generate new samples.

๐Ÿ’ก Deep Learning โ€” 20 Questions

Q26 โญ What is backpropagation?

A: Algorithm to compute gradients of loss with respect to all parameters using the chain rule. Forward pass: compute loss. Backward pass: compute โˆ‚loss/โˆ‚w for each weight by chaining partial derivatives from output layer back to input. Parameters updated by optimizer using these gradients.

Q27 What are activation functions? Why do we need non-linearity?

A: Non-linear functions applied after each layer. Without them, stacking layers collapses to a single linear transformation. Common: ReLU (default, fast, avoids vanishing gradient), Sigmoid (output layer for binary classification), Softmax (multi-class probabilities), GELU (used in transformers).

Q28 โญ What is the vanishing gradient problem?

A: During backpropagation through deep networks, gradients are multiplied repeatedly. If they're < 1 (like sigmoid derivatives), they exponentially shrink โ†’ early layers get near-zero gradient โ†’ don't learn. Fix: ReLU activations, residual connections (ResNets), batch normalization, LSTM/GRU for sequences.

Q29 What is batch normalization? What does it actually do?

A: Normalizes layer inputs to zero mean, unit variance within each mini-batch, then applies learned scale (ฮณ) and shift (ฮฒ). Reduces internal covariate shift โ€” layers don't have to adjust to shifting input distributions. Speeds training by ~10x, allows higher learning rates, slight regularization effect.

Q30 โญ What is dropout? How does it work at inference?

A: Randomly sets neurons to 0 during training with probability p (typically 0.2โ€“0.5). Forces network to learn redundant representations โ€” reduces co-adaptation. At inference: ALL neurons are active, but their activations are scaled by (1-p) to match expected training behavior (or use inverted dropout during training).

Q31 What is transfer learning? Name 3 strategies.

A: Use pretrained model weights (trained on large dataset) for new task. Strategies: (1) Feature extraction: freeze all layers, train only new head. (2) Fine-tuning: unfreeze top layers, train with low LR. (3) Full fine-tuning: update all weights (only if you have lots of data).

Q32 Explain CNNs โ€” what do different layers learn?

A: Convolutional layers detect local patterns with shared weights. Early layers: edges, textures. Mid layers: shapes, patterns. Deep layers: semantic concepts (faces, objects). Pooling layers downsample, providing spatial invariance. Dramatically fewer parameters than fully connected via weight sharing.

Q33 โญ What is the transformer architecture?

A: Multi-head self-attention + feed-forward layers, repeated N times. Self-attention: each token attends to all other tokens (attention = softmax(QKแต€/โˆšd_k) ร— V). Positional encoding added since attention is permutation-invariant. Encoder-only (BERT), decoder-only (GPT), or encoder-decoder (T5).

Q34 What is the difference between RNNs, LSTMs, and Transformers?

A: RNN: sequential processing, vanishing gradient on long sequences. LSTM: gates (forget, input, output) control information flow โ€” handles longer dependencies. Transformer: full self-attention, processes all tokens in parallel, scales better, handles long-range dependencies โ€” now dominant.

Q35 What is weight initialization? Why does it matter?

A: Poor initialization โ†’ vanishing or exploding gradients before training even starts. Xavier/Glorot: scales by โˆš(1/fan_in) โ€” good for tanh. He initialization: scales by โˆš(2/fan_in) โ€” designed for ReLU. Batch norm reduces sensitivity to initialization.

Q36โ€“Q45 (Quick-fire)

  • Q36: What is learning rate scheduling? โ†’ Reduce LR over training (step decay, cosine annealing, warm-up + decay). Prevents overshooting minima.
  • Q37: Adam optimizer โ€” how does it work? โ†’ Adaptive learning rates per parameter. Maintains running averages of gradient (momentum) and squared gradient (RMSprop). Usually best default choice.
  • Q38: What is an autoencoder? โ†’ Encoder compresses input to latent space; decoder reconstructs. Used for: dimensionality reduction, anomaly detection, denoising.
  • Q39: What is a GAN? โ†’ Generator creates fake samples; discriminator distinguishes real from fake. Both trained adversarially. Used for: image generation, data augmentation.
  • Q40: What is the difference between model capacity and model complexity? โ†’ Capacity = ability to learn diverse functions (more params = more capacity). Complexity = how complex the actual learned function is. High capacity + regularization โ‰  high complexity.
  • Q41: What is early stopping? โ†’ Stop training when validation loss stops improving. Use patience parameter (e.g., stop if no improvement for 5 epochs).
  • Q42: What is knowledge distillation? โ†’ Train small "student" model to mimic large "teacher" model's output probabilities. Gets close to teacher performance with much smaller model.
  • Q43: What is the difference between generalization and memorization? โ†’ Generalization: model learns underlying patterns. Memorization: model learns noise/specific examples. Test set performance measures generalization.
  • Q44: What is residual connection (skip connection)? โ†’ Add input of a layer directly to its output: output = F(x) + x. Allows gradients to flow directly, enables training of very deep networks (ResNet).
  • Q45: What is mixed precision training? โ†’ Use 16-bit floats for forward/backward pass, 32-bit for weight updates. 2โ€“3x faster on modern GPUs, reduces memory by ~50%.

๐Ÿ’ก NLP / LLMs โ€” 20 Questions

Q46 โญ What is self-attention? Walk me through the computation.

A: Each token creates Query (Q), Key (K), Value (V) vectors via learned weight matrices. Attention scores = Q ร— Kแต€ / โˆšd_k. Apply softmax โ†’ attention weights. Output = weighted sum of V vectors. Multi-head: run attention H times in parallel with different weight matrices, concatenate.

Q47 What is tokenization? What is BPE?

A: Tokenization splits text into subword units. BPE (Byte Pair Encoding): starts with characters, iteratively merges the most frequent adjacent pair. "playing" โ†’ ["play", "ing"]. Handles rare/novel words gracefully. WordPiece (BERT) is similar but merges pairs that maximize language model likelihood.

Q48 โญ What is RLHF and why is it used?

A: Reinforcement Learning from Human Feedback. Step 1: supervised fine-tuning on demonstrations. Step 2: humans rank model outputs โ†’ train reward model. Step 3: use PPO (RL algorithm) to fine-tune LLM to maximize reward. Makes models helpful, harmless, and honest โ€” standard training for Claude, GPT-4.

Q49 What are hallucinations in LLMs? How do you mitigate them?

A: Model generates plausible but factually wrong content. Mitigations: RAG (ground output in retrieved documents), chain-of-thought prompting (reason step by step), temperature=0 for factual tasks, self-consistency (sample multiple times, pick majority), calibration training (RLHF with honesty reward).

Q50 โญ RAG vs fine-tuning โ€” when to choose each?

A: RAG: when knowledge is dynamic (changes frequently), knowledge is too large for context window, you need citations/sources, latency is acceptable. Fine-tuning: when you need a specific response style/format, task is very specialized, need fastest possible inference, knowledge is stable.

Q51โ€“Q65 (Quick-fire โ€” know these cold)

  • Q51: What is prompt engineering? โ†’ Crafting inputs to elicit better LLM outputs. Techniques: zero-shot, few-shot, chain-of-thought, role prompting, structured output (JSON mode).
  • Q52: What is chain-of-thought prompting? โ†’ Ask model to "think step by step" before answering. Dramatically improves multi-step reasoning. Emerges in models >100B parameters.
  • Q53: What is temperature in LLMs? โ†’ Controls randomness. T=0: deterministic (always highest probability token). T=1: standard sampling. T>1: more random/creative. Use T=0 for factual tasks, T=0.7โ€“1.0 for creative.
  • Q54: What is the context window? โ†’ Maximum tokens (input + output) the model can process in one call. GPT-4: 128K. Claude 3.5: 200K. Longer context = more expensive + slower.
  • Q55: What is LoRA? โ†’ Adds low-rank trainable matrices to frozen pretrained weights. Trains only ~0.1% of parameters. Standard method for efficient LLM fine-tuning. QLoRA: 4-bit quantized LoRA โ€” fine-tune 70B models on a single GPU.
  • Q56: What is the difference between BERT and GPT architectures? โ†’ BERT: encoder-only, bidirectional attention, good for understanding tasks (classification, NER). GPT: decoder-only, causal attention, good for generation tasks.
  • Q57: What is semantic search vs keyword search? โ†’ Keyword: exact string matching (BM25, TF-IDF). Semantic: embeds text to vectors, searches by meaning similarity (cosine distance). Semantic better for paraphrase/synonyms; hybrid search often best.
  • Q58: What is a vector database? โ†’ Database optimized for storing and querying high-dimensional embeddings using approximate nearest neighbor (ANN) search. Examples: Pinecone, ChromaDB, Weaviate, pgvector.
  • Q59: What is embedding? โ†’ Dense vector representation of text (or image) capturing semantic meaning. Similar meanings โ†’ similar vectors (close in vector space). GPT models use 1536-dim embeddings; BERT uses 768-dim.
  • Q60: What is instruction tuning? โ†’ Fine-tune a base LLM on (instruction, response) pairs to follow user instructions. Without it, LLMs just complete text; with it, they answer questions and follow directions.
  • Q61: What is PEFT? โ†’ Parameter-Efficient Fine-Tuning. Fine-tunes a small subset of parameters instead of all. Includes LoRA, prefix tuning, prompt tuning, adapter methods.
  • Q62: What is the difference between GPT-3.5 and GPT-4? โ†’ GPT-4 is multimodal, better at reasoning (especially math/coding), fewer hallucinations, larger context window. GPT-3.5 is faster and cheaper.
  • Q63: What is token prediction? How do LLMs generate text? โ†’ LLMs predict the next token probability distribution over vocabulary. Sample or argmax to select next token. Repeat until EOS token or max length.
  • Q64: What is Constitutional AI (Anthropic)? โ†’ Training approach where model critiques and revises own outputs according to a set of principles, reducing need for human feedback at scale.
  • Q65: What is the difference between LLM completion and chat APIs? โ†’ Completion: single text in, text out. Chat: structured conversation with system/user/assistant roles. Chat API enables multi-turn conversations with maintained context.

๐Ÿ’ก AI Agents & RAG โ€” 15 Questions

Q66 โญ What is an AI agent?

A: An LLM that can take actions โ€” call tools (web search, code execution, API calls), observe results, and decide next action โ€” autonomously until a goal is achieved. Key components: LLM (brain), tools (hands), memory (context), planning loop (ReAct, CoT).

Q67 โญ What is the ReAct framework?

A: Thought โ†’ Action โ†’ Observation loop. LLM: (1) Thinks about what to do next, (2) Calls a tool, (3) Receives observation, (4) Updates its thinking. Enables multi-step reasoning with real-world grounding via tools.

Q68 What is LangGraph and when would you use it over LangChain?

A: LangGraph builds agents as directed graphs (nodes = LLM/tool calls, edges = routing logic). Use when: you need branching logic, multi-agent systems, cycles/loops, persistent state across turns. LangChain chains are linear; LangGraph handles complex agent flows.

Q69 โญ What is chunking strategy in RAG? What chunk size do you use?

A: Splitting documents for embedding. Fixed-size (most common): 256โ€“1024 tokens with ~10โ€“20% overlap. Semantic chunking: split at sentence/paragraph boundaries. Larger chunks: more context, but lower retrieval precision. Smaller chunks: more precise retrieval, but might miss context. 512 tokens + 10% overlap is a solid default.

Q70 How do you evaluate a RAG system?

A: Use the RAGAs framework: (1) Context Precision: were retrieved chunks actually relevant? (2) Context Recall: were all relevant chunks retrieved? (3) Answer Faithfulness: does the answer stick to the retrieved context? (4) Answer Relevance: does the answer address the question? Track all 4 metrics in a regression test suite.

Q71 What is hybrid search in RAG?

A: Combine sparse retrieval (BM25 โ€” keyword matching) with dense retrieval (vector similarity). Weighted combination of both scores (RRF โ€” Reciprocal Rank Fusion is common). Outperforms either alone because: BM25 is great for exact terms/names, vector search handles semantic similarity.

Q72 What is an agent tool? Give 3 examples.

A: Functions the agent can call to interact with the world. Examples: web_search (get current information), code_interpreter (run Python), read_file (access documents), send_email (take action), sql_query (query database). Tools are defined with a name, description, and parameter schema (JSON Schema).

Q73 What is memory in AI agents?

A: (1) In-context memory: conversation history in the context window โ€” simple but limited by context length. (2) External memory: store summaries or full history in vector DB, retrieve relevant memories. (3) Entity memory: extract and store key facts (user preferences, names, facts). (4) Episodic memory: store past interaction summaries.

Q74 What is a multi-agent system?

A: Multiple specialized agents collaborating โ€” each handles a subtask. Example: orchestrator agent routes queries โ†’ researcher agent searches web โ†’ writer agent drafts response โ†’ editor agent reviews. Reduces complexity per agent, enables parallelism, improves specialization.

Q75 What are the failure modes of AI agents?

A: (1) Tool call errors (agent keeps retrying in a loop). (2) Hallucinated tool arguments (passes wrong parameters). (3) Context window overflow (too many tool results fill the context). (4) Infinite loops (no clear termination condition). Mitigations: max iterations, structured output for tool calls, summarize tool results.

Q76โ€“Q80 (Quick-fire)

  • Q76: What is function calling in LLMs? โ†’ Structured way for LLMs to output tool calls. You define available functions as JSON schemas; model outputs JSON instead of natural language to call them.
  • Q77: What is a system prompt? โ†’ Instructions given to the LLM before user conversation. Sets role, behavior, output format, tools available. Persistent across turns in chat.
  • Q78: What is query expansion in RAG? โ†’ Rephrase user query into multiple variants before retrieval to improve recall. Use LLM to generate synonyms or sub-questions.
  • Q79: What is a guardrail in AI agents? โ†’ Input/output validation that blocks harmful, off-topic, or malformed content. Examples: Anthropic's Constitutional AI, Nvidia NeMo Guardrails, custom regex/classifier checks.
  • Q80: What is the difference between a workflow and an agent? โ†’ Workflow: predefined, deterministic sequence of steps. Agent: LLM decides the next step dynamically. Workflows are predictable; agents are flexible but less reliable.

๐Ÿ’ก Python & SQL โ€” 20 Questions

Q81 What is the GIL in Python?

A: Global Interpreter Lock โ€” CPython only allows one thread to execute Python bytecode at a time. Limits CPU parallelism for CPU-bound tasks. Fix: use multiprocessing (separate processes), or async/await for I/O-bound tasks. NumPy operations release the GIL.

Q82 โญ What is a decorator? Write one.

A: Function that wraps another function to add behavior. Classic example:
import time
def timer(func):
    def wrapper(*args, **kwargs):
        start = time.time()
        result = func(*args, **kwargs)
        print(f"{func.__name__} took {time.time()-start:.2f}s")
        return result
    return wrapper

@timer
def my_function(): pass

Q83 What is the difference between a list and a generator in Python?

A: List: stores all items in memory at once. Generator: produces items one at a time on demand (lazy evaluation). Generators use far less memory for large sequences. Use yield instead of return.

Q84 โญ Explain list comprehension and dictionary comprehension.

A: squares = [x2 for x in range(10) if x % 2 == 0] โ†’ list. sq_dict = {x: x2 for x in range(10)} โ†’ dict. Faster than equivalent for loops due to Python internals.

Q85 What are *args and **kwargs?

A: *args: captures variable positional arguments as tuple. **kwargs: captures variable keyword arguments as dict. Allows flexible function signatures. Commonly used in wrappers and decorators.

Q86 What is the difference between is and ==?

A: == checks value equality. is checks identity (same object in memory). a = [1,2]; b = [1,2]; a == b is True, but a is b is False. Exception: small integers and strings are cached โ€” 1 is 1 is True.

Q87 โญ Write SQL to find the 2nd highest salary.

A: SELECT MAX(salary) FROM employees WHERE salary < (SELECT MAX(salary) FROM employees); Or: SELECT salary FROM employees ORDER BY salary DESC LIMIT 1 OFFSET 1;

Q88 What is a window function in SQL?

A: Performs calculation across related rows without collapsing them into groups. Example: ROW_NUMBER() OVER (PARTITION BY dept ORDER BY salary DESC) โ€” ranks employees within department without losing individual rows.

Q89 โญ Write SQL to find duplicate records.

A: SELECT email, COUNT() as cnt FROM users GROUP BY email HAVING COUNT() > 1;

Q90 What is the difference between JOIN types?

A: INNER JOIN: only matching rows. LEFT JOIN: all left rows + matching right (null if no match). RIGHT JOIN: opposite. FULL OUTER JOIN: all rows from both. CROSS JOIN: cartesian product. For AI: LEFT JOIN is most common โ€” keep all records, flag missing.

Q91โ€“Q100 (Quick-fire Python/SQL)

  • Q91: What is vectorization in NumPy? โ†’ Replace Python loops with array operations โ€” runs in C, 100x faster. np.sum(arr) vs sum(arr).
  • Q92: What is the difference between .loc and .iloc in pandas? โ†’ .loc: label-based indexing (row name/column name). .iloc: integer position-based indexing.
  • Q93: How do you handle missing values in pandas? โ†’ df.isnull().sum() to find. Fill: df.fillna(value). Drop: df.dropna(). Impute with median/mean for ML pipelines.
  • Q94: What is multiprocessing vs multithreading in Python? โ†’ Multithreading: same process, shared memory, limited by GIL. Multiprocessing: separate processes, separate memory, true parallelism. Use multiprocessing for CPU-bound ML tasks.
  • Q95: What is async/await in Python? โ†’ Asynchronous programming for I/O-bound tasks. async def defines coroutine; await pauses execution until I/O completes without blocking. Essential for high-throughput LLM API calls.
  • Q96: Write a Python function to calculate cosine similarity. โ†’ def cosine(a, b): return np.dot(a,b) / (np.linalg.norm(a) * np.linalg.norm(b))
  • Q97: What is the difference between deepcopy and copy? โ†’ copy: shallow copy โ€” copies object but references nested objects. deepcopy: recursively copies everything. Critical when working with nested data structures in ML pipelines.
  • Q98: How do you write efficient SQL for large tables? โ†’ Use indexes, avoid SELECT *, filter early with WHERE before JOIN, use EXPLAIN to check query plan, avoid functions on indexed columns in WHERE clause.
  • Q99: What is CTEs (Common Table Expressions) in SQL? โ†’ WITH cte AS (SELECT ...) SELECT * FROM cte. Named subquery that improves readability. Can be recursive.
  • Q100: What is the difference between GROUP BY and PARTITION BY? โ†’ GROUP BY collapses rows to one per group. PARTITION BY (in window functions) keeps all rows but adds a grouping context for calculations.

โš ๏ธ Common Mistakes to Avoid

  • Memorizing answers without understanding โ€” interviewers always ask "why" or "what if" follow-ups
  • Skipping easy questions โ€” hiring decisions are often made on how you handle basics, not just hard questions
  • Not knowing your own resume โ€” any project or skill you listed is fair game for deep-dive
  • Python-only focus โ€” SQL is tested at almost every company; don't neglect it
  • No practical examples โ€” every theory answer should end with "in practice, this is used for..."

๐Ÿš€ Quick Reference: Key Formulas

FormulaExpression
MSE(1/n) ร— ฮฃ(y_pred - y_true)ยฒ
PrecisionTP / (TP + FP)
RecallTP / (TP + FN)
F1 Score2 ร— (Precision ร— Recall) / (Precision + Recall)
Cosine Similarity(AยทB) / (โ€–Aโ€– ร— โ€–Bโ€–)
Attention Scoresoftmax(QKแต€ / โˆšd_k) ร— V
Softmaxe^x_i / ฮฃe^x_j
Cross-Entropy Loss-ฮฃ y_true ร— log(y_pred)

๐Ÿ“‹ Action Checklist

  • Answer all 10 ML Fundamentals questions (Q1โ€“Q10) out loud without notes
  • Work through all 5 deep learning starred questions (Q26, Q28, Q30, Q33, Q34)
  • Answer the 5 LLM starred questions (Q46, Q48, Q49, Q50, Q55)
  • Write the cosine similarity and softmax functions from memory
  • Practice the 3 SQL questions (Q87, Q88, Q89) in a live SQL environment
  • Write a Python decorator from scratch (Q82)
  • Review all quick-fire answers and mark which ones need more study
  • Do a 30-minute timed drill: pick 10 random questions, answer out loud
โ† ๐Ÿ’ป Technical Interview Preparation๐Ÿ“š Technical Concepts โ€” Master Reference โ†’