Arrays, dtypes, views vs copies, axis operations, and broadcasting — the two foundational NumPy topics. Tests both conceptual clarity and the edge cases that trip up experienced engineers.

Series, DataFrame indexing, dtypes, groupby, merge, and the traps that cause production bugs. Covers the distinction between views and copies and every variant of SettingWithCopyWarning.

Feature engineering, missing values, categorical encoding, time-series splits — then visualization: figure anatomy, confusion matrix heatmaps, learning curves, and publication-quality formatting.

Statistical testing, sparse matrices, distance calculations, then the full sklearn Estimator API — Pipeline, ColumnTransformer, cross-validation, and hyperparameter search. Essential for every ML interview.

Tensors, autograd, computation graphs, then the full training loop: zero_grad, backward, mixed precision, model.train/eval modes. The two topics every ML engineer must master before any PyTorch interview.

Keras model lifecycle, tf.function tracing, BatchNorm training modes, then PyTorch DataLoader internals and tf.data pipelines — prefetch, shuffle buffers, DistributedSampler, and bottleneck detection.

Your first ML libraries interview simulation. Covers fundamentals across NumPy, Pandas, sklearn, PyTorch, and TensorFlow. Tests the concepts interviewers ask in every round — common traps included.

Second easy-difficulty interview simulation. A fresh set of fundamental questions across all 12 ML library topics. Complete both Easy mocks before moving to Medium for maximum coverage.

First medium-difficulty mock interview. Covers applied reasoning — broadcasting bugs, leakage in CV pipelines, groupby transform vs agg, gradient accumulation, and tf.data prefetch strategy. FAANG-style questions.

Second medium-difficulty mock. A different angle on applied reasoning: view flags, left vs inner merge semantics, apply bottleneck, pairplot scale, Welch vs Student t-test, DataLoader Windows crash, and Batch Norm training mode.

First hard-difficulty mock. Covers int32 overflow, fancy indexing diagonal trap, temporal leakage with KFold, BFGS finite differences, custom sklearn transformer protocol, double backward error, full reproducibility checklist, GradientTape disconnected graph, and DistributedSampler epoch seeding.

Second hard-difficulty mock. Covers reshape copy after transpose, covariance centering error, broadcast_to read-only trap, non-monotonic index slicing, concat O(N²) anti-pattern, ROC below diagonal diagnosis, nested CV selection bias, DataParallel .module access, NaN loss debugging, @tf.function retrace, scaler leakage in cross-validation, and KS test for production drift.

Elite assessment for senior engineers. Tests deep internals: NumPy memory layout and overflow, broadcast_to read-only semantics, fancy indexing paired vs grid selection, covariance centering, pandas Copy-on-Write, concat O(N²), sparse matrix memory math, custom sklearn clone protocol, nested CV bias, PyTorch double backward, full reproducibility sources, GradientTape disconnection, @tf.function retrace cost, and einsum contraction. Expect multi-step reasoning on every question.

Second elite assessment. A production-failure focused gauntlet: non-monotonic index slicing, ROC curve inversion diagnosis, twinx legend bug, Pipeline double-underscore naming, DataParallel .module access, NaN loss systematic diagnosis, DistributedSampler epoch seeding, padded_batch for variable-length sequences, scaler leakage in cross-validation, seaborn mask upper triangle, KS test for production drift, FeatureUnion vs ColumnTransformer, model.half() dtype mismatch, Keras save/load inference drift, TFRecord cache strategy, Categorical dtype add_categories trap, df.query @variable scope failure, and DataLoader __len__ contract.