Interview Prep

Attention internals, transformer training dynamics, and the RLHF stack — from softmax to DPO, with safety-flavoured objectives at the end.

LLM training and decoding fundamentals, scaling-aware parameter counting, and alignment losses used to fine-tune and evaluate frontier models.

Math-heavy screens in the Google style: matrix decompositions, eigen-computations, convex optimization steps, and probability foundations.

Reinforcement learning theory with the probability and optimization toolkit behind policy evaluation, value iteration, and trust-region methods.

Ranking and recommendation metrics, collaborative filtering, and computer-vision evaluation — the measured half of Meta's ML loops.

Kernel-level efficiency: quantization scales, attention tiling, occupancy, and the memory math that decides whether a model fits.

Classical probability and statistical inference with stochastic processes: distributions, Bayes, Markov chains, and estimator quality.

Market-math puzzles: combinatorics, paradoxes, betting games, and optimization under uncertainty, closing with option-pricing steps.

Hypothesis testing, experiment sizing, and quality metrics for the data-science loop — the statistics Amazon asks before the ML questions.

A broad data-science path: descriptive statistics, probability, regression and classification metrics, then inference and resampling.

The full applied pipeline: splitting and scaling data, cross-validation, optimizer updates, and the calibration and importance checks before shipping.

Classic supervised learning and deep learning: losses and gradients, k-means and trees, normalization layers, and batch optimizers.

Serving a model in production: quantization and KV-cache accounting, batching and latency budgets, then fused kernels and speculative decoding.