Interview prep
Applied Scientist: 7-day sprint
Seven days on what applied science loops test - derivations, modeling depth, deep learning, transformers, post-training, and experiment design.
Rapid-fire this sprint0/39 done
Reading mode for lessons:
The job. Invent or adapt modeling approaches for a product problem and prove they work: the math, the experiments, and enough engineering to ship.
What the loop usually tests
| Round | What they probe | Where it's covered |
|---|---|---|
| ML breadth | Every classical model's assumptions, objective, and failure modes | Days 2 and 3 |
| ML depth / derivations | Gradients, likelihoods, bias-variance, why an algorithm converges | Days 1 to 4 (read the math sections in full) |
| Deep learning and LLMs | Architectures, optimization, attention, pretraining, post-training | Days 4 to 6 |
| Research and experiment design | Baselines, ablations, metrics, statistical significance | Day 6 |
| Paper or project deep dive | Defend your choices and results | Day 7 |
This sprint uses full lessons, not quick read: the math sections are the interview.
Day 1: The math you'll be asked to use
- Read: Probability and statistics, Calculus for ML, Vectors and matrices, Overfitting and bias-variance
- Also read: Information theory, Numerical stability
- Lab: underfitting vs overfitting
- Cheat sheet: formulas
- Be ready to: derive the bias-variance decomposition, apply Bayes' rule to a base-rate problem, and explain MLE vs MAP.
Day 2: Linear models, derived
- Read: Linear regression, Regularization, Logistic regression, Support vector machines, Naive Bayes
- Labs: linear regression, classification threshold
- Be ready to: derive the normal equation and the logistic-loss gradient, explain L1 sparsity geometrically, and connect L2 to a Gaussian prior.
Day 3: Trees, clustering, and dimensionality
- Read: Decision trees, Random forests and boosting, K-means, Density, hierarchical, and mixture clustering, PCA
- Labs: decision-tree builder, gradient boosting, K-means, PCA
- Be ready to: explain boosting as gradient descent in function space, K-means as coordinate descent (and EM for GMMs), and PCA via eigenvectors of the covariance matrix.
Day 4: Deep learning
- Read: Gradient descent, Backpropagation, Training and regularization, Convolutional networks, Recurrent networks
- Also read: Loss functions, Debugging neural networks, ZeRO sharding
- Labs: gradient descent, backpropagation, convolution
- Be ready to: backpropagate through a small network by hand, explain vanishing gradients and why residual connections and LSTMs help, and compare normalization schemes.
Day 5: Transformers and generative models
- Read: Self-attention, Transformer architecture, Tokenization and pretraining, Autoencoders, diffusion, and transfer learning
- Also read: BERT, Attention variants, Scaling laws, Latent diffusion
- Labs: self-attention, tokenizer
- Be ready to: count a transformer's parameters and FLOPs, explain the attention cost in sequence length, and contrast autoregressive and diffusion generation.
Day 6: Post-training, evaluation, and experiments
- Read: Fine-tuning in practice, Evaluating LLM systems, Experimentation and A/B testing
- Also read: Policy gradients, PPO, Mixture of experts
- Labs: LoRA and fine-tuning memory, A/B test power and peeking
- Be ready to: compare RLHF with PPO against DPO, design an ablation, and compute a sample size for an online test.
Day 7: Mock interview
- Rapid-fire: 20 mixed cards from this sprint, 90 seconds each (open the drill).
- Whiteboard, 15 minutes each: the logistic-regression gradient; the nine backpropagation steps; the attention worked example with shapes.
- Research design, 40 minutes: Search relevance dropped for long-tail queries after a model update. Propose a modeling fix and an experiment that would convince a skeptical reviewer.
- Deep dive: pick a project from the project ladder and defend every choice against "why not X?"