Comprehensive AI Architecture Matrices & Interview Mastery

This master reference synthesizes the core mathematical derivations, algorithmic implementations, and architectural decisions across the entire AI & Machine Learning Engineering Curriculum.

Whether designing enterprise AI platforms, optimizing low-latency LLM serving runtimes, or defending design decisions in technical interviews, this guide provides the definitive comparative matrices and mathematical identity cheat sheets.

Pedagogical StandardCourse Capstone

100% Zero-Dependency

From linear classifiers to PagedAttention and PPO, every algorithm in this curriculum was derived algebraically and implemented from first principles.

1. End-to-End Algorithmic Decision Matrices ("When to Use What")#

1. Classical Machine Learning Selection Guide#

Problem Domain	Data Characteristics	Recommended Model	Why Selection Wins	Key Trade-offs & Failure Modes
Tabular Regression / Classification	Medium size ( $N < 100\text{k}$ ), heterogeneous numerical & categorical features	XGBoost / LightGBM / CatBoost	Handles missing values, non-linear feature interactions, and feature scaling automatically; state-of-the-art accuracy on tabular data	Weak on unstructured raw pixel/text data; sensitive to extreme hyperparameter misconfiguration
High-Dimensional Linear Data	Large feature count ( $d > N$ ), sparse text term counts	Logistic Regression / ElasticNet	Fast training/inference; linear decision boundaries; regularized weights ( $L_1 / L_2$ ) prevent overfitting	Cannot capture non-linear feature interactions without explicit polynomial feature engineering
Ultra-Fast Baseline / Text Filtering	Discrete word counts, spam classification	Multinomial Naive Bayes	$O(N \cdot d)$ single-pass training; requires minimal sample size; fast prediction speeds	Conditional independence assumption fails if features are strongly correlated
Auditable Interpretable Decision Rules	Small dataset, strict compliance/regulatory auditing	CART Decision Tree (Pruned)	Highly interpretable tree splits; clear IF-THEN rule structures	High variance; unstable split trees prone to overfitting without depth constraints
Dense Metric Retrieval	Low-dimensional continuous vectors ( $d < 20$ )	k-NN with KD-Tree / Ball Tree	Non-parametric lazy learning; adapts instantly to evolving distributions	Suffer from Curse of Dimensionality ( $d > 20$ degrades to brute-force $O(N \cdot d)$ )
Small Sample Size, Complex Boundary	Complex continuous features ( $N < 10\text{k}$ )	Support Vector Machine (RBF Kernel)	Maximizes geometric margin; robust against overfitting via kernel trick	Quadratic programming complexity $O(N^2 \cdot d)$ to $O(N^3)$ makes it unscalable for large $N$

2. Deep Learning & Perception Architecture Selection Guide#

Task Domain	Input Modality	Recommended Architecture	Core Mechanism	Primary Alternative & Trade-off
Image Classification / Feature Extraction	High-resolution 2D Images	ResNet / ConvNeXt	Spatial convolutions with residual skip connections ( $\mathcal{F}(x) + x$ ) eliminating vanishing gradients	Vision Transformer (ViT): Requires massive pre-training datasets due to lack of spatial inductive bias
Sequence Encoding & Bi-directional Context	Text, code tokens, molecular strings	BERT (Encoder-Only)	Masked Language Modeling (MLM) with bidirectional self-attention	Decoder-Only (GPT): Sub-optimal for sentence similarity due to causal triangular attention masking
Auto-Regressive Generation & Dialogue	Free-form natural language text	Decoder-Only Transformer (LLaMA, GPT)	Causal masked self-attention predicting next-token probabilities auto-regressively	Encoder-Decoder (T5): Effective for translation/summarization but higher serving complexity
Non-Euclidean Relational Data	Knowledge graphs, citation networks, molecules	GraphSAGE / GAT	Inductive neighborhood sampling / Dynamic self-attention weights	GCN: Transductive requirement limits execution to fixed static full adjacency matrices
Multimodal Vision-Language Alignment	Paired images and text captions	CLIP (Dual-Encoder)	Contrastive InfoNCE alignment projecting vision and text to shared unit hypersphere	LLaVA (Adapter): Greater conversational depth but cannot pre-compute fast vector indices

3. Generative AI, Serving & Production MLOps Matrix#

Operational Goal	Constraints	Optimal Technical Stack	Key Architectural Mechanism
Custom Enterprise Knowledge Retrieval	Constantly evolving corporate docs, zero retraining budget	Hybrid RAG (Dense + BM25 + Cross-Encoder)	HNSW vector index + BM25 lexical search combined via Reciprocal Rank Fusion (RRF)
Domain Adaptation & Style Alignment	Specialized corpus, limited GPU memory budget	QLoRA (4-bit NormalFloat)	Low-rank adapter matrices ( $\Delta W = B \cdot A$ ) attached to frozen quantized base weights
Human Preference Alignment	Paired feedback dataset $(y_w \succ y_l)$	Direct Preference Optimization (DPO)	Closed-form implicit reward loss derived directly from policy ratio without training reward models
High-Throughput LLM Serving	Multi-tenant GPU cluster, constrained VRAM	vLLM with PagedAttention	Non-contiguous virtual memory block paging eliminating KV cache physical fragmentation
Multi-Step Tool-Using Autonomous Tasks	Unpredictable environment, external APIs	ReAct Agent with Reflexion Loop	Interleaved Thought-Action-Observation loops with verbal episodic self-correction memory
Ultra-Low Latency Mobile/Edge Deployment	Embedded devices, strict memory limits	AWQ / GPTQ INT4 + ONNX Runtime	Salient channel protection / Hessian error compensation with fused integer kernels

2. Master Mathematical Equation Reference#

1. Classification & Information Theory#

Sigmoid Logistic Function:

\sigma(z) = \frac{1}{1 + e^{-z}}, \quad \frac{d\sigma(z)}{dz} = \sigma(z)(1 - \sigma(z))

Softmax Categorical Function:

\text{Softmax}(z_i) = \frac{e^{z_i}}{\sum_{j=1}^K e^{z_j}}, \quad \frac{\partial \text{Softmax}(z_i)}{\partial z_j} = \text{Softmax}(z_i)(\delta_{ij} - \text{Softmax}(z_j))

Cross-Entropy / Negative Log-Likelihood Loss:

\mathcal{L}_{\text{CE}} = -\sum_{k=1}^K y_k \log(\hat{y}_k)

Shannon Entropy:

H(X) = -\sum_{i=1}^n p(x_i) \log_2 p(x_i)

Gini Impurity:

\text{Gini}(S) = 1 - \sum_{k=1}^C p_k^2

2. Deep Learning Initialization & Normalization#

Xavier / Glorot Initialization Variance:

\text{Var}(W) = \frac{2}{n_{\text{in}} + n_{\text{out}}}

He / Kaiming Initialization Variance (for ReLU):

\text{Var}(W) = \frac{2}{n_{\text{in}}}

Scaled Dot-Product Attention:

\text{Attention}(Q, K, V) = \text{softmax}\left( \frac{Q K^T}{\sqrt{d_k}} \right) V

RMSNorm (Root Mean Square Normalization):

\text{RMS}(x) = \sqrt{\frac{1}{d} \sum_{i=1}^d x_i^2 + \epsilon}, \quad \text{RMSNorm}(x) = \frac{x}{\text{RMS}(x)} \odot \gamma

3. Reinforcement Learning & Bellman Equations#

Bellman Expectation Equation for $V^\pi(s)$ :

V^\pi(s) = \sum_{a \in \mathcal{A}} \pi(a \mid s) \left[ R(s, a) + \gamma \sum_{s' \in \mathcal{S}} P(s' \mid s, a) V^\pi(s') \right]

Bellman Optimality Equation for $Q^*(s, a)$ :

Q^*(s, a) = R(s, a) + \gamma \sum_{s' \in \mathcal{S}} P(s' \mid s, a) \max_{a' \in \mathcal{A}} Q^*(s', a')

Policy Gradient Theorem:

\nabla_\theta J(\theta) = \mathbb{E}_{\tau \sim \pi_\theta} \left[ \sum_{t=0}^T \nabla_\theta \log \pi_\theta(a_t \mid s_t) G_t \right]

PPO Clipped Surrogate Objective:

L^{\text{CLIP}}(\theta) = \hat{\mathbb{E}}_t \left[ \min\left( r_t(\theta)\hat{A}_t, \, \text{clip}(r_t(\theta), 1 - \epsilon, 1 + \epsilon)\hat{A}_t \right) \right]

4. Multimodal, GNNs & Systems Optimization#

Symmetric InfoNCE Contrastive Loss (CLIP):

\mathcal{L}_{\text{CLIP}} = -\frac{1}{2N} \sum_{i=1}^N \left[ \log \frac{\exp(s_{ii}/\tau)}{\sum_k \exp(s_{ik}/\tau)} + \log \frac{\exp(s_{ii}/\tau)}{\sum_k \exp(s_{ki}/\tau)} \right]

GCN Renormalized Layer:

H^{(l+1)} = \sigma \left( \tilde{D}^{-1/2} \tilde{A} \tilde{D}^{-1/2} H^{(l)} W^{(l)} \right), \quad \tilde{A} = A + I_N

KV Cache VRAM Consumption:

\text{Bytes per Token} = 2 \times n_{\text{layers}} \times n_{\text{kv\_heads}} \times d_{\text{head}} \times P_{\text{bytes}}

Uniform Affine Quantization:

q = \text{clamp}\left( \text{round}\left( \frac{x}{S} \right) + Z, \, q_{\min}, \, q_{\max} \right), \quad S = \frac{x_{\max} - x_{\min}}{q_{\max} - q_{\min}}

3. High-Frequency Technical Interview Scenarios & Master Responses#

Architecture Diagram

Rendering diagram...

Scenario 1: Mitigating Catastrophic Forgetting in LLMs#

Question: "When fine-tuning a foundation model on proprietary legal documents, performance on general reasoning collapses. How do you detect and mitigate this?"

Master Response:

Diagnosis: Measure downstream performance on standard general benchmarks (MMLU, GSM8k) concurrently with legal task evaluation. The degradation stems from large unconstrained weight updates that disrupt pre-trained representations.
Mitigation 1 (Parameter Isolation): Employ LoRA (Low-Rank Adaptation). Freezing the base model parameters $W_0$ entirely guarantees that general knowledge encoded in $W_0$ remains pristine, while task adaptation is constrained to low-rank delta matrices $\Delta W = B \cdot A$ .
Mitigation 2 (Replay Mixture): Blend 10% to 20% of general domain pre-training data into the fine-tuning dataset to anchor the loss landscape.
Mitigation 3 (KL Penalty): Add a Kullback-Leibler divergence penalty against the frozen reference model: $\mathcal{L}_{\text{total}} = \mathcal{L}_{\text{task}} + \beta D_{\text{KL}}(\pi_\theta \parallel \pi_{\text{ref}})$ .

Scenario 2: Double Q-Learning vs Standard Q-Learning#

Question: "Prove why Q-learning overestimates action values, and explain how Double DQN resolves this mathematically."

Master Response:

Mechanism: The target in Q-learning contains the term $\max_{a'} Q(s', a')$ . Since the maximum operator $g(x) = \max_a x_a$ is convex, Jensen's inequality guarantees that:

\mathbb{E}\left[ \max_a \hat{Q}(s, a) \right] \ge \max_a \mathbb{E}\left[ \hat{Q}(s, a) \right]

Even if the estimates $\hat{Q}(s, a)$ have zero-mean unbiased noise, the expected value of their maximum is strictly greater than the maximum of their true expectations.

Resolution: Double DQN decouples action selection from action evaluation:

Y_t^{\text{Double}} = R + \gamma Q\left(S', \arg\max_{a'} Q(S', a'; \theta); \theta^-\right)

The online network $\theta$ selects the best action candidate, while the frozen target network $\theta^-$ evaluates its value. Overestimations in $\theta$ do not correlate with overestimations in $\theta^-$ , eliminating the systematic positive bias.

Scenario 3: DPO vs PPO in LLM Alignment#

Question: "Why has Direct Preference Optimization (DPO) largely replaced PPO in modern LLM alignment pipelines?"

Master Response:

PPO Complexity: Standard RLHF with PPO requires loading four distinct large models into GPU memory simultaneously: the Actor (policy being trained), Critic (value network), Reference Model (frozen base for KL penalty), and Reward Model. It also suffers from high training instability and sensitive hyperparameter tuning.
DPO Innovation: Rafailov et al. proved mathematically that under the Bradley-Terry preference model, the optimal policy $\pi^*$ can be expressed analytically as a function of the ground-truth reward:

r(x, y) = \beta \log \frac{\pi^*(y \mid x)}{\pi_{\text{ref}}(y \mid x)} + \beta \log Z(x)

Substituting this identity directly into the Bradley-Terry likelihood eliminates the reward model entirely, yielding the closed-form DPO objective:

\mathcal{L}_{\text{DPO}}(\theta) = -\mathbb{E}_{(x, y_w, y_l)} \left[ \log \sigma \left( \beta \log \frac{\pi_\theta(y_w \mid x)}{\pi_{\text{ref}}(y_w \mid x)} - \beta \log \frac{\pi_\theta(y_l \mid x)}{\pi_{\text{ref}}(y_l \mid x)} \right) \right]

DPO trains stably with standard binary cross-entropy on the policy network alone, cutting GPU VRAM requirements in half.

4. Final Curriculum Knowledge Check#

Concept check

Question

question: "In high-throughput LLM serving, what is the primary difference between the roles of FlashAttention and PagedAttention?"

options:

"(Recommended) FlashAttention optimizes GPU SRAM/HBM IO memory access during dense attention computation, while PagedAttention eliminates physical VRAM fragmentation of the KV cache across multi-tenant sequences"
"FlashAttention is for text while PagedAttention is for images"
"PagedAttention trains the model while FlashAttention only works during inference"
"FlashAttention quantizes weights to 4 bits while PagedAttention uses 16 bits"

explanation: "FlashAttention is an exact IO-aware attention kernel that computes softmax without materializing the N x N attention matrix into slow HBM. PagedAttention is a memory management architecture for serving engines (vLLM) that allocates KV cache in non-contiguous virtual memory blocks, eliminating allocation waste."

:::

Concept check

Question

question: "Under the Message Passing Neural Network (MPNN) framework, why must the neighborhood aggregation operator be permutation invariant?"

options:

"(Recommended) Because nodes in an undirected graph have no canonical or natural ordering; swapping the arbitrary indexing of neighboring nodes must not alter the resulting representation"
"Because permutation invariance guarantees that the adjacency matrix is invertible"
"Because it prevents models from generating complex numbers"
"Because gradient descent cannot optimize non-symmetric matrices"

explanation: "Graph topology is fundamentally unordered. If aggregation were sensitive to permutation (like standard sequence RNNs), relabeling the arbitrary indices of neighboring nodes would alter the node's updated representation, violating graph isomorphism."

:::