Skip to content

PragmaticML — The Anti-LLM Playbook for Principal ML Engineers

PragmaticML is a specialized, zero-cost Machine Learning study platform and architectural reference manual. It addresses "LLM Reflex Syndrome"—the industry tendency to deploy expensive, slow, non-deterministic 70B+ parameter generative LLMs for tasks that classical and targeted machine learning algorithms solve in 2 milliseconds with zero API costs.

💡 Architecture & Key Features

1. Config-Driven Modular Curriculum (lib/curriculum/)

  • Declarative Schema: Every machine learning concept is declared via a strongly-typed TypeScript configuration (types.ts), decoupling mathematical theory, diagrams, and code recipes from page routing.
  • Automated Static Generation (SSG): All concept routes (/concepts/[slug]) are pre-rendered at build time with 100% type safety and zero server-side rendering latency.

2. The 10 Principal ML Curriculum Pillars

  1. Graph & Network Algorithms: PageRank (Random walk stationary distributions), Dijkstra's Algorithm (Greedy priority-queue pathfinding), A* Search (Heuristic guidance), and Bellman-Ford (Negative cycle arbitrage).
  2. Distance & Similarity Metrics: Cosine Similarity & Angular Distance, Euclidean ($L_2$) & Manhattan ($L_1$), Levenshtein Dynamic Programming, Jaccard Index & Bitwise Hamming, and Mahalanobis Covariance Distance.
  3. Classical Supervised Learning: Logistic Regression, Linear Regression with Ridge/Lasso/ElasticNet, Decision Trees (CART), Random Forests (Bagging & OOB), Gradient Boosted Trees (XGBoost/LightGBM), SVMs (RBF Kernel Trick), Naive Bayes, and k-NN.
  4. Unsupervised Learning & Dimensionality Reduction: Principal Component Analysis (PCA & SVD), K-Means & K-Means++, DBSCAN (Density reachability & outlier rejection), and t-SNE & UMAP (Manifold learning).
  5. Recommenders & Collaborative Filtering: User/Item Collaborative Filtering, Matrix Factorization (ALS / SVD), and Two-Tower Dual Encoders for billion-scale candidate generation.
  6. Classical NLP & Information Extraction: GLiNER (Zero-shot bidirectional span NER), TF-IDF statistical weighting, Word2Vec (Skip-gram & CBOW), GloVe & FastText, and VADER rule-based sentiment.
  7. Search, Retrieval & Ranking: Okapi BM25 (Term saturation & length normalization), Dense Vector Retrieval (Bi-encoders), Cross-Encoder Re-Ranking, Reciprocal Rank Fusion (Hybrid Search RRF), and HNSW (Hierarchical Navigable Small World).
  8. Deep Learning & Neural Architectures: Convolutional Neural Networks (CNNs & ResNet skip connections), Multi-Layer Perceptrons & Backprop, RNN/LSTM/GRU, Transformers & Scaled Dot-Product Self-Attention, and Encoder vs Decoder vs Seq2Seq archetypes.
  9. Evaluation, Loss Functions & Calibration: Loss Functions (Cross-Entropy, MSE, Focal Loss for extreme class imbalance), Classification Metrics (ROC-AUC vs PR-AUC & $F_1$), and Probability Calibration (Brier score).
  10. Production ML Systems & Optimization: Model Quantization (FP16, INT8, INT4, PTQ vs QAT), Model Distillation & Pruning, Covariate Shift & Concept Drift Detection (PSI & KS-test), and Inference Latency vs Throughput Engineering.

3. Integrated Architectural & Helper Diagrams

  • Every concept features a native dark-mode Mermaid architectural diagram or visual schema illustrating exact tensor transformations, memory layouts, and data pipelines.
  • Formulations and loss objectives rendered via KaTeX LaTeX formatting.

4. Principal ML Engineer Interview Focus

  • Each module includes real-world Principal ML system design interview questions, failure modes in production, and architectural tradeoffs.

🛠 Tech Stack & Port Mapping

  • Framework: Next.js 16 (App Router), React 19, Tailwind CSS v4, TypeScript 5
  • Diagram Engine: Mermaid.js client-side rendering
  • Math Typesetting: KaTeX (katex.renderToString)
  • Local Port Mapping: 3008 (http://localhost:3008)
  • Production Domain: https://pragmatic-ml.anandmuraleedharan.com

🚀 Deployment & Subdomain Binding

1. Vercel Deployment

cd apps/pragmatic-ml
npx vercel --prod --yes

2. Custom Subdomain Binding

npx vercel domains add pragmatic-ml.anandmuraleedharan.com pragmatic-ml

3. Spaceship DNS Configuration

Add a DNS CNAME record in Spaceship DNS: * Host: pragmatic-ml * Type: CNAME * Target: cname.vercel-dns.com