AI & Machine Learning

Overview

Notes span classical machine learning through modern large language models, covering both theory (algorithms, model evaluation) and applied practice (deployment, agent design, LLM integrations).

---

Core Areas

Classical Machine Learning

• Supervised learning — regression, classification (SVM, decision trees, random forests)
• Unsupervised learning — clustering (k-means, DBSCAN), dimensionality reduction (PCA, t-SNE)
• Model evaluation — ROC curves, AUC, precision/recall, confusion matrices, cross-validation
• Feature engineering — normalisation, one-hot encoding, feature selection

See: Machine Learning

Deep Learning & Neural Networks

• Feedforward networks, CNNs, RNNs, LSTMs
• Backpropagation and gradient descent variants (Adam, SGD, RMSProp)
• Transfer learning and fine-tuning pre-trained models
• Regularisation: dropout, batch normalisation, weight decay

Large Language Models

• Transformer architecture: attention mechanism, positional encoding
• Prompting strategies: zero-shot, few-shot, chain-of-thought
• RAG — grounding LLM outputs with retrieved knowledge
• Embeddings and vector similarity search
• Fine-tuning vs prompting trade-offs
• Reasoning Effort Control — RLVR-trained reasoning traces, and effort as an inference-scaling knob separate from model size
• Context Engineering — what to put in front of an agent, and why capable models need less scaffolding

MLOps & Deployment

• Model versioning and experiment tracking (MLflow, Weights & Biases)
• Model serving: REST APIs, batch inference, streaming predictions
• A/B testing models in production
• Monitoring: data drift, concept drift, prediction quality

AI Agent Frameworks

• Multi-agent orchestration patterns
• Tool use and function calling
• Claude Code / Anthropic API usage
• Agent memory: short-term (context window), long-term (vector stores)

---

Relationships

• Cloud & AWS Infrastructure — SageMaker, GPU instances, EKS for ML workloads
• Entrepreneurship & SaaS — AI-powered SaaS products, LLM integrations
• Trading & Finance — quant models, predictive analytics for trading
• Retrieval-Augmented Generation (RAG) — key LLM application pattern
• Neural Network — perceptron to CNN, the basis of deep learning
• Transformer (GPT Architecture) — the decoder-only architecture behind GPT/LLMs
• Data Science Methodology — iterative process from business problem to deployment
• Regression and Statistical Inference — sampling, EDA, ANOVA, multicollinearity, and suppression effects
• Gradient Boosting — XGBoost, the leading model for tabular data
• Data Leakage — validation pitfall that inflates apparent accuracy
• AI Engineering Pitfalls — common product/eval mistakes building on foundation models
• OWASP Top 10 for LLM Applications — security risks specific to LLM apps
• Agentic Cyber Attack — autonomous agent swarms on offense, and AI-assisted forensics on defense
• Data Poisoning and Adversarial ML — attacking the model through its training data; 0.04% contamination can be enough
• System Design — ML system design (feature stores, model registries)
• Prediction Machines — the economic framing of AI as cheap prediction, and what stays scarce (judgment, data, execution)
• AI Inference Hardware — serving silicon designed against tokens-per-megawatt, and the disaggregation trade-off
• Agent Skill Evolution — improving an agent without touching weights, by compiling its traces into skills