AI (Generative & Agentic)

7 subjects · 19 chapters · 76 topics

1. 01

   Foundations of Modern AI

   1.1 How models learn

   • Supervised, unsupervised and reinforcement learning
   • Loss functions and gradient descent
   • Training, validation and test splits
   • Overfitting, regularisation and early stopping

   1.2 Neural networks in practice

   • Perceptrons to deep networks
   • Activation functions and initialisation
   • Backpropagation, step by step
   • Batching, epochs and learning-rate schedules

   1.3 Working with text and tokens

   • Tokenisation and vocabularies
   • Word vectors and semantic distance
   • Context windows and truncation
   • Cost and latency of long inputs
2. 02

   Transformers and Large Language Models

   2.1 The transformer architecture

   • Self-attention and multi-head attention
   • Positional encoding
   • Encoder, decoder and encoder-decoder models
   • Why scale changed everything

   2.2 Using an LLM well

   • Temperature, top-p and sampling
   • System, user and assistant roles
   • Structured output and JSON mode
   • Streaming and token budgeting

   2.3 Open and hosted models

   • Comparing model families
   • Running a model locally
   • Quantisation and hardware limits
   • Choosing a model for a given job
3. 03

   Prompt Engineering and Fine-Tuning

   3.1 Prompting techniques

   • Zero-shot, few-shot and chain-of-thought
   • Role and constraint prompting
   • Decomposing a task into steps
   • Common prompt failures and fixes

   3.2 Adapting a model

   • When fine-tuning beats prompting
   • Preparing and cleaning a dataset
   • LoRA and parameter-efficient tuning
   • Measuring whether tuning helped
4. 04

   Retrieval-Augmented Generation

   4.1 Embeddings and vector search

   • What an embedding represents
   • Cosine similarity and nearest neighbours
   • Vector databases: FAISS, Chroma, pgvector
   • Indexing and re-indexing strategy

   4.2 Building a RAG pipeline

   • Chunking documents sensibly
   • Query rewriting and hybrid search
   • Re-ranking retrieved passages
   • Citing sources in the answer

   4.3 Making RAG reliable

   • Handling 'not in the documents'
   • Freshness and cache invalidation
   • Evaluating retrieval separately from generation
   • Cost control at scale
5. 05

   Generative Media

   5.1 Image generation

   • Diffusion models explained simply
   • Prompting for images, and negative prompts
   • Inpainting, outpainting and img2img
   • Control methods and reference images

   5.2 Audio and video

   • Speech-to-text and text-to-speech
   • Voice cloning and its limits
   • Generated video: current capability
   • Watermarking and provenance
6. 06

   Agentic AI

   6.1 Agents that plan and act

   • The reason-act loop
   • Tool and function calling
   • Short-term and long-term memory
   • Multi-step plans and self-correction

   6.2 Multi-agent systems

   • Splitting work between agents
   • Passing state between agents
   • Supervisor and worker patterns
   • Where multi-agent adds nothing

   6.3 Frameworks

   • LangChain and LangGraph
   • Model Context Protocol (MCP)
   • Building an agent without a framework
   • Debugging and tracing an agent run
7. 07

   Evaluation, Safety and Deployment

   7.1 Guardrails

   • Prompt injection and how to resist it
   • Input and output filtering
   • Handling personal data in prompts
   • Human-in-the-loop checkpoints

   7.2 Measuring quality

   • Building an evaluation set
   • LLM-as-judge, and its blind spots
   • Regression testing a prompt change
   • Tracking cost, latency and failure rate

   7.3 Shipping it

   • Serving an API around a model
   • Caching and rate limiting
   • Monitoring in production
   • A capstone project, end to end