Core Topics

• Transformer architecture overview: attention mechanism, tokens, and embeddings
• Pre-training, fine-tuning, and RLHF: how ChatGPT and Gemini are built
• Tokens and tokenisation: why token count matters for cost and context
• Context window: what it is, its limits, and how models use it
• Model families: GPT-4o, Gemini 1.5 Pro, Claude, Llama, Mistral — a comparison
• Hallucination, bias, and why LLMs are probabilistic, not deterministic
• Practical impact of temperature, top-p, and sampling parameters

Tools & Tech

Learning Outcomes

• Explain how a large language model generates text, why hallucinations occur, and how context window size and temperature affect output quality.

Core Topics

• Zero-shot, few-shot, and many-shot prompting with examples
• Chain-of-Thought (CoT) prompting and self-consistency techniques
• System prompts: role, instructions, output format, constraints
• Structured output: prompting for JSON, markdown tables, and consistent formats
• Tree-of-Thought and ReAct (Reason + Act) prompting patterns
• Prompt injection attacks and defensive prompting strategies
• Prompt versioning and A/B testing prompts for production quality

Tools & Tech

Learning Outcomes

• Write system prompts and few-shot examples that produce consistent, structured, high-quality LLM outputs — and detect and mitigate prompt injection attempts.

Core Topics

• OpenAI API: chat completions, function calling, streaming responses
• Google Gemini API: multimodal inputs (text + image), grounding with search
• Structured output with response_format (JSON mode) and Pydantic parsing
• Tool/function calling: defining tools, handling tool results, multi-turn conversations
• Vision models: analysing images, screenshots, and documents with GPT-4o / Gemini
• Rate limiting, error handling, and retry strategies for production API calls
• Cost estimation and token optimisation strategies for real applications

Tools & Tech

Learning Outcomes

• Build a FastAPI backend that exposes an AI feature using OpenAI or Gemini function calling — with structured JSON output, vision support, and rate-limit-aware error handling.

Core Topics

• What are embeddings: vector representations of meaning in high-dimensional space
• Text embedding models: text-embedding-3-small, all-MiniLM-L6-v2, BGE
• Cosine similarity and dot product: measuring semantic distance between texts
• Building a semantic search engine from scratch with embeddings
• Embedding documents, queries, and comparing with numpy
• Chunking strategies for long documents: fixed-size, recursive, semantic chunking
• Evaluating embedding quality: retrieval precision, recall, MRR metrics

Tools & Tech

Learning Outcomes

• Build a semantic search system that finds the most relevant document passages for a natural language query using embeddings and cosine similarity.

Core Topics

• Why vector databases: beyond keyword search, scalable similarity search
• Chroma: local persistent vector store, collection management, metadata filtering
• Pinecone: cloud-managed vector DB, namespaces, upsert, query, and delete
• HNSW and IVF indexing algorithms: understanding the speed/recall tradeoff
• Hybrid search: combining vector similarity with keyword (BM25) search
• Metadata filtering: combining semantic search with structured filters
• pgvector: vector similarity search inside PostgreSQL for simpler stacks

Tools & Tech

Learning Outcomes

• Index a document corpus in both Chroma (local) and Pinecone (cloud), query with metadata filtering, and measure retrieval quality between the two setups.

Core Topics

• RAG architecture: document ingestion → chunking → embedding → retrieval → generation
• LangChain RAG chain: RetrievalQA, ConversationalRetrievalChain
• LlamaIndex query engine: document loading, indexing, retrieval modes
• Advanced RAG: re-ranking, HyDE (Hypothetical Document Embeddings), multi-query
• Chat memory: ConversationBufferMemory, ConversationSummaryMemory
• Multi-document RAG: routing queries to different knowledge bases
• Exposing a RAG pipeline as a FastAPI endpoint for a React frontend

Tools & Tech

Learning Outcomes

• Build and deploy a production-quality RAG pipeline that answers questions over a private document corpus using LangChain or LlamaIndex, exposed as a REST API.

Core Topics

• When to fine-tune vs few-shot vs RAG: the decision framework
• OpenAI fine-tuning: JSONL dataset format, uploading, training, and evaluation
• PEFT: LoRA and QLoRA for fine-tuning open-source models on limited hardware
• RAG evaluation with RAGAS: faithfulness, answer relevancy, context precision
• LLM evaluation: G-Eval, LLM-as-judge patterns, human evaluation rubrics
• AI guardrails: output validation, topic restriction, PII detection with Guardrails AI
• Observability: tracing LLM calls with LangSmith for debugging production issues

Tools & Tech

Learning Outcomes

• Fine-tune a small GPT model, evaluate a RAG pipeline with RAGAS metrics, and add input/output guardrails to an AI application before production deployment.

Core Topics

• Streamlit for rapid AI app prototyping: building a chat UI in 50 lines
• FastAPI production setup: async endpoints, streaming responses with Server-Sent Events
• Dockerising an AI application: managing model clients, environment secrets
• Deploying to Hugging Face Spaces, Railway, or Render with environment variables
• Cost monitoring: tracking API spend, setting usage limits, optimising prompt tokens
• Responsible AI considerations: transparency, disclosure, and avoiding harm
• Capstone project: full AI application from RAG pipeline to deployed web UI

Tools & Tech

Learning Outcomes

• Deploy a complete Generative AI application — RAG pipeline + API + chat UI — to a live public URL with observability, guardrails, and cost monitoring in place.