LLM Engineering & Generative AI Masterclass

• 1.1. What is AI, ML, and Deep Learning?
• 1.2. Neural Networks: Perceptrons, Layers, Activation Functions
• 1.3. The Transformer Architecture: Encoder & Decoder
• 1.4. Attention Mechanisms: Self-Attention & Multi-Head Attention
• 1.5. Tokenization: BPE, WordPiece, SentencePiece
• 1.6. Embedding Concepts: Word2Vec, Positional Encodings
• 1.7. Pre-training vs Fine-tuning vs Prompting
• 1.8. Overview of Popular LLMs: GPT, LLaMA, Mistral, Gemma

Understand the foundational theory behind large language models including transformers, attention, and tokenization that underpin all modern LLMs.

A solid theoretical foundation enables you to make informed engineering decisions, debug model behaviour, and understand why certain prompting and fine-tuning strategies work.

• 2.1. Python Essentials: Data Structures, OOP, Type Hints
• 2.2. NumPy: Array Operations, Broadcasting, Linear Algebra
• 2.3. PyTorch Tensors: Creation, Operations, Autograd
• 2.4. Building Neural Networks with torch.nn
• 2.5. Training Loops: Loss Functions, Optimisers, Backprop
• 2.6. GPU Training: CUDA, device management, mixed precision
• 2.7. Hugging Face Datasets & Tokenizers library
• 2.8. Model Checkpointing & Reproducibility

Write production-quality Python and use PyTorch to build, train, and evaluate neural network models on GPU, ready for LLM engineering tasks.

PyTorch is the dominant framework for LLM research and engineering. Proficiency here unlocks fine-tuning, custom training pipelines, and the entire Hugging Face ecosystem.

• 3.1. Open-Source LLM Landscape: LLaMA, Mistral, Phi, Gemma
• 3.2. Ollama Setup: Install, Pull, Run models locally
• 3.3. llama.cpp: GGUF format, quantization levels (Q4, Q8)
• 3.4. vLLM Server: PagedAttention, continuous batching
• 3.5. OpenAI-compatible API serving with vLLM
• 3.6. Model Benchmarking: speed, memory, quality comparison
• 3.7. Context Windows & Prompt Formatting
• 3.8. Structured Output: JSON mode, grammar constraints

Set up and serve open-source LLMs locally and on a server, compare quantized models, and expose an OpenAI-compatible inference endpoint.

Running LLMs locally with tools like Ollama and vLLM eliminates API costs, ensures data privacy, and gives you full control over the inference stack.

• 4.1. LangChain Architecture: Chains, Runnables, LCEL
• 4.2. LLM Wrappers: OpenAI, Ollama, HuggingFace integrations
• 4.3. Memory: ConversationBufferMemory, SummaryMemory
• 4.4. Document Loaders: PDF, CSV, Web, Notion
• 4.5. Text Splitters & Chunking Strategies
• 4.6. LlamaIndex: Data Connectors, Index Types
• 4.7. Query Engines: SimpleQuery, RouterQuery
• 4.8. Agents with Tools in LangChain

Build end-to-end LLM applications using LangChain's LCEL chains and LlamaIndex query engines with persistent memory and document ingestion.

LangChain and LlamaIndex are the two most widely adopted LLM orchestration frameworks — mastering them is essential for any LLM engineering role.

• 5.1. RAG Architecture: Indexing, Retrieval, Generation
• 5.2. Embedding Models: sentence-transformers, OpenAI, BGE
• 5.3. Vector Stores: Milvus setup, collections, index types (IVF, HNSW)
• 5.5. Chunking Strategies: fixed, recursive, semantic
• 5.6. Similarity Search: cosine, dot product, Euclidean
• 5.7. Hybrid RAG: dense + sparse (BM25) retrieval
• 5.8. Metadata Filtering & Namespaces

Design and implement production-grade RAG pipelines using Milvus vector database with hybrid retrieval and metadata filtering.

RAG is the most in-demand LLM engineering skill in the industry — it powers enterprise chatbots, knowledge assistants, and search systems at scale.

• 6.1. Naive RAG vs Advanced RAG vs Modular RAG
• 6.2. Re-ranking: Cross-encoders, Cohere Rerank, FlashRank
• 6.3. HyDE: Hypothetical Document Embeddings
• 6.4. Query Transformation: Decomposition, Step-back prompting
• 6.5. Knowledge Graphs: Concepts, Ontology, Triplets
• 6.6. Neo4j: Cypher queries, graph schema, LLM integration
• 6.7. GraphRAG: Microsoft GraphRAG patterns & pipeline
• 6.8. Combining Vector & Graph retrieval

Implement advanced retrieval techniques like re-ranking and HyDE, and build GraphRAG pipelines that combine vector search with Neo4j knowledge graphs.

Advanced RAG and GraphRAG dramatically improve LLM answer quality on complex, multi-hop questions — a key differentiator for production AI systems.

• 7.1. Agentic AI: What makes an agent vs a chain?
• 7.2. ReAct Pattern: Reason + Act with tool calls
• 7.3. LangGraph: Stateful graphs, nodes, edges, cycles
• 7.4. Building Multi-Agent Workflows with LangGraph
• 7.5. AutoGen: Conversation-based multi-agent system
• 7.7. MCP (Model Context Protocol): tool & resource servers
• 7.8. Tool Use: web search, code executor, file I/O agents

Design and implement multi-agent AI systems using LangGraph and AutoGen with tool use, MCP integration, and stateful agentic loops.

AI Agents are transforming software engineering — companies are building entire autonomous workflows that replace manual processes using multi-agent LLM systems.

• 8.1. Fine-Tuning vs Prompt Engineering vs RAG: when to use what
• 8.2. Supervised Fine-Tuning (SFT): dataset preparation, formats
• 8.3. Unsloth: 2x faster fine-tuning, memory optimisation
• 8.4. PEFT library: adapter merging, saving, loading
• 8.5. Training on Custom Datasets: Alpaca, ShareGPT formats
• 8.6. Pushing fine-tuned models to Hugging Face Hub

Fine-tune open-source LLMs on custom datasets using Unsloth for 2x faster training, and publish trained adapters to Hugging Face Hub.

Fine-tuning lets you specialise general LLMs for specific domains and tasks — creating proprietary models that outperform general-purpose models at a fraction of the cost.

• 9.1. FastAPI: REST endpoints, async, Pydantic models
• 9.2. Docker: Containerising LLM apps, multi-stage builds
• 9.3. Redis: Caching LLM responses, semantic caching
• 9.4. LLM Evaluation: RAGAS metrics (faithfulness, relevance)
• 9.5. DeepEval: Unit testing LLM outputs
• 9.6. Langfuse: LLM observability, tracing, prompt management
• 9.7. CI/CD for LLM Apps: GitHub Actions, automated evals
• 9.8. Cloud Deployment: Hugging Face Spaces, Modal, Render

Deploy production-ready LLM applications with FastAPI and Docker, implement semantic caching with Redis, and set up comprehensive evaluation and observability pipelines.

LLM engineering doesn't stop at building — deploying, monitoring, and evaluating LLM applications in production is what separates junior from senior LLM engineers.

• Build RAG chatbot on custom docs
• Deploy fine-tuned LLM API
• Multi-agent research assistant
• LLM evaluation dashboard
• Final LLM Engineer Certification

Complete 2 production-grade LLM projects for portfolio. Showcase real RAG pipelines and fine-tuned models to employers and clients.