Generative AI Masterclass — 100 Hours of LLMs, RAG, Agents and Production LLMOps

• System vs user vs tool messages; role design and instruction hierarchy
• Few-shot, chain-of-thought, self-consistency, decomposition, ReAct
• Structured outputs: JSON schema, function calling, constrained decoding, Pydantic validation
• Prompt versioning, templating, regression testing, prompt injection basics
• Context engineering: what to put in the window and what to leave out
• Lab: take a flaky prompt from 62% to 94% task accuracy on a golden set

• Embedding models compared; dimensionality, normalisation, multilingual, domain adaptation
• Vector indexes: HNSW, IVF, ScaNN; recall vs latency vs memory trade-offs
• Vector stores: pgvector, Pinecone, Chroma, Weaviate, Qdrant, Databricks Vector Search, OpenSearch
• Hybrid search: BM25 + dense, reciprocal rank fusion, metadata filtering
• Reranking with cross-encoders; query rewriting and HyDE
• Lab: measured retrieval quality across five chunking strategies

• Document ingestion: PDFs, tables, images, HTML; layout-aware parsing; OCR
• Chunking strategies: fixed, recursive, semantic, hierarchical, contextual retrieval
• Grounding, citations, refusal behaviour, handling 'I don't know'
• Advanced patterns: parent-document, multi-vector, GraphRAG, agentic RAG
• Freshness: incremental indexing, deletion, permissions-aware retrieval (row-level security)
• Serving architecture: API layer, caching, streaming responses, rate limits
• Lab: Project 1 end to end

• Building golden datasets; task-specific metrics; inter-annotator agreement
• RAGAS: faithfulness, answer relevance, context precision/recall
• LLM-as-judge: rubrics, position bias, calibration, when it lies to you
• Tracing with LangSmith / MLflow / OpenTelemetry; token and latency dashboards
• Guardrails: input/output filters, PII redaction, jailbreak and prompt-injection defence
• Regression gates in CI: fail the build when quality drops
• Lab: build an eval harness and wire it into GitHub Actions

• Agent loops: ReAct, plan-and-execute, reflection; when an agent is the wrong abstraction
• Function/tool calling in depth; parallel tools; error and timeout handling
• LangGraph: state machines, checkpointing, human-in-the-loop, subgraphs
• Multi-agent patterns: supervisor, hand-off, debate — and their failure modes
• Model Context Protocol (MCP): building an MCP server, resources, tools, prompts
• Computer-use and browser agents; sandboxing and blast-radius control
• Lab: Project 2 agent with red-team suite

• Decision framework: prompt → RAG → fine-tune → pretrain (and the cost of each)
• SFT dataset construction, formatting, deduplication, contamination checks
• PEFT: LoRA, QLoRA, adapters; rank/alpha selection; catastrophic forgetting
• Preference tuning: DPO, ORPO; RLHF at a conceptual level
• Quantisation: GGUF, AWQ, GPTQ, bitsandbytes; accuracy vs memory trade-offs
• Distillation to a smaller model for cost reduction
• Lab: Project 3 fine-tune with a proper eval comparison

• Inference servers: vLLM, TGI, Ollama, TensorRT-LLM; continuous batching, paged attention
• GPU sizing and economics: A10G vs L4 vs A100 vs H100; self-host vs API break-even maths
• Model routing and cascades; semantic caching; prompt caching; batch APIs
• Deployment on Databricks Model Serving, Amazon Bedrock, Azure AI Foundry, Vertex AI
• Versioning, canary rollouts, shadow traffic, rollback
• Lab: cut cost per 1,000 requests by 70% without losing quality

• Vision, audio and document understanding; multimodal RAG
• Image and video generation overview; when it matters for enterprise work
• OWASP Top 10 for LLM applications; supply-chain risk in model weights
• Data privacy, residency, DPAs, retention; what you can and cannot send to an API
• EU AI Act and NIST AI RMF at a practitioner level; model cards and documentation

• Reference architectures for enterprise GenAI on AWS, Azure, GCP and Databricks
• Build vs buy; platform selection scorecard
• Capstone presentation and architecture defence
• GenAI engineer interview questions, portfolio framing, resume rewrite
• Staying current: how to read papers and release notes without drowning

Frontier APIs (OpenAI GPT-5.x, Anthropic Claude 4.x, Google Gemini 2.x) plus open-weight models (Llama 4, Mistral, Qwen) run locally with Ollama/vLLM. Model versions move fast — the course tracks current releases each batch and the principles are deliberately model-agnostic.

No. API work needs nothing special, and fine-tuning labs run on free Colab/Kaggle T4 GPUs using QLoRA. If you have a 16 GB+ GPU you can run everything locally.

Budget roughly $20–$40 in API credits across the course. We use small models for iteration and teach cost control from Module 1 — several students finish under $15.

Developers and architects. There is real code in every module. If you need a leadership-level overview, the 2-day executive workshop under Corporate Training is the better fit.