Generative AI and Large Language Models Training Course

Participants can create tested prompts that produce consistent, task-specific outputs.

• Prompt anatomy: instructions, context, examples, constraints and output format
• Zero-shot, few-shot and chain-of-thought prompting compared for task types
• System prompts establish roles, behavioural boundaries and response priorities
• JSON schemas constrain extraction, classification and workflow-ready model outputs
• Prompt templates parameterise recurring tasks with approved business variables
• Output failures diagnosed through ambiguity, missing context and conflicting instructions

Participants can design a retrieval-augmented generation solution for controlled knowledge access.

• Embeddings represent semantic meaning as searchable numerical vectors
• Chunking strategies balance document context, retrieval precision and token cost
• Vector databases support similarity search through cosine distance ranking
• Hybrid retrieval combines keyword search, metadata filters and vector similarity
• RAG pipelines ground answers in approved source documents and citations
• Retrieval quality measured through recall, precision, relevance and attribution

Participants can select an LLM and deployment approach against defined technical and business requirements.

• Proprietary, open-weight and domain-specific models compared against requirements
• Model selection criteria cover quality, latency, context, cost and residency
• API calls manage messages, parameters, authentication, quotas and error handling
• Function calling connects models with approved systems and business actions
• Structured output validation prevents malformed data entering downstream workflows
• Cloud, private cloud and on-premises deployment trade-offs assessed systematically

Participants can evaluate LLM applications using repeatable datasets, rubrics and performance measures.

• Evaluation datasets represent priority tasks, edge cases and prohibited requests
• Quality rubrics score factuality, completeness, relevance, style and safety
• Golden datasets establish expected outputs for regression testing
• Human evaluation calibrates judgement criteria between reviewers and subject experts
• Automated evaluation compares outputs using rule-based and model-based graders
• Error taxonomies classify retrieval, reasoning, formatting and safety failures

Participants can identify major LLM risks and specify proportionate technical and governance controls.

• Prompt injection attacks manipulate model instructions through untrusted content
• Data classification determines permitted inputs, retention and model access
• Personally identifiable information requires minimisation, redaction and access controls
• Content filters detect harmful, restricted or policy-violating generated material
• Human-in-the-loop approval thresholds match consequence, uncertainty and user authority
• NIST AI RMF functions organise govern, map, measure and manage activities

Participants can define an application architecture that connects LLMs safely to enterprise services.

• Reference architecture separates user interface, orchestration, retrieval and model services
• Orchestration layers manage prompts, tools, context assembly and response validation
• Identity and access management restricts data retrieval by user permissions
• Audit logs capture prompts, sources, outputs, decisions and user actions
• Workflow integration identifies handoffs, exceptions, approvals and service-level requirements
• Observability tracks latency, token usage, errors, quality and safety events

Participants can estimate operating costs and define monitoring requirements for production LLM services.

• Token-based pricing models estimate input, output and embedding expenditure
• Caching reduces repeated inference costs for stable prompts and responses
• Latency budgets allocate response time across retrieval, inference and validation
• Rate limits, retries and fallback models maintain service availability
• Model monitoring detects quality drift, changing usage and rising costs
• Service operating procedures define incident response, escalation and model updates

Participants can prioritise LLM opportunities using a defensible business case and risk assessment.

• Process decomposition identifies tasks involving language, knowledge and judgement
• Use-case canvases document users, inputs, outputs, controls and success measures
• Value estimation quantifies time saved, quality improvement and risk reduction
• Feasibility scoring tests data readiness, integration complexity and model suitability
• Risk assessment considers harm severity, affected groups and control effectiveness
• Prioritisation matrices rank candidates by value, effort, risk and strategic fit

Participants can prepare a phased plan for piloting, governing and scaling an LLM application.

• Pilot scope defines user group, workflow boundary, data sources and measures
• Acceptance criteria specify quality thresholds, safety checks and escalation routes
• Cross-functional roles assign product, engineering, legal, security and business accountabilities
• Vendor due diligence examines model terms, data handling and service commitments
• Change management prepares users for limitations, review duties and feedback processes
• Ninety-day action plans sequence discovery, prototype, evaluation and deployment decisions