Advanced Generative AI Applications Training Course

Participants can design versioned prompt specifications that produce controlled and machine-readable outputs.

• Instruction hierarchies distinguish system, developer, user, and tool messages
• Structured prompt specifications define role, task, constraints, inputs, and outputs
• Few-shot example selection covers boundaries, counterexamples, and format consistency
• Chain-of-thought handling uses concise rationales without exposing sensitive reasoning
• JSON Schema constrains extraction, classification, and downstream application inputs
• Prompt versioning records purpose, variables, tests, owners, and release status

Participants can design multimodal AI workflows for extracting, interpreting and generating business content.

• Multimodal input design combines documents, images, audio, video, and metadata
• Document intelligence extracts tables, fields, clauses, and exceptions from scanned files
• Vision prompts request grounded observations, regions, counts, and uncertainty
• Audio transcription pipelines handle diarisation, timestamps, redaction, and quality checks
• Image generation controls composition, style references, text rendering, and provenance
• Multimodal evaluation tests OCR errors, visual ambiguity, and inaccessible content

Participants can design grounded retrieval-augmented generation solutions using trusted enterprise knowledge sources.

• RAG architecture separates ingestion, chunking, embedding, retrieval, generation, and citations
• Chunking strategies tune semantic boundaries, overlap, metadata, and document hierarchy
• Embedding selection evaluates domain fit, dimensions, latency, and retrieval cost
• Hybrid retrieval combines dense vectors, BM25, metadata filters, and reranking
• Grounded answer policies require citations, abstention, and source-date checks
• Retrieval evaluation measures recall, precision, faithfulness, and answer relevance

Participants can connect generative AI models to enterprise systems through controlled APIs and workflow patterns.

• API request patterns manage authentication, rate limits, retries, and idempotency
• Function calling maps model intents to validated enterprise service operations
• JSON Schema and Pydantic validate tool arguments before execution
• Workflow orchestration coordinates approvals, queues, notifications, and exception routes
• Webhook and event patterns connect generative AI to business processes
• Integration logging captures prompts, tool calls, outputs, latency, and failures

Participants can design bounded agent workflows with appropriate controls over actions and approvals.

• Agent designs distinguish single-turn workflows, assistants, planners, and tool-using agents
• State management maintains conversation memory, task context, and user preferences
• Tool selection policies restrict actions by role, confidence, and risk
• Human-in-the-loop checkpoints govern payments, communications, and record changes
• Multi-agent patterns assign researcher, reviewer, executor, and supervisor responsibilities
• Stop conditions prevent loops, budget overruns, duplicated actions, and unsafe escalation

Participants can evaluate generative AI quality and establish monitoring for performance, cost and failures.

• Evaluation datasets pair representative inputs with expected outputs and scoring rubrics
• LLM-as-judge methods require calibrated criteria, anchors, and human spot checks
• Regression suites detect prompt, model, retrieval, and integration degradation
• A/B tests compare completion quality, handling time, and operational error rates
• Observability dashboards track token spend, latency, refusal, and grounding metrics
• Error analysis prioritises fixes using severity, frequency, detectability, and business impact

Participants can apply practical governance controls to protect data, users and business processes.

• OWASP LLM Top 10 identifies prompt injection and data disclosure risks
• Threat modelling maps assets, actors, attack paths, controls, and residual risk
• Data classification sets permitted inputs, retention periods, and redaction requirements
• PII handling applies minimisation, masking, access controls, and audit trails
• Model risk documentation records intended use, limitations, validation, and approvals
• Copyright review distinguishes training rights, output ownership, and attribution duties

Participants can assess and prioritise high-value generative AI applications using operational and risk criteria.

• Use-case prioritisation scores value, feasibility, data readiness, risk, and sponsorship
• Claims triage design extracts evidence, flags missing information, and drafts correspondence
• Procurement copilot workflows compare bids, identify deviations, and cite contract clauses
• Customer-service assistants combine knowledge retrieval, CRM context, and escalation rules
• Analyst workflows generate research briefs, variance narratives, and source-linked summaries
• Public-sector case handling applies eligibility rules, records controls, and review gates

Participants can prepare a phased deployment plan with ownership, controls, measures and release decisions.

• Operating models define product owner, model owner, risk owner, and support roles
• Release gates require evaluation evidence, security review, user guidance, and rollback
• Change management plans address adoption, training, feedback, and service desk readiness
• Vendor due diligence examines data terms, model changes, availability, and exit options
• Value tracking links baseline measures, benefits hypotheses, costs, and realised outcomes
• Application roadmap sequences pilot scope, controls, integration, evaluation, and scale decisions