Applied Generative AI & Agentic Systems

Learning outcomes

• Map tokens, context windows, inference and model capabilities
• Compare model families, modalities, latency and cost trade-offs
• Use hosted model APIs with streaming and structured responses
• Manage instructions, messages and bounded conversation state
• Handle rate limits, retries, secrets and structured errors
• Recognise hallucination, prompt injection and privacy risks early

Learning outcomes

• Design context with explicit goals, boundaries and evidence
• Use examples, decomposition and self-checking selectively
• Generate schema-valid JSON with typed structured outputs
• Build reusable prompt templates and version them like code
• Create golden datasets, rubrics and automated evaluation checks
• Measure quality, latency, cost and failure modes across models

Learning outcomes

• Prepare documents with parsing, chunking and metadata strategies
• Create embeddings and compare vector search approaches
• Combine semantic, keyword and hybrid retrieval
• Improve relevance with query rewriting, reranking and filters
• Enforce document permissions and metadata-based access controls
• Generate cited answers with source-aware response policies
• Evaluate retrieval quality separately from answer quality

Learning outcomes

• Design clear tool schemas and reliable function contracts
• Build and connect MCP clients and servers for reusable capabilities
• Validate arguments and control tool permissions before execution
• Connect agents to APIs, databases, files, browsers and search
• Model sequential, parallel and conditional workflow steps
• Handle timeouts, retries, idempotency and partial failure
• Keep humans in the loop for consequential actions

Learning outcomes

• Decide when an agent is justified and when a fixed workflow is better
• Represent state with graphs, checkpoints and typed transitions
• Separate short-term context from durable user or task memory
• Resume long-running work safely across queues, failures and approvals
• Apply planning, reflection and recovery patterns selectively
• Set budgets, stop conditions and recursion limits
• Inspect traces to understand why an agent chose each step

Learning outcomes

• Choose supervisor, handoff, router and collaborative patterns
• Use A2A and MCP for agent-to-agent and agent-to-tool interoperability
• Define agent roles, shared context and responsibility boundaries
• Route work by intent, capability, confidence and cost
• Run independent tasks concurrently and reconcile their outputs
• Prevent circular delegation, context drift and duplicated work
• Evaluate the whole system and recognise when one agent is enough

Learning outcomes

• Threat-model prompt injection, memory poisoning and unsafe tool chains
• Apply least-privilege access, approval gates and sandboxed execution
• Trace prompts, retrieval, tool calls, latency and cost with OpenTelemetry
• Build adversarial evaluations, regression suites and release thresholds
• Deploy APIs with containers, queues, caching and streaming
• Monitor drift, incidents, user feedback and cost in production

Learning outcomes

• Scope a capstone around a valuable and bounded user workflow
• Write the architecture, threat model and evaluation plan first
• Build retrieval, tools and orchestration with review checkpoints
• Test quality, safety, latency, reliability and cost trade-offs
• Publish a technical case study with traces, controls and design decisions
• Practise system design, LLM fundamentals and project walkthroughs