Master the AI Stack

Why several specialised agents often beat one large prompt, and when they do not. Planner-executor-reviewer structures, supervisor crews, shared state, and LangGraph, CrewAI and AutoGen compared on their real differences. You'll build: A multi-agent workflow where a planner delegates, executors work in parallel and a reviewer validates the result before it ships.

• Specialisation beats one giant prompt
• Working around context-window limits
• Separation of concerns & testability
• The cost & latency you trade away
• When NOT to go multi-agent
• Planner → Executor → Reviewer
• Supervisor / worker crews
• Sequential vs parallel execution
• Conditional routing & handoffs
• Shared state & memory synchronisation
• Goal decomposition & planning
• Self-review and critique loops
• Agent-to-Agent (A2A) messaging
• Failure recovery & loop detection
• LangGraph · CrewAI · AutoGen compared

The model layer itself. Transformers, tokens and context windows, the reasoning models, prompt and function calling, structured output, and the fine-tuning-versus-RAG-versus-prompting decision — plus what each option costs. You'll build: An LLM application with function calling and validated structured output — plus a fine-tuned model trained on our RTX 5090 GPU lab.

• Foundation models & transformers
• Tokens, embeddings & context windows
• Inference, temperature & sampling
• Reasoning models
• Open-source vs commercial
• Prompt engineering & system prompts
• Function calling & tool use
• Structured outputs (JSON mode)
• Fine-tuning vs RAG vs prompting
• Cost, latency & rate limits
• GPT · Claude · Gemini
• Llama · Mistral · DeepSeek
• Choosing the right model per task
• Running open models locally
• How enterprises deploy them

Retrieval in depth, because this is what most enterprise AI actually is. Chunking and embeddings, vector databases, hybrid search and re-ranking, citation assembly, and permission-aware retrieval on real corpora. You'll build: A complete RAG pipeline — ingest → chunk → embed → retrieve → re-rank → answer with citations — then evaluate its accuracy.

• Hallucination & stale knowledge
• Grounded, cited answers
• RAG vs fine-tuning vs context
• Keeping company data private
• Evaluating retrieval quality
• Chunking strategies & embeddings
• Vector databases & indexing
• Semantic + hybrid search
• Re-ranking & metadata filtering
• Context assembly & citations
• Internal knowledge assistants
• PDF & document chatbots
• Policy, contract & compliance search
• Support & FAQ grounding
• Permission-aware retrieval

Model Context Protocol: the standard that lets one agent reach GitHub, a database, Drive, Slack or a CRM without a bespoke integration each time — and the auth, injection defence and audit trail that has to come with it. You'll build: Working MCP-connected tools inside your agents — so one assistant can search, look up and act across real systems safely.

• What MCP is & why it's becoming standard
• MCP clients & MCP servers
• Tools, resources & prompts
• Request → context → response flow
• Standardising across providers
• GitHub & developer tooling
• Databases & internal APIs
• Drive, Notion & Slack
• Ticketing & CRM systems
• Line-of-business systems
• Auth & permission checks
• Preventing context leakage
• Prompt-injection defence
• Audit trails & observability
• Human-in-the-loop approvals

Automation as the point of all the preceding layers. Trigger-process-deliver pipelines, scheduled and event-driven runs, approval checkpoints, and the tools that carry it across email, CRM, HR and support. You'll build: End-to-end automations that classify, generate and deliver — running on their own, without you.

• Workflow & intelligent automation
• Trigger → process → deliver pipelines
• Scheduled & event-driven runs
• Decision automation
• Human approval checkpoints
• Email & inbox automation
• CRM & sales workflows
• HR & onboarding processes
• Customer support routing
• Reporting & data processing
• n8n · Zapier · Make
• LangGraph · CrewAI · AutoGen
• LangChain & LLM APIs
• Power Automate & Copilot
• Custom API integrations

The layer that decides whether any of this is safe to ship. Groundedness and RAG metrics, golden datasets and regression testing, injection and PII defence, and tracing, cost and audit trails once it is running. You'll build: An evaluation harness and guardrail layer over your own RAG assistant — so you can measure accuracy before and after a change instead of guessing.

• Groundedness & hallucination rate
• RAG metrics: context precision, recall, faithfulness
• Answer relevance & citation accuracy
• Golden datasets & regression testing
• Comparing prompts, models & versions
• Prompt injection & jailbreak defence
• PII detection, redaction & data privacy
• Input / output validation & schemas
• Scoping tool & data permissions
• Content safety and refusal behaviour
• Tracing multi-step agent runs
• Token cost & latency monitoring
• Audit trails for every action taken
• Bias & fairness review
• Human-in-the-loop approval design