Learn how AI connects to your own data — then build and deploy a live RAG chatbot with citations and memory.

Topics covered

• Embeddings intuition: words to vectors, cosine similarity as semantic distance

• Embedding model comparison: OpenAI, Cohere, Voyage AI — choosing by task, cost, and latency

• Vector database landscape: Pinecone, Chroma, Weaviate, Qdrant — comparison and setup

• RAG pipeline anatomy: ingestion → chunking → embedding → storage → retrieval → augmentation → generation

• Chunking strategies: fixed-size, semantic, hierarchical, sentence-window — quality comparison

• Hybrid search: dense + sparse (BM25) + reranking with Cohere Rerank

• RAGAS evaluation framework: Faithfulness, Answer Relevancy, Context Precision, Context Recall

• Document loaders: PDF, DOCX, web pages, YouTube transcripts, Notion — including complex PDFs

You'll learn:

• How to connect any AI model to your own documents and data

• How to structure documents for better retrieval accuracy

• How to evaluate RAG quality and systematically fix failure cases

• When retrieval is better than fine-tuning — and vice versa

Practical

• Build a production RAG chatbot in FlowiseAI: ingest your own documents, deploy to a live URL

• Add multi-turn memory and source citations to your chatbot

• Run RAGAS evaluation over 20 test questions and fix the top 3 failures

• Semantic product search: embed 50 items into Pinecone, build natural-language search interface

Assignment

• Add 3 additional documents to your RAG bot. Fix 2 identified failure cases. Share improved eval scores.

Move from a chatbot that answers questions to an agent that takes actions — and build your first working agent.

Topics covered

• GenAI vs. Agentic AI: OODA loop, autonomy spectrum (copilot → assistant → operator → fully autonomous)

• 4 core agent design patterns: Reflection, Tool Use, Planning, Multi-Agent (Andrew Ng framework)

• ReAct framework deep dive: Reason → Act → Observe loop, Thought/Action/Observation trace

• Agent memory systems: in-context, entity, summary, external DB, semantic (vector search over past interactions)

• Tool ecosystem: web search (Tavily, Serper), calculator, file reader, custom REST APIs

• Agent safety and reliability: failure modes, max iterations, confirmation gates, output validators

• Human-in-the-loop: when and how to require human approval before an agent acts

You'll learn:

• How to make an AI think step by step — not just generate answers

• How to give an agent persistent memory across sessions

• How to let an agent use external tools safely

• How to handle mistakes, loops, and unexpected agent behaviour

Practical

• Build a web search + calculator agent in FlowiseAI: Tavily + calculator + custom weather tool

• Add conversation memory buffer and a self-review step — compare output quality before and after

• Build a custom domain agent for your own use case using the Agent Design Canvas

• Document 3 interesting ReAct traces and share in the cohort workspace

LangChain's mental model, LangGraph state machines, LangSmith observability — and your first graded project.

Topics covered

• LangChain mental model: LLMs, Chains, Agents, Tools, Memory, Retrievers, Callbacks, LCEL pipe operator

• Chain patterns: simple, sequential, router, map-reduce, transform — demonstrated in FlowiseAI

• LangGraph — state machines for complex agent flows: nodes, edges, state dict, conditional routing, human interrupt/resume

• Tool design principles: name, description, parameter schema — writing tool descriptions the model actually reads

• Serper, Wikipedia, and custom REST API integration as agent tools

• LangSmith observability: full chain traces, token costs, latency per step, error diagnosis, prompt versioning

• Building a router agent: classify intent, dispatch to specialist agents, share user context via memory

You'll learn:

• How to use LangChain abstractions in FlowiseAI without writing code

• How to build conditional, looping, and branching agent flows with LangGraph

• How to trace, debug, and optimise agent pipelines using LangSmith

• How to route queries intelligently across multiple specialist agents

Practical

• LangSmith setup: connect FlowiseAI, read first full chain trace

• Graded Project 2 Part 1: Research Agent with Serper + page-fetcher + Pinecone memory

• Graded Project 2 Part 2: Customer Care Agent with Google Sheets order lookup tools

• Router integration: combine both agents behind a single endpoint with shared memory buffer

Design collaborative AI systems where specialist agents divide work, coordinate tasks, and produce outputs no single agent could.

Topics covered

• Why multi-agent: specialisation, parallelism, fault isolation, and scalability

• CrewAI deep dive: Agent (role, goal, backstory, tools, LLM), Task, Crew, Process — YAML configuration

• Crew design patterns: Pipeline (A→B→C), Parallel (A+B → C), Hierarchical (Manager→Workers), Looping (until threshold)

• Memory strategies in multi-agent systems: short-term task context, long-term Pinecone persistence, shared vs. private

• Testing and debugging agent handoffs: tracing in LangSmith, isolating failures, sequential vs. parallel timing

• Human-in-the-loop approval gates in CrewAI-based workflows

You'll learn:

• How to design AI systems where different agents have different specialist roles

• How agents collaborate, delegate tasks, and share context

• How to trace and debug failures in multi-agent handoffs

• How to add shared and private memory to a multi-agent crew

Practical

• Build a 3-Agent Content Crew in FlowiseAI: Topic Researcher → Blog Writer → SEO Reviewer

• Add human-in-the-loop approval before the final output is released

• Build your own domain crew: e.g. PM → Developer → QA, or Researcher → Copywriter → Designer Brief

• Trace agent-to-agent handoffs in LangSmith and identify the top failure mode

Enterprise-grade multi-agent patterns from Microsoft and Google — and your second graded project.

Topics covered

• Microsoft AutoGen: ConversableAgent, AssistantAgent, UserProxyAgent, GroupChat, GroupChatManager

• Google Agent Development Kit (ADK): coordinator dispatches sub-agents, FunctionTool, AgentTool, streaming

• Enterprise AI patterns: API key management, input sanitisation, audit logs, PII masking, cost controls

• Scalability patterns: horizontal scaling, queue-based orchestration, caching, cheap-model routing for simple steps

• Graded Project 3: 4-agent CrewAI pipeline orchestrated by n8n — Trend Researcher, Content Writer, Brand Reviewer, Social Scheduler

You'll learn:

• How enterprise teams structure AI agent systems for security and compliance

• How to control costs across a multi-agent system

• How to build human approval gates in a real publishing or content workflow

• How to orchestrate a full multi-agent system using n8n as the outer automation layer

Practical

• Graded Project 3 full build: 4-agent pipeline with CrewAI + n8n

• Human approval webhook: n8n waits for approval before Social Scheduler fires

• All outputs logged to Google Sheets with status tracking

• Full LangSmith trace for all 4 agents

No-code automation at scale and the Model Context Protocol — connect Claude to your real business tools.

Topics covered

• n8n architecture: triggers (webhook, schedule, app), AI Agent node, OpenAI/Claude/Gemini LLM nodes

• n8n advanced patterns: sub-workflows, loop over items, error workflows, merge node, expression editor

• n8n + FlowiseAI integration: n8n as orchestrator, FlowiseAI as the AI reasoning brain

• Make.com advanced: router, iterator, aggregator, data stores, real-time webhooks

• Model Context Protocol (MCP): host-client-server architecture, JSON-RPC 2.0, stdio vs. SSE transport

• MCP server ecosystem: Gmail, Google Calendar, HubSpot, Figma, Supabase, GitHub, Slack, Google Drive

• MCP in production: OAuth 2.0, tool discoverability, sandboxing, audit logs, Cloudflare Workers hosting

• Agent observability deep dive: LangSmith trace viewer, latency P50/P95/P99, cost per task, alert design

You'll learn:

• How to build production automation workflows in n8n with AI nodes and error handling

• How the Model Context Protocol lets Claude interact with any tool without custom code

• How to connect Claude to Gmail, Calendar, and CRM tools via MCP in a single workflow

• How to monitor live agent systems with meaningful dashboards

Practical

• n8n Automation Sprint: build a lead scorer, competitive intel workflow, or support triage system

• Claude + Figma MCP demo: read a design file, suggest copy, export component specs — all in natural language

• Graded Project 4 full build: Gmail + Calendar + HubSpot MCP → Claude classification → CRM update → auto-reply → daily digest

• LangSmith observability dashboard: latency, tool call accuracy, error rate, cost/day

Assignment

• Automation Portfolio: 3 working n8n or Make.com automations you'd actually use at work. Document: trigger, actions, tools, estimated time saved per week.

Simulate a real-world enterprise AI system design review — design, document, and defend your agentic AI architecture.

Topics covered

• Enterprise AI system design review process: architecture decisions, trade-offs, documentation standards

• End-to-end capstone architecture: RAG layer, agent layer, automation layer, observability, security

• Cost modelling for AI products: inference cost, engineering overhead, expected ROI

• SLA definition and incident response planning for agentic systems

• AI governance and responsible AI: prompt injection defence, PII handling, audit trail design

• Presentation skills: communicating agentic AI architecture to technical and non-technical stakeholders

You'll learn:

• How to design an enterprise AI system that can be trusted in a production environment

• How to document architecture decisions with technical and business rationale

• How to defend design choices against challenging questions from stakeholders

• How to articulate the business value and ROI of an agentic AI system

Practical

• Design an enterprise-grade AI agent system for a chosen vertical (marketing, operations, finance, or HR)

• Write complete architecture documentation: system design, data flow, security model, cost model, observability plan

• Deliver a 20-minute architecture defence presentation to peers and mentors

• Incorporate live feedback and iterate on your documentation

Assignment

• Final architecture document submission with all required components. Peer review and instructor sign-off.