AI Engineering Bootcamp
Cohort 01 — May 2026

Tokens, context windows, temperature, system prompts, model families (GPT-4o, Claude, Gemini). Understanding what you're building on.

A spec is a structured document — requirements, behaviour, constraints — that gives an AI agent the context to generate correct code. Introduced here as an evolution of prompt engineering. You'll apply it every week from here forward. Live comparison: vague prompt vs. structured spec, same model, dramatically different output.

Call the Anthropic and OpenAI APIs, chain prompts, handle structured JSON output, basic error handling. Submit your spec alongside the code.

What RAG is and why it matters — embeddings, semantic search, vector databases (ChromaDB, Pinecone), chunking strategies and tradeoffs, the full retrieve-augment-generate pipeline. How to use your RAG system from code and from your IDE.

Build the full RAG pipeline in Python live. Then connect to it from Kiro — so you can query your documents directly from your IDE. This is how you'll use RAG in your Week 3 and Week 4 projects.

Ingest PDFs → embed → store in a vector DB → retrieve → answer questions. Full pipeline. Graded project submission. Built inside Kiro using MCP-connected tools.

ReAct agentic loops, function/tool calling. Agent architecture patterns — orchestrated vs. decentralized — and how to choose before you build. Agent memory taxonomy: in-context (active window), semantic (your RAG from Week 2), episodic (history persisted across sessions), and procedural (baked into model weights). Introduction to Claude as the reasoning engine powering your agents.

LangSmith for tracing agent execution. Setting up evaluation datasets. Understanding agent behavior through observability. Introduction to evaluation metrics and testing strategies.

Build and run an agent live that receives a goal, reasons step by step, calls real tools, reflects on results, and loops until done. Watch the full reasoning trace in LangSmith as it executes. The spec that drives it is written first — students see exactly how the specification shapes every decision the agent makes.

Write a specification first, then wire an LLM agent to external tools. The spec is part of the graded submission — not just the code. Built with Claude + Kiro.

Run evaluation experiments in LangSmith using golden datasets, comparing agent runs, analyzing traces, and identifying regressions and performance differences across versions.

Introduce evaluation frameworks (DeepEval + RAGAS) to systematically measure LLM and agent quality using metrics such as faithfulness, relevance, correctness, and retrieval grounding.

Design multi-agent systems using CrewAI, focusing on role-based agents, task decomposition, and orchestration patterns such as sequential pipelines, hierarchical supervisor models, and peer collaboration.

Connect evaluation + orchestration into a production optimization loop, where multi-agent systems are continuously improved through evaluation feedback (cost, latency, reliability, and output quality).

Design and implement a multi-agent system using CrewAI. Run evaluation experiments with LangSmith to optimize agent performance. Assess output quality with DeepEval and RAGAS metrics. Document agent roles, task orchestration, and cost optimization strategies.

Model Context Protocol (MCP) — standardized way for agents to call tools and access data across systems. MCP servers, tools, resources. Wrapping RAG and memory as MCP tools. Connecting agents to external systems.

Containerisation, Railway/Render deployment, API wrapping, structured logging, cost monitoring, production readiness checklist.

Present your deployed capstone project live. Walk through architecture decisions, defend tradeoffs, receive detailed expert feedback. Certificate awarded on completion.

Second day of capstone presentations for the cohort. Live demos, Q&A, and peer feedback.