30-Day AI Engineering Roadmap

A full-time, immersive upskilling program combining AI fundamentals with agentic development tooling. All resources are free. Built for 30+ engineers with varying experience levels.

30 Days
20 Learning Days
10 Project Days
50+ Project Ideas
6+ Hrs / Day

Fast Track Active. You're already comfortable with Cursor, Claude Code, and basic LLM APIs. Skip the intro days (marked with ⚡) and move directly to intermediate/advanced content. Use freed-up days for deeper project work.

Learn
Apply / Practice
Capstone Project

Foundations — LLMs, Prompts & Pair Programming

Get grounded in how LLMs work, master prompt engineering, and build pair-programming fluency with Cursor.

Day 01 LLM Fundamentals — How Models Work Beginner
⚡ Fast Track: Skim the transformer overview, skip to Day 2.
Objective: Understand transformer architecture, tokenization, attention mechanisms, and how LLMs generate text. Grasp the difference between open and closed models.
Transformer architecture Tokenization Attention mechanism Temperature & top-p Context window Open vs closed models
Day 02 LLM API Integration — OpenAI, Anthropic, Google Beginner
⚡ Fast Track: Jump to the Anthropic/OpenAI docs. Try structured outputs & function calling.
Objective: Write production-grade Python code that calls LLM APIs. Understand messages format, system prompts, streaming, token limits, and error handling.
Messages API System prompts Streaming responses Structured outputs Token counting Error handling
Day 03 Prompt Engineering — Techniques That Matter Intermediate
Objective: Master zero-shot, few-shot, chain-of-thought, ReAct, and structured prompting. Learn to write system prompts that produce consistent, high-quality output.
Zero-shot Few-shot Chain of Thought ReAct pattern Tree of Thought System prompt design Output formatting
Day 04 Cursor AI — Pair-Programming Fluency Beginner
⚡ Fast Track: Skip if already using Cursor daily. Read the Settings article for optimization tips.
Objective: Install Cursor, configure rules, and build something from scratch using Agent Mode and Plan Mode.
.cursor/rules/ folder Plan Mode Agent Mode YOLO Mode Composer Model selection
Day 05 Claude Code — Terminal Agentic Workflows Intermediate
⚡ Fast Track: Read Best Practices doc. Focus on CLAUDE.md setup and TDD loop.
Objective: Install Claude Code, create CLAUDE.md for your codebase, and practice the TDD agentic loop.
CLAUDE.md Thinking budget TDD agentic loop Context management /clear Custom slash commands
Day 06 Practice — Build with Cursor + Claude Code Hands-on
Objective: Spend the full day building. Create a small tool or internal utility using Cursor + Claude Code. Practice Plan Mode → Build → Test → Iterate. Write CLAUDE.md for a real repo.
Suggested Exercises
  • Build a CLI tool that uses an LLM API with proper error handling and streaming
  • Set up .cursor/rules/ for an existing project (ui.md, logic.md, testing.md)
  • Create CLAUDE.md for your team's main codebase
  • Use Claude Code's TDD loop: write failing tests → let Claude implement → iterate
Day 07 MCP — Model Context Protocol Intermediate
Objective: Understand MCP as the universal tool protocol. Install 3+ MCP servers and use them with Claude Code and Cursor on real tasks.
MCP protocol GitHub MCP Playwright MCP Context7 PostgreSQL MCP Brave Search Sequential Thinking

RAG, Function Calling & Agents

Build retrieval pipelines, connect LLMs to external tools, and design your first autonomous agents.

Day 08 RAG Foundations — Embeddings, Vectors & Retrieval Intermediate
Objective: Understand how RAG works end-to-end: document loading, chunking, embedding, vector storage, retrieval, and generation.
Embeddings Vector stores Chunking strategies FAISS / pgvector Similarity search MMR
Day 09 Advanced RAG — Hybrid Search, Reranking & Evaluation Advanced
Objective: Go beyond basic RAG. Learn semantic chunking, hybrid search (vector + keyword), reranking with Cohere, self-RAG, and corrective RAG patterns.
Hybrid search Semantic chunking Reranking Self-RAG Corrective RAG RAG evaluation Faithfulness scoring
Day 10 Practice — Build a RAG Pipeline Hands-on
Objective: Build a complete RAG system from scratch: load documents, chunk, embed, store in FAISS or pgvector, retrieve, and generate answers. Use Claude Code to accelerate.
Suggested Exercises
  • Build a Q&A chatbot over your team's internal docs using LangChain + FAISS
  • Experiment with different chunking strategies and measure retrieval quality
  • Add a reranking step and compare results
  • Deploy as a FastAPI endpoint
Day 11 Function Calling & Tool Use Intermediate
Objective: Understand how LLMs use tools via function calling. Build agents that can search the web, query databases, and call APIs.
Function calling Tool schemas Parallel tool use Error recovery LangChain tools
Day 12 LLM Agents — Design Patterns & Architecture Intermediate
Objective: Learn agent architectures: ReAct, reflection, planning. Understand context management, memory, and when to use agents vs chains.
ReAct pattern Reflection Planning agents Memory systems Context management Human-in-the-loop
Day 13 Practice — Build a Tool-Using Agent Hands-on
Objective: Build an agent that uses 3+ tools (web search, database, calculator, etc.) to answer complex questions. Practice with LangChain or raw function calling.
Suggested Exercises
  • Build a research agent that searches the web, summarizes pages, and writes a report
  • Create a database agent that translates natural language to SQL queries
  • Build a code review agent that reads PRs via GitHub MCP
Day 14 Claude Code — Subagents, Hooks & Git Worktrees Advanced
Objective: Go deeper into Claude Code: build subagents, set up Git worktrees for parallel sessions, create custom slash commands, and configure hooks for automation.
Subagents Git worktrees Custom /pr /review Hooks ultrathink Parallel sessions

LangGraph, Multi-Agent & Orchestration

Master stateful workflows, build multi-agent systems, and coordinate AI teams for complex tasks.

Day 15 LangGraph Foundations — State, Nodes & Edges Intermediate
Objective: Learn LangGraph's core abstractions — State, Nodes, Edges, Conditional Edges, Checkpointing — and build your first graph-based workflow.
StateGraph Nodes Edges Conditional edges Checkpointing Reducers
Day 16 LangGraph Advanced — Agentic RAG, Reflection & Memory Advanced
Objective: Build agentic RAG with LangGraph. Learn reflection patterns, persistent memory, and self-correcting workflows.
Agentic RAG Reflection loops Persistent memory Self-RAG Adaptive RAG Query routing
Day 17 Practice — Build a LangGraph Workflow Hands-on
Objective: Build a complete LangGraph application — an agentic RAG system with query routing, retrieval, generation, and self-correction.
Suggested Exercises
  • Build an agentic RAG system with query routing (web search vs vector store)
  • Add a reflection step that grades document relevance and retries if needed
  • Implement checkpointing for conversation persistence
Day 18 Multi-Agent Systems — Architecture & Coordination Advanced
Objective: Understand multi-agent architectures: supervisor, swarm, hierarchical, and debate patterns. Learn when to use each pattern.
Supervisor pattern Swarm pattern Hierarchical agents CrewAI Inter-agent communication Task delegation
Day 19 Claude Code Agent Teams & Multi-Agent Orchestration Advanced
Objective: Learn Claude Code's Agent Teams, subagent patterns, and third-party orchestrators (Gas Town, Multiclaude). Understand when each approach is appropriate.
Agent Teams flag Team Lead + Shared Task List Contract-first planning File ownership Gas Town Multiclaude Token cost management
Day 20 Practice — Build a Multi-Agent System Hands-on
Objective: Build a multi-agent system using LangGraph supervisor pattern OR Claude Code Agent Teams. Real feature, real codebase.
Choose One Track
  • Track A (LangGraph): Build a supervisor + 3 specialist agents (researcher, writer, reviewer) for content generation
  • Track B (Claude Code): Enable Agent Teams flag, define a feature spec, and orchestrate frontend + backend + test agents on a real repo

Capstone Project Sprint

Choose a project from the bank below. 10 full days of building. Ship something real.

Day 21 Project Kickoff — Select, Scope & Architect
What to do: Pick a project from the Project Bank below. Break it down into a technical architecture. Write a CLAUDE.md. Set up your repo with Cursor rules. Define your stack and plan the build.
Days 22–29 Build Sprint — Daily Progress, Ship Incrementally
Daily rhythm: Build features using Cursor + Claude Code. Use TDD agentic loops. Integrate MCP servers where useful. Apply RAG, agents, or multi-agent patterns from the course. Push working code daily.
Build Guidelines
  • Use Claude Code subagents for parallel workstreams (frontend / backend / tests)
  • Apply the techniques you learned: RAG pipelines, tool-calling agents, LangGraph workflows
  • Keep a build log — what worked, what didn't, what you'd do differently
  • Ship incrementally: working MVP by Day 25, polish and edge cases from Day 26–29
Day 30 Demo Day — Present Your Build
What to do: Demo your project to the team. Walk through the architecture, the AI techniques used, what you learned, and what you'd improve. Share your CLAUDE.md and build log.

50+ Client-Style Project Briefs

Each brief is written as a non-technical client would describe it. Your job: translate the business need into a working AI-powered solution in 10 days.

Showing 55 projects

Learning & Project Assessment

This framework is designed to give engineers structured self-assessment to guide their growth, with formal evaluation conducted by their engineering manager or team lead. Ratings are qualitative — the goal is honest development feedback, not number-chasing. Evaluation outcomes feed directly into performance reviews and career progression conversations.

Part A

Learning Assessment (Days 1–20)

Evaluated at the end of Week 3, before the capstone begins. Covers both conceptual understanding and practical ability to apply what was learned. The engineer completes a self-assessment first, then the lead reviews and has a 1-on-1 conversation.

Category Needs Improvement Meets Expectations Exceeds Expectations
LLM Fundamentals & API Integration Cannot explain how LLMs generate text or has not built a working API integration. Struggles with basic concepts like tokenization, context windows, or message formatting. Can explain transformer basics, token economics, and model tradeoffs. Has built at least one working LLM API integration with proper error handling and streaming. Deeply understands model selection tradeoffs (cost, latency, capability). Builds production-grade integrations with structured outputs, retries, and fallback models. Can teach others.
Prompt Engineering Writes vague or unstructured prompts. Doesn't know the difference between zero-shot, few-shot, and chain-of-thought. Outputs are inconsistent. Applies appropriate prompting techniques for different tasks. Writes clear system prompts that produce consistent, well-structured output. Understands ReAct and when to use it. Designs prompt architectures for complex workflows. Systematically tests and iterates on prompts. Creates reusable prompt templates that the team adopts.
RAG & Retrieval Systems Cannot build a basic RAG pipeline. Doesn't understand embeddings, chunking strategies, or the difference between vector and keyword search. Builds end-to-end RAG pipelines: loading, chunking, embedding, storing, retrieving, generating. Understands tradeoffs between chunking strategies and can evaluate retrieval quality. Implements advanced patterns — hybrid search, reranking, self-RAG, or corrective RAG. Can diagnose retrieval failures and systematically improve pipeline performance.
Agents & Function Calling Cannot explain what an agent is or how function calling works. Has not built a working agent that uses tools. Builds agents that use multiple tools to complete tasks. Understands agent loops, tool schemas, and error recovery. Can implement ReAct or reflection patterns. Designs agent architectures for complex, multi-step workflows. Implements human-in-the-loop patterns, memory systems, and robust error handling. Agents handle edge cases gracefully.
LangGraph & Orchestration Cannot build a basic LangGraph workflow. Doesn't understand states, nodes, edges, or checkpointing. Builds stateful LangGraph workflows with conditional routing, tool nodes, and checkpointing. Can implement an agentic RAG system or a multi-step processing pipeline. Designs complex multi-agent LangGraph applications with shared state, reflection loops, and persistent memory. Understands when LangGraph is the right tool versus simpler alternatives.
Multi-Agent Systems Cannot explain the difference between single-agent and multi-agent patterns. Has not coordinated multiple agents on a task. Understands supervisor, swarm, and hierarchical patterns. Has built a working multi-agent system where agents coordinate on a shared task with clear role separation. Designs multi-agent architectures with sophisticated coordination — task delegation, conflict resolution, shared state management, and cost-aware orchestration. Can evaluate when multi-agent adds value versus overhead.
How to assess: The lead should have a 20–30 minute conversation with the engineer. Ask them to explain a concept (testing understanding), then walk through something they built (testing application). Look at their code from the practice days. The self-assessment should be completed by the engineer before this conversation — the lead's role is to calibrate, not just assign ratings from scratch.
Part B

Capstone Project Assessment (Days 21–30)

Evaluated after Demo Day. Covers four dimensions: technical quality, product thinking, development process, and communication. The engineer self-assesses first, the lead evaluates based on the live demo, code review, and Q&A.

Dimension Needs Improvement Meets Expectations Exceeds Expectations
Technical Quality Code is disorganized, hard to follow, or doesn't work reliably. AI techniques (RAG, agents, LangGraph) are applied superficially or incorrectly. Architecture is unclear or not thought through. Clean, well-structured code with a clear architecture. AI techniques are applied appropriately — the choice of RAG, agents, or multi-agent patterns makes sense for the problem. Error handling exists and the system is reasonably robust. Production-grade code quality. Thoughtful architectural decisions with clear reasoning. AI techniques are combined effectively — e.g., agentic RAG with reflection, multi-agent coordination with human-in-the-loop. Handles edge cases and failure modes.
Product Thinking Built something technically interesting but didn't address the client's actual problem. Requirements from the brief are ignored or misunderstood. The solution wouldn't work for real users. The solution clearly addresses the client's stated problem and requirements. Reasonable assumptions were made and documented. The system would be usable by the described client with minimal changes. Goes beyond the brief — anticipates edge cases the client didn't mention, identifies risks, and proposes a realistic roadmap for production deployment. Demonstrates genuine understanding of the business problem, not just the technical challenge.
Development Process No clear approach — jumped straight into coding without planning. Build log is empty or sparse. Didn't iterate — the final product looks like a first attempt. Didn't leverage the AI tools or course content effectively. Started with architecture and planning before building. Iterated meaningfully — there's evidence of testing, refining, and improving. Applied course concepts (TDD loop, CLAUDE.md, plan-then-build) during the build. Build log shows daily progress. Exemplary process — clear architecture document, disciplined iteration, meaningful use of Git history showing progression. Used advanced workflows (subagents, parallel sessions, MCP integrations) naturally and effectively. Build log is detailed and reflective.
Communication & Demo Demo is unstructured or confusing. Can't clearly explain what the system does or why design choices were made. Struggles to answer Q&A questions. Presentation runs over time or is unprepared. Clear, structured demo: problem → approach → live walkthrough → what was learned. Explains design decisions with reasoning. Handles Q&A competently. Stays within time. Compelling presentation that tells a story — from the client's problem through to the solution. Live demo is polished and handles unexpected inputs. Answers tough Q&A questions thoughtfully. Shares insights that help the whole team learn.
Part C

Demo Day Format

Day 30 is a live presentation to the full team. Each engineer presents their capstone project. The lead evaluates; peers ask questions.

Presentation Structure

Each engineer gets a fixed slot. The format is designed to be tight, focused, and useful for the whole team — not a lecture, but a live walkthrough.

Duration
15 minutes presentation + 10 minutes Q&A
Audience
Full engineering team + leads/managers
Segment 1
The Problem (2 min) — Read the client brief aloud. Explain what the client actually needs and any assumptions you made.
Segment 2
Architecture (3 min) — Walk through your technical design. What AI techniques did you use and why? What was the stack?
Segment 3
Live Demo (5 min) — Show it working. Use a real scenario from the client brief. Don't just show the happy path — show how it handles an edge case.
Segment 4
Reflections (3 min) — What worked, what didn't, what would you do differently with more time? What did you learn about AI development that surprised you?
Segment 5
Build Process (2 min) — How did you use the tools (Cursor, Claude Code, MCP)? Share one specific workflow or technique the team should steal.
Q&A
10 minutes — Open floor. Leads should ask at least one probing technical question and one product question.

What to Submit Before Demo Day

Each engineer should have these ready before their presentation. These form the basis for the lead's code review and evaluation.

Code
Git repository with clean commit history showing progression over the 10 days
CLAUDE.md
The project's CLAUDE.md file — shows how well they configured their AI workflow
Build Log
A daily log (even just bullet points) of what they built, what problems they hit, and what they learned
Self-Assessment
Completed self-assessment using the rubric above — both Part A (Learning) and Part B (Project)