Lock in your room for 3 month to become AI Engineer (step-by-step plan below)

Eight months ago I dropped $2,400 on a 12-week AI bootcamp, kept a $59/month subscription to a "prompt mastery" platform running for four months straight, and bought a $349 lifetime deal for a course that turned out to be 80% outdated slides

I had a folder full of PDFs, three half-finished certificates, and zero projects I could actually show anyone

Then I found something that changed everything: the companies that actually build AI, Google, Anthropic, OpenAI, had started giving away their training for free. Not watered-down intro videos. Full courses with certificates. Meanwhile, GitHub had repositories with 95,000+ stars that taught better than any bootcamp I'd paid for

I cancelled the subscription, ate the bootcamp loss as a lesson, and rebuilt my entire learning path from scratch. Built an AI agent that manages my morning routine. And it cost me $0

This is the exact system, step by step, from environment setup to a paying AI business on the other side. Follow it in order

The Three Rules Before You Start

Don't skip ahead. If you jump to LLMs without understanding gradients, you'll be copying code you don't understand

Take notes in Obsidian, free and local. After every session write down what you learned, what surprised you, and what's still unclear. This is non-negotiable

Build at every step. Each phase ends with a checkpoint. If you can't complete it, go back before moving forward

Step 1: Environment Setup, Day One

This is the one evening you're not allowed to overthink. Install Python 3.11 or later from python.org, checking "Add to PATH" during install so your terminal can find it later. Install VS Code with the Python extension, this becomes your workspace for everything from here on

Set up Git and a GitHub account, this is where every project you build gets pushed and where employers will eventually look. Install Obsidian and build the folder structure below inside it, this vault becomes your second brain for the entire roadmap

Install Ollama now even though you won't touch it until Step 4, this lets you run open models locally on your own machine with zero API costs

Then create free accounts on four platforms. Anthropic Academy gives you 16 free courses with real certificates, arguably the most underrated AI learning platform in 2026. OpenAI Academy offers free workshops and an AI Foundations course straight from the team that built ChatGPT. Google AI hosts the Google AI Professional Certificate, seven modules, free through Coursera audit mode. And Coursera itself, when you look for the small "Audit this course" link at checkout instead of paying, unlocks every video and material for free, the only thing missing is the Coursera-branded certificate, which doesn't matter because you'll walk away with certificates directly from Anthropic, OpenAI, and Google instead

Checkpoint: Python, VS Code, and Ollama installed. GitHub account created. Obsidian vault ready. Accounts live on Anthropic Academy, OpenAI Academy, Google AI, and Coursera

Step 2: AI Fundamentals, Weeks 1-2

A 2025 WEF analysis found AI-literate workers command a 15-22% salary premium over non-literate peers. This step alone puts you ahead of 90% of applicants before you write a single line of code

Week one is about vocabulary, not code. Start with the Google AI Professional Certificate, modules 1 through 3, the gentlest possible on-ramp covering what AI actually is and how to brainstorm and research with it. Then move to Anthropic Academy's AI Fluency: Framework & Foundations course, built around the 4D AI Fluency Framework and co-developed with university professors. It takes 2-3 hours and the certificate carries real weight on LinkedIn because it comes from Anthropic, the company behind Claude, not a generic MOOC platform

Week two is your first code. Fork microsoft/generative-ai-for-beginners, a repository sitting at 95,000+ stars with 21 full lessons, and work through lessons 1 through 6: what generative AI actually is, how large language models work under the hood, how to write your first prompts, and how to build your first chat app. After every session, log what you learned, what surprised you, and what's still unclear in your Obsidian vault, using terms like LLM, transformer, and token as your running glossary

Checkpoint: you can explain LLMs, tokens, and transformers in your own words without looking anything up. Your first Jupyter notebooks run. Your Obsidian vault has 4-6 real notes in it

Step 3: ML Foundations, Weeks 3-5

Companies pay $150K+ for engineers who understand why a model underperforms, not just how to call an API endpoint. This is the gap between someone who copies tutorials and someone who actually debugs models, and it's the step most beginners try to skip

Your primary track is microsoft/ML-For-Beginners, a 44,900+ star repository built as a full 12-week curriculum covering regression, classification, clustering, and NLP basics, complete with quizzes and challenge notebooks. Compress it to three weeks by doing two lessons a day instead of the intended pace

Run IBM Machine Learning on Coursera in parallel, in audit mode so it's free, this is a more traditional video format and gives you a second angle on the exact same concepts, which dramatically improves retention compared to studying one source alone

Keep mlabonne/llm-course open as a math reference throughout, a 40,000+ star repository whose first section covers only the linear algebra, calculus, and probability that's actually relevant to machine learning, no wasted academic detours

By week five, pick a dataset directly from the Microsoft repo and build your own classification model completely from scratch, then push the finished project to GitHub

Checkpoint: you understand regression, classification, clustering, gradient descent, loss functions, and overfitting well enough to explain them to someone else. You've trained a model on real data. You have one live project on GitHub

Step 4: Deep Learning & Neural Networks, Weeks 6-8

Your primary resource here is karpathy/nn-zero-to-hero, both the video lectures and the accompanying code repository, built by Andrej Karpathy, former Director of AI at Tesla and an OpenAI co-founder. He builds neural networks from absolute zero, no frameworks, just raw Python and math, and you build alongside him: micrograd, makemore, and eventually nanoGPT

Week six covers lectures 1 through 3, micrograd and makemore. Code along in real time, pause constantly, type every single line yourself instead of copy-pasting, run it, then deliberately break it to see what happens. Week seven is dense, lectures 4 and 5 on activations, BatchNorm, and backpropagation, one lecture per day with detailed notes since this is the most conceptually loaded stretch of the entire roadmap. Week eight is the payoff, lectures 6 and 7, where you build a GPT-style transformer from scratch and learn tokenization

Run a parallel experiment while you build nanoGPT: open a second terminal and run ollama run llama3.2:3b, then compare your own "toy" model's output side by side with a real 3-billion-parameter model. This is genuinely eye-opening, it's the moment theory ("I understand transformers") becomes felt experience ("I can see exactly what scale actually buys you")

Supplement with microsoft/AI-For-Beginners for weeks 7-12 content on CNNs and RNNs, which expands past Karpathy's curriculum, especially useful if you want computer vision exposure. Then bridge to production with Anthropic Academy's Building with the Claude API course, now that you understand what's happening inside a model, you learn how to actually call one via API, covering auth, system prompts, tool use, and streaming

Checkpoint: you've built a neural network entirely from scratch. You understand backprop, attention, and transformers well enough to explain how GPT works to a non-technical friend. You can run models locally with Ollama. You know the Claude API

Step 5: LLMs & Prompt Engineering, Weeks 9-10

mlabonne/llm-course's LLM Scientist track is the single most comprehensive free LLM curriculum available, with a Colab notebook for every topic. It walks through LLM architecture, which connects directly back to what you built with Karpathy, fine-tuning techniques like LoRA and QLoRA for customizing models to specific tasks, quantization for running models locally which ties into your existing Ollama setup, and evaluation methods for actually measuring whether a model is good or just looks good

On the prompting side, OpenAI Academy's "Intro to Prompt Engineering" and "ChatGPT for any role" come straight from the team that built ChatGPT. Anthropic's official prompt engineering documentation is arguably the best-written guide of its kind on the internet, not a course, a deeply detailed reference you'll keep returning to

Go back to microsoft/generative-ai-for-beginners and finish lessons 7 through 21, which you skipped earlier. With the deep knowledge you now have, advanced topics like RAG, function calling, design patterns, and fine-tuning finally click in a way they wouldn't have on day one

Your week 10 project: build a RAG system directly over your own Obsidian vault using ChromaDB or LanceDB, both free and fully local. You end up with a tool that answers questions about everything you've studied so far, literally a second brain built on top of your second brain. Push it to GitHub

Step 6: AI Agents, Weeks 11-12

microsoft/ai-agents-for-beginners covers 12 full lessons on tool use, memory, multi-agent systems, and orchestration. Go deeper with Anthropic Academy's MCP courses, "Introduction to Model Context Protocol" and "MCP: Advanced Topics", where MCP is Anthropic's open standard for connecting AI to external tools and has become the 2026 standard for agent tool-use. These courses teach you to build both MCP servers and clients from scratch

Spend two to three sessions with LangGraph, LangChain's framework for building stateful, multi-step agent workflows, in free Colab notebooks. It complements Anthropic's MCP approach perfectly: LangGraph handles orchestration, MCP handles tool connections. Study the Anthropic Cookbook as a bonus resource, it contains the best real-world examples of tool use and MCP patterns available anywhere

Your final agent project: build something that uses MCP and Claude to work with your own local files, for example an agent that reads your Obsidian vault, checks the web for updates on topics you're studying, and generates a daily summary sent straight to your Telegram

Checkpoint: you've built a working AI agent with MCP. You understand agent architecture, tool use, and multi-step workflows well enough to design your own from scratch. Your portfolio keeps growing

Step 7: Production, Portfolio & Responsible AI, Weeks 13-14

Deploy your best project entirely for free. Gradio paired with Hugging Face Spaces is the fastest way to share an ML demo with free hosting. Streamlit Community Cloud works well for data-focused apps on its free tier. Vercel handles web-based AI tools on its free tier

A deployed model without evaluation is a liability, not a product. Learn DeepEval, an open-source framework for LLM evaluation, RAGAS specifically for evaluating your RAG pipelines from Step 5, and LLM-as-Judge, where you use one LLM, Claude works excellently here, to evaluate another's outputs

This is where 90% of free guides fail completely, they teach you to build but never to build responsibly. Study Constitutional AI to understand how modern models are actually aligned, Anthropic's core approach to safety. Learn prompt injection defense to protect your apps from adversarial inputs, and red-teaming to stress-test your own systems before real users ever get the chance to break them

Your GitHub profile is your resume in this field. Write a professional profile README plus project READMEs with architecture diagrams and live demo links. Draft two to three short LinkedIn case studies covering what problem you solved, what you built, and what you learned. The career ladder from here runs Junior AI Engineer at $80-120K, up to Prompt or Agent Engineer at $120-180K, up to AI Product Engineer at $150-250K

Your capstone project is a production-grade AI agent solving a real problem in your own life, deployed, evaluated, and safety-checked. This is what you show employers. This is what you post about. This is the proof

Maintenance Mode: Staying Current After You Finish

AI moves fast enough that the roadmap doesn't end at week 14, it just changes shape. Monday, spend 10 minutes checking Anthropic, OpenAI, and Google release notes. Wednesday, spend 15 minutes browsing arxiv-sanity-lite and read one abstract that catches your eye. Friday, spend 20 minutes on a Yannic Kilcher or 1littlecoder video covering a new paper or tool. Once a month, build one small project with something new and push it to GitHub. Total time investment is roughly an hour a week, and it's enough to keep you in the top 10% of practicing AI engineers

What Comes After The Roadmap: Turning The Skill Into Income

Finishing the 14-week roadmap gives you the technical floor. What actually turns that into money is a delivery system, and this is where the "Start a 1-Person Business with Claude" framework picks up exactly where the roadmap leaves off

Roughly 84% of the planet has never used AI once, not a single prompt, around 16% touched a free chatbot and stopped there, about 0.3% pay for a real model subscription, and a rounding error of humanity knows tools like Claude Code even exist. Finishing the roadmap above already puts you ahead of nearly everyone, which means the businesses you'll sell to, roofers, dentists, law firms, e-commerce shops, experience what you build as pure magic, and magic is billable

The business model runs on three levels, and everyone starts at level one with no exceptions. Level one is AI-built websites, sold not to the 30% of small businesses with no website at all, who are the hardest sell in the market, but to the other 70% running bad, outdated, non-converting websites, who are proven buyers because they already paid for one once. A typical first sale runs around $500, rising to $3-10k once you have proof and testimonials

Level two is automations and agents: invoicing, lead follow-up, onboarding, review collection, and support, sold into niches that are visibly bleeding money from slow response times, home services, dental clinics, real estate, law firms, and solar contractors. The same four automations, instant follow-up, review collection, appointment reminders, and dead-lead reactivation, get rebuilt and resold across every one of these niches. Typical tickets run $1-5k for setup plus $200-500 a month recurring

Level three bundles 5-25 automations behind a single client dashboard, turning the whole thing into a product rather than a gig, typically priced at $10-20k per project or $2-5k a month in retainers

The entire operation runs out of one Claude Code folder acting as an operating system, with skills organized like departments: engineering, design, sales, delivery, and operations, backed up to a private GitHub repo so the business survives even if your laptop doesn't. Client acquisition runs on two parallel machines, a personal brand built on proof-first content, building and posting real demos before anyone asks for them, and a cold email engine run through Instantly, Apollo, and Claude connected via MCP to draft and manage entire campaigns autonomously

Closing follows what the framework calls the Doctor Approach: never give price before a call, ask questions until the client states their own pain and prices their own solution out loud, then state one clean number and stay silent. The proof-of-concept close, payment upfront with a full satisfaction guarantee, removes essentially all of the client's risk and consistently raises close rates

Also check this article which can improve your knowledge in loop engineering:

https://x.com/keloneoneal/status/2074484119845605379

Book, like, quote this article and send it to your close friend who want to start by still don't moving in this term

Links

• python.org
• x.com/keloneoneal/status/2074484119845605379