A 24-week self-study curriculum · fully free

From Python basics to a
shipped LLM agent.
In 24 weeks.

5–8 hours a week, nights and weekends. The best free resources on the internet, sequenced in the order they actually make sense, with a knowledge graph that connects every concept. Built by someone learning this alongside you.

Start week 0 → Why this roadmap ↓
24
weeks
132
tasks
146
curated links
21
projects to ship
$0
total cost
00 / WHY THIS EXISTS

I built this for myself, first.

I'm Haran, and I'm rebuilding my AI engineering skills from the ground up — not because I have to, but because I want to understand the whole stack, not just the parts I use at work.

I started the usual way: a few Coursera courses, a half-finished book, a bookmarks folder with 200 links I'd never click again. I realized the problem wasn't missing resources — the best stuff in AI is free. Karpathy's videos. 3Blue1Brown. Lilian Weng's essays. The original transformer paper. It's all out there.

The problem was sequencing. I didn't know what to learn first, what to skip, or when I was ready to move on. So I designed the path I wish I'd had when I started: 24 weeks, 10 phases, every task linked to the right free resource, everything cross-referenced in a knowledge graph I could open in Obsidian.

I'm sharing it with a small group of friends first to see if it's actually useful. If you're reading this, you're one of them — thanks for trying it out. I'd love your honest feedback: what worked, what didn't, what's missing, what felt like a waste of time.

— Haran · haran.prithvii@gmail.com

01 / WHY THIS ROADMAP

There are 50 AI roadmaps on GitHub. Here's what's different.

01
Sequenced for the part-time learner.
Most roadmaps assume you're studying full-time. This one is built for 5–8 hours a week, nights and weekends. Each week has a realistic scope — you won't fall behind, you won't burn out.
02
Every link is vetted, not just listed.
146 resources, all free, all checked. Karpathy's Zero-to-Hero, the Attention paper, Lilian Weng's surveys, Hamel's evals essay — the stuff that actually teaches you, not just the stuff that shows up first on Google.
03
Concepts are linked, not siloed.
175+ concept notes in Obsidian, all cross-referenced. When you're reading about attention, you can jump to the tokens it acts on, the backprop that trained it, the LoRA that fine-tunes it. One keyboard shortcut shows the whole graph.
04
You ship, you don't just study.
21 projects across the 24 weeks — from a scratch-built MLP to a production RAG agent with evals. By week 24 you have a GitHub full of real work, not just completion certificates.
05
Built to stay current.
The tracker has a curated sidebar of the sources that working AI engineers actually read — Raschka, Interconnects, HF Daily Papers, Artificial Analysis, Simon Willison. One click, their latest content in a new tab.
06
Honest about what it isn't.
This won't make you a research scientist. It won't replace a PhD. What it will do: get you to where you can read a paper, build with an LLM, and hold your own in a technical AI interview.
02 / ROADMAP

From linear algebra to shipped agent.

PHASE 00
Toolchain & Mindset
WEEK 0
Python, uv, VS Code, Git, Obsidian, Docker. Zero-friction setup.
PHASE 01
Python & CS
WEEKS 1–2
CS50P, idiomatic Python, NeetCode, Git workflow. The floor you build on.
PHASE 02
Math for ML
WEEKS 3–4
Linear algebra, calculus, probability, information theory. Enough to read papers.
PHASE 03
Classical ML
WEEKS 5–6
Regression, trees, SVMs, clustering. Full lifecycle on tabular data.
PHASE 04
Deep Learning
WEEKS 7–9
Karpathy's Zero-to-Hero. Micrograd → PyTorch → CNNs & RNNs.
PHASE 05
Transformers & LLMs
WEEKS 10–12
Attention from scratch. Build nanoGPT. Tokenization. Pre- and post-training.
PHASE 06
RAG Engineering
WEEKS 13–15
Chunking, embeddings, HNSW, hybrid search, reranking, HyDE, ColBERT.
PHASE 07
Eval & Production
WEEKS 16–17
Golden sets, RAGAS, LLM-as-judge, FastAPI, Redis, Langfuse, Docker.
PHASE 08
Agents & MCP
WEEKS 18–19
ReAct, tool use, MCP servers, CRAG, prompt injection defense.
PHASE 09
Fine-Tune & Research
WEEKS 20–22
LoRA, QLoRA, quantization, Ollama, vLLM. Reading papers on schedule.
PHASE 10
Capstone
WEEKS 23–24
Ship one thing so good it speaks for you at interviews.
NEXT
Get hired.
WEEK 25 →
AI Engineer. AI Product Engineer. ML Researcher. Portfolio does the talking.
03 / WHAT'S INSIDE

Four things. All free.

1
Sequenced curriculum
Every phase, every week. Every resource curated and linked — Karpathy videos, 3Blue1Brown, ISLP, the original papers. No filler.
2
Obsidian knowledge graph
175+ concept notes with wikilinks. Open in Obsidian (free), hit Cmd+G, and watch the entire field of AI engineering appear as an interactive graph.
3
Interactive progress tracker
A standalone HTML page (this site!) that persists your progress in localStorage. 210+ tasks, 11 phase progress bars. Nothing to install.
4
21 shipped projects
Every phase has builds. House-price predictor → MLP on MNIST → nanoGPT → production RAG → CRAG agent → LoRA fine-tune → capstone. Your portfolio emerges.

Most AI "bootcamps" teach you a framework in six weeks.

This roadmap teaches you the field in six months — from vectors to agents — with every hour of your time
pointing at the highest-leverage material on the internet.

04 / GOLDEN RULES

Four rules. Non-negotiable.

  1. Read the theory, implement it from scratch, use the library, ship a project — every topic.
  2. If you can't explain it, you don't understand it. Write a note.
  3. Papers before libraries. The arXiv link loads faster than the docs.
  4. One project per phase, minimum. Your GitHub is your résumé.
05 / KNOWLEDGE GRAPH

Every concept, connected.

The repo ships with a pre-built Obsidian vault — 175+ concept notes, all cross-linked with [[wikilinks]]. Open it in Obsidian and hit one keyboard shortcut to see the entire field of AI engineering as an interactive graph.

  1. 01
    Download Obsidian
    Free. macOS / Windows / Linux. obsidian.md ↗
  2. 02
    Open folder as vault
    Point it at obsidian-vault/ inside the cloned repo.
  3. 03
    Press ⌘+G (or Ctrl+G)
    Graph view opens. Every concept is a node. Every [[link]] is an edge. Click any node to jump to the note.
Get Obsidian ↗
06 / HOW IT WORKS

One platform: study, track, stay current.

The tracker isn't just a to-do list — it's a study environment. Every task has the right resource linked inside it. No tab-switching, no bookmarks folder, no hunting. Sign in once and your progress syncs everywhere.

01 THE ROADMAP
132 tasks, sequenced across 24 weeks.
Ten phases, from "install Python" to "ship a production agent." Click any task to expand it — the video, paper, or doc you need is right there. Check it off when you're done. Your progress bar fills as the phase completes.
02 THE LIBRARY
Every resource, in one place.
Sometimes you want to browse rather than follow the roadmap. The Library tab flattens all 146 resources into one grid, grouped by type: videos, papers, courses, essays, and build tools.
03 THE PROJECTS
21 things you'll actually build.
From a scratch-built MLP on MNIST to a production RAG system with evals. Each project has the tools pre-linked — Kaggle datasets, HF models, the right libraries. You don't waste time on setup.
04 STAY CURRENT
Curated sources, always in view.
The right side of the tracker is a launchpad to the best AI sources: HF Daily Papers, Sebastian Raschka, Nathan Lambert, Lilian Weng, Artificial Analysis, LMSys Arena. Each link opens the live site in a new tab — bookmark the ones you'll actually read.
07 / WHO IT'S FOR

Is this actually for you?

You'll get value from this if…
  • You're a working engineer or CS student who wants to move into AI.
  • You've touched Python before but are new to ML/DL fundamentals.
  • You have 5–8 hours a week, nights and weekends, consistently.
  • You learn better from doing than from watching — you want to ship code.
  • You want to understand how things work, not just call APIs.
  • You're okay reading a paper every now and then.
You should probably skip this if…
  • You're looking for a quick "build an AI app this weekend" shortcut.
  • You already work as an ML engineer — this will be too basic for weeks 1–10.
  • You want a credential or certificate at the end. (There isn't one.)
  • You prefer structured, synchronous, taught-by-a-human learning.
  • You want to do research. This is an engineering path, not an ML research path.
  • You can't commit to at least 3 months. The value compounds — week 1 alone won't change your life.
Still unsure? Open the tracker, glance at Phase 0 and Phase 5 (Transformers & LLMs). If the Phase 5 resources look interesting but above your current level, this roadmap is probably right for you — Phase 0 starts where you are, and the path is designed to get you to Phase 5 and beyond.
Start today.
What else were you going to do?
Open the tracker →