An opinionated roadmap · revised May 2026

From Python day one to shipping AI in production. Five stages, in honest order.

Most "AI roadmaps" online are dumps of links. They tell you what exists, not what to do. This is the opposite. Five stages. Time budget per stage. The three projects that prove you cleared it. And the one trap at each stage that wastes most learners a year. Written by the people who have walked 11,000+ Modern Age Coders students through it.

5 stagesFrom beginner to shipping
14–24 monthsHonest end-to-end time
3 projectsPer stage, that prove fluency
1 trapPer stage. Marked clearly.
What this roadmap is not

This is not a YouTube playlist disguised as a career plan.

We have watched students lose entire years to "AI roadmaps" that are sixty links to free courses. The links are not the problem. The problem is they are unranked, untimed, and unsequenced. A real path through this material is ruthless about ordering and ruthless about cutting.

This page picks one path. You will not learn everything in AI by following it. You will learn the right things, in the right order, for the goal of becoming a working AI engineer in 14–24 months. Other goals (research, mathematics depth, MLOps specialisation) need other paths. We say so when it matters.

Every stage has a clear "you are done with this when..." signal. No checkbox. A concrete behavioural milestone — the day you sit down and the work just flows. That is when you graduate to the next stage. Not earlier.

The five stages

One stage at a time. In order. Don't skip.

The single most common mistake we see — and we see it monthly — is students trying to do Stage 4 (deep learning) before they have actually finished Stage 1 (Python fluency). The result is six months of confused frustration. The order below is non-negotiable. Each stage compounds on the previous one.

I
Stage 01 · weeks 1 — 12

Python fluency, not Python literacy.

Time budget · 3 months · 8 hours/week · ~96 hours total

The first stage is unglamorous and the most important. Almost everyone underestimates it. You are not done with Python when you can write a for-loop. You are done with Python when reading other people's code feels easy, when writing a 200-line script feels small, and when debugging is a five-minute habit rather than an hour-long ordeal.

This stage uses no AI at all. No PyTorch. No LLM API. Just Python — language, ecosystem, command line, git. We are building the keyboard fluency that every later stage takes for granted.

What you will learn

  • Core language — data types, control flow, functions, classes
  • Standard library essentials — pathlib, json, datetime, collections
  • Virtual environments, pip, requirements.txt
  • Git, GitHub, basic command line
  • Reading and debugging error tracebacks
  • Writing tests with pytest

Three milestone projects

  • A CLI tool that does something real — a budget tracker, a habit logger
  • A small Flask or FastAPI web app with a real route
  • Read and modify someone else's open-source Python repo
Trap to avoid Jumping into ML libraries before this stage is done. If you can not write a working Python function from scratch without Googling, you are not ready for Stage 2. Six months of "learning numpy" will not patch over weak Python. Finish Stage 1 first. The students who do this finish the full roadmap on time.
II
Stage 02 · weeks 13 — 28

Mathematics, the working amount.

Time budget · 4 months · 6 hours/week · ~96 hours total

You do not need a maths degree to be an AI engineer. You need a precise, small subset of mathematics, learned well, so the libraries and papers later in this roadmap stop being opaque. We will tell you the exact subset and let you skip the rest. Three subjects. No more.

Linear algebra (vectors, matrices, dot products, projections). Calculus (derivatives, the chain rule, gradient descent). Probability (distributions, expectation, Bayes' rule). That is it. Twenty hours per subject. Sixty hours total. Then a few weeks consolidating with code.

What you will learn

  • Linear algebra — matrix multiplication, dot product, eigenvalues at intuition level
  • Calculus — derivatives, chain rule, gradient descent
  • Probability — distributions, expectation, variance, Bayes' rule
  • How to translate maths notation into numpy code
  • Just enough statistics to read a confusion matrix

Three milestone projects

  • Implement matrix multiplication in pure numpy. Compare with numpy's own.
  • Hand-code gradient descent on a one-parameter function. Plot the loss curve.
  • Build a small Monte Carlo simulation of a real problem — Plinko, dice odds, an election poll.
Trap to avoid Buying a "machine learning maths" textbook and trying to read it cover-to-cover. You will lose three months and gain little. The right approach is to learn maths in tight loops alongside numpy code — every concept landed via implementation. Khan Academy and 3Blue1Brown are excellent first passes. The MAC Mathematics for AI block is built exactly around this loop.
III
Stage 03 · weeks 29 — 44

Classical machine learning, not yet deep learning.

Time budget · 4 months · 8 hours/week · ~128 hours total

Before neural networks. Yes, before. Linear regression. Logistic regression. Decision trees. Random forests. K-means clustering. The same arc Andrew Ng's old course took you through, but tighter and with modern tooling. scikit-learn is your home for this stage.

You are learning two things in parallel: classical ML techniques (which still solve 70% of real industry problems) and the surrounding workflow — data cleaning, splits, cross-validation, metrics, the bias-variance tradeoff. The workflow is more important than any individual algorithm.

What you will learn

  • Pandas, numpy, matplotlib — the trio of working data scientists
  • Regression, classification, clustering — the algorithm families
  • Cross-validation, train/val/test splits, common pitfalls
  • Metrics — accuracy, precision, recall, F1, ROC
  • Feature engineering and feature selection
  • Hyperparameter tuning and the curse of dimensionality

Three milestone projects

  • Solve a Kaggle "Titanic" or "Housing Prices" problem from scratch, end-to-end, with honest cross-validation
  • Build a recommendation system on a public dataset — books, movies, anything
  • Deploy one of these as a small FastAPI service. The first time your model talks to the world.
Trap to avoid Skipping classical ML because "everyone is doing deep learning now." Classical ML is what you will actually use in your first job. It is also where you learn the discipline of honest evaluation. The students who skip Stage 3 are the ones who later cannot tell the difference between a working model and a leaking one.
IV
Stage 04 · weeks 45 — 64

Deep learning, built up from scratch.

Time budget · 5 months · 10 hours/week · ~200 hours total

Now the neural networks. Now PyTorch. Now the architectures that power every modern AI system. The right way to learn this is from the bottom up — perceptron, multilayer perceptron, backpropagation by hand, then CNN, then RNN, then transformer. Our Inside Neural Networks pillar page is the conceptual companion to this stage.

The middle of this stage is where you will build your first real model from scratch — a small image classifier, a tiny language model. The end of this stage is where you can read a deep learning paper and roughly understand what it is saying. That is a non-trivial milestone.

What you will learn

  • PyTorch — tensors, autograd, nn.Module, training loops
  • Perceptron through multilayer perceptron through backpropagation
  • Convolutional networks for images
  • Recurrent networks (briefly) then transformers
  • Loss functions, optimisers (SGD, Adam, AdamW)
  • Regularisation, batch normalisation, dropout
  • How to use a GPU. How to debug an OOM error.

Three milestone projects

  • Train an MNIST classifier from scratch — the deep learning equivalent of "Hello, World"
  • Build a small image classifier on a non-trivial dataset (CIFAR, custom)
  • Implement a tiny transformer language model — see our build-your-own-GPT walkthrough
Trap to avoid Watching too many tutorial videos. Deep learning intuition does not transfer from watching someone else write code. Type every line yourself. Break it. Fix it. Read the error. The students who learn this stage by typing are six months ahead of the students who learn by watching, every time.
V
Stage 05 · weeks 65 — 96

LLMs, agents, and shipping AI in production.

Time budget · 6–8 months · 10 hours/week · ~280 hours total

The home stretch. You can now read papers, train small models, deploy services, and write production Python. Stage 5 brings these together into the actual job: building AI products. RAG systems. Fine-tuned models. Agents that do work. Evaluation pipelines. The full LLM application stack.

This is also the longest stage and the most open-ended. The field is moving fast. The skill at this stage is not what you know — it is how quickly you can learn the next thing. The previous four stages were about building that learning muscle on a stable foundation. Now you use it.

What you will learn

  • The full LLM application stack — APIs, prompts, RAG, fine-tuning
  • Embeddings and vector databases
  • Agent architectures — ReAct, planning, memory, tool use (see our agents pillar)
  • Evaluation — how to know your AI app is actually working
  • Latency, cost, caching — production realities
  • Reading and reproducing the latest papers

Three milestone projects

  • Build a RAG application on a real corpus — your company's docs, a textbook, your own notes
  • Fine-tune a small open-weight model on a domain-specific task
  • Build and deploy a multi-step agent that genuinely saves you an hour a week
Trap to avoid Treating LLM APIs as a substitute for the previous stages. Many "AI engineers" of 2025 know how to call OpenAI's API but cannot debug a tensor dimension mismatch. Their first hard problem ends them. The students who climb to this stage on top of Stages 1–4 are the ones who become senior in two years rather than six.
The students who finish this roadmap are not the ones who studied the hardest. They are the ones who refused to skip a stage.
— What we have learned from teaching this for six years
Modern Age Coders cohorts · stage-mapped

The MAC courses that walk you through this stage by stage.

Each course below maps to one or two stages of the roadmap. You can self-study every stage if you have the time and discipline; or you can join a live small-batch cohort that runs alongside other learners with a mentor checking your work weekly. Both work. The cohort path finishes 4–6 months faster on average.

Place yourself on the roadmap · free 25-min call

Tell us where you are. We will tell you what to do next.

The hardest part of any roadmap is honestly figuring out which stage you are actually on. Most learners overestimate by one stage. A senior MAC mentor will spend 25 minutes with you, sample your code, and tell you the truth — and the exact next three months of work. Free. No sales call.

Place me on the roadmap

A senior MAC mentor reads each form. Reply within 24 hours.
Honest follow-ups

The questions a serious learner asks next.

I am 14 — is this roadmap really for me?

Yes. We have students who started Stage 1 at age 12 and were on Stage 4 by age 16. The roadmap was designed to be age-agnostic above 12; the only thing that changes with age is the weekly time budget and the maturity of the projects. The teen courses listed above are this roadmap, teen-paced.

I am 45, mid-career, no coding background. Should I attempt this?

Yes — with one adjustment. Extend each stage's time budget by 30%. The cognitive part is no harder; what changes is the hours-per-week you can sustain alongside a job. We have placed adults on Stage 5 within 18 months. The full path is 22–28 months for most career-changers. The Coding Comeback for Women programme is built around this exact pacing.

Do I really need to do Stage 3 (classical ML) before Stage 4?

Yes — and this is the single most common skip. Classical ML is where you learn honest evaluation, the bias-variance tradeoff, and what overfitting actually feels like. Without that intuition, your deep learning models will silently leak and you won't know it. Two weeks of classical ML saves you six months of confused debugging in Stage 4. We will fight you on this in the orientation call. Listen to us.

Can I parallelise stages?

Sparingly. Stages 2 and 3 can run in parallel (maths-with-numpy, ML-with-sklearn) and that works well. Trying to run Stage 1 alongside Stage 4 is the disaster pattern — you spend three months looking up syntax instead of learning deep learning. The general rule: the next stage requires the previous one to feel automatic, not just possible.

What if I want to do research, not engineering?

This roadmap is built for engineering. For research, swap Stage 5 for a research path — read papers daily, do reproduction studies, pick a sub-field early, find a mentor in academia. Stages 1–4 are identical. The branch happens at Stage 5.

How much can I expect to earn after Stage 5?

Honestly variable. In India, a Stage-5 finisher with one shipped project commands ₹8L–₹18L for a first AI engineer role; the ceiling is much higher and rising. Globally — US, EU, UK — entry-level AI engineering ranges from $90k to $180k. We are quoting honest median ranges from MAC alumni 2023–2025. The frontier salaries you read about in the news are outliers, not the median.

What about JEE / college simultaneously?

If you are a Class 11-12 JEE aspirant, follow our JEE coding track instead. It compresses Stages 1 and partial-2 into a JEE-friendly 3 hours/week and resumes the full roadmap after JEE Advanced. For college students, the full roadmap fits inside a four-year degree if you give it ten hours a week.

Where do the MAC courses fit if I am self-studying?

You can self-study every stage. The courses give you three things you cannot get alone: a fixed weekly cadence so you do not drift, a mentor who reads your code, and a cohort of peers at the same stage. Roughly 70% of MAC's AI students said in our 2025 survey they would have dropped off without the cohort. Self-study is harder; not impossible. Pick honestly.

What should I read alongside this page?

Three things. How AI Actually Works for the conceptual map. Build Your Own GPT in Python for the Stage 4 capstone. Inside Neural Networks for the deep-learning depth that Stage 4 needs. And, in print, Andrej Karpathy's blog plus Sebastian Raschka's "Build a Large Language Model From Scratch" book.

Ask Misti AI
Chat with us