How to become an AI engineer: a realistic six-month plan

Most six-month plans on the internet are reading lists. A reading list is not a plan, because it has no output. This one is organised around six things you build, and the theory arrives when you need it.

Weeks 1 to 5: get fluent, not clever

You need Python you do not have to think about, and enough SQL to pull your own data. This is the least glamorous stretch and the one people skip. Every later problem you hit — a broken join, a silent type coercion, a model trained on leaked data — traces back to a gap here.

Build: a command-line tool that reads a public dataset, cleans it, and answers three questions about it. No notebooks. Write it as modules with functions you can test.

Weeks 6 to 11: supervised learning end to end

Now the modelling. Regression, classification, ensembles. Implement linear regression and a decision tree by hand once, then use the library forever after — one pass is enough to stop treating .fit() as magic.

The part that matters more than the algorithm is validation. If you cannot explain why your cross-validation split is honest, your accuracy number is decoration.

Build: a classifier on a messy, imbalanced dataset, with a written justification of the threshold you picked and what it costs the business when it is wrong.

Weeks 12 to 15: structure without labels

Clustering, dimensionality reduction, and your first neural network. The mental shift here is that you no longer have a score to optimise toward, so you have to argue that your segments mean something.

Build: a segmentation of a real customer or product dataset, presented as five slides a non-technical person could act on.

Weeks 16 to 22: pick a depth

You cannot be deep in everything by month six. Choose one:

• Language. Embeddings, transformers, retrieval, evaluation of generated text.
• Vision. Convolutional networks, augmentation, transfer learning.
• Scale and time. Spark, pipelines, forecasting.

Build: one substantial project in your chosen direction, deployed behind an API with logging. Deployment is not optional. A model nobody can call has not been finished, and interviewers know the difference immediately.

Weeks 23 to 26: convert it into an offer

Interview preparation is a skill separate from the work. ML case studies have a structure, SQL rounds have recurring patterns, and behavioural rounds reward stories with a measured result attached.

Build: a portfolio page for each project — problem, approach, trade-off you made, result — plus twelve mock interviews. Twelve, not two. The improvement curve is steep and it is entirely practice.

What does not work

Collecting certificates. Nobody in a hiring loop counts them. They open your GitHub.

Tutorials on cleaned data. Every dataset in a tutorial has been laundered. The job is mostly the laundering.

Only prompting. Prompting is a useful skill sitting on top of an engineering discipline. On its own it is not a role, and the market has already worked that out.

Learning alone. The single biggest predictor of finishing is having someone expecting your work on Thursday.

What to do this week

Pick the dataset for your week-one project. Something you actually care about — your city's transport data, match results, energy prices. Curiosity is the fuel that gets you through week nine, when the novelty has gone and the gradient still is not descending.