Roadmap¶
Nine months, three trimesters, fifteen milestones. This page is the whole programme in one place: not the syllabus, but what you will be able to do at each point, and what you will have to show for it.
1 Data Analyst Months 1 to 3
You arrive able to write some Python. You leave able to take a messy file that nobody in the room understands and produce an answer you can defend.
1. Set up and ship. A working Linux environment, notebooks, and version control. Your first project pushed to GitHub in week one, because the habit matters more than the code in it.
2. Read the maths. Linear algebra and probability treated as the language models are written in. Vectors, matrices, distributions, uncertainty. Enough to read a formula and know what it is doing, not enough to sit an exam.
3. Move data with code. NumPy and pandas until reshaping a dataset is faster than opening a spreadsheet. Filtering, grouping, joining, aggregating.
4. Get it and clean it. Pull data from files, APIs, SQL and NoSQL databases. Then handle what is missing, duplicated, mistyped or quietly wrong, which is most of it.
5. Explain it, then predict. Visualisation and statistics that hold up to questioning. Then your first models: linear and tree based methods for when you have labels, clustering and dimensionality reduction for when you do not.
Python NumPy pandas Matplotlib SQL MongoDB scikit-learn Git and GitHub Statistics Linear algebra
What you can show at the end
An analysis of a real dataset, from collection to conclusion, with the models you built and an honest account of what they get wrong.
2 Machine Learning Engineer Months 4 to 6
You arrive able to call a model. You leave able to build one, train it properly, and say why it is failing when it does.
1. Build a network. Feed forward networks layer by layer, from a single unit upward, until nothing inside one is a black box to you.
2. Make training work. Loss functions, optimisers, learning rates, initialisation and batching. The settings that decide whether a model converges at all, and what to change when it does not.
3. Diagnose failure. Error analysis and regularisation. Given a model that underperforms, work out whether the problem is the data, the capacity, the training, or the question, and fix the right one.
4. Teach it to see. Convolutions and pooling, deep vision architectures, augmentation, transfer learning and object detection. A vision problem solved end to end.
5. Teach it to remember. Recurrent networks and sequences, built fluently in Keras and TensorFlow rather than copied from a tutorial.
TensorFlow Keras Neural networks CNNs RNNs Optimisation Regularisation Transfer learning Error analysis
What you can show at the end
A trained deep learning model on a problem you chose, with the training curves, the failure analysis, and the record of what you changed and why.
3 AI Engineer Months 7 to 9
You arrive able to train models. You leave able to build products on top of them, which is a different job and the one most companies are hiring for.
1. Forecast what comes next. Time series, where the order of the data is the whole problem and the usual shuffling tricks quietly destroy the answer.
2. Turn language into numbers. NLP fundamentals and embeddings. How text becomes something a model can work with, and what gets lost on the way.
3. Open up the chatbot. Transformers and attention. What a large language model is actually doing when it answers, and what that means for when to trust it.
4. Ground it in your own data. Retrieval augmented generation, so a system answers from documents you control instead of from what it happens to have memorised.
5. Let it act. Agents, tool use and MCP. Systems that do things in the world rather than only describe them, and the judgment about when that is a good idea.
NLP Embeddings Transformers LLMs RAG Vector databases Agents MCP Time series APIs
What you can show at the end
A working AI application someone else can use, deployed, with your reasoning about what it does well and where it should not be trusted.
The finish¶
The programme ends with a portfolio project of your own choosing, built to a standard you would be comfortable defending in an interview. Not a tutorial followed to completion. Something you decided to make.
By then you will have three trimesters of projects behind it, which is the difference between a candidate who says they learned machine learning and one who can open a repository and show it.
How material appears¶
Each module opens on this site the week the cohort reaches it. That is deliberate. The order is what makes the nine months work, and having everything visible at once encourages skipping ahead to the interesting-sounding part before the foundation is there.
The roadmap stays complete, so you always know what is coming.