Sefism Machine Learning Roadmap

What this roadmap is

The Machine learning roadmap is Sefism's path for programmers and students who want to train models on data, not only call someone else's API. It starts with Python and the maths you actually need, moves through data cleaning, classical models and honest evaluation, then deep learning in PyTorch, and ends with shipping a model as a reproducible, monitored service.

You choose a focus when you set it up:

• General machine learning: the broad path, from tabular data to deep learning.
• Computer vision: images and video, detection and segmentation. Adds a Computer vision in practice stage.
• NLP: text and language models. Adds an NLP in practice stage.

The shared stages are the same for everyone. The focus adds a stage, adjusts topics in the first stage (images as NumPy arrays, or working with text in Python), changes the deep learning and production milestones, and changes the suggested projects.

What you can do at the end

• Load, clean and explore a messy real dataset in pandas and explain what you found with charts.
• Train, tune and compare scikit-learn models with proper validation and no data leakage.
• Pick the right metric for a problem and explain bias, variance and leakage using your own projects.
• Build, train and fine-tune neural networks in PyTorch for images or text.
• Serve a model behind an API with tracked experiments, a versioned model and drift monitoring.

Who it suits, and who should start elsewhere

It suits programmers and students, including those planning an AI final year project, who want to understand and build models rather than only use them.

If you have never programmed, start with Programming foundations in Python. If you want to build products on hosted LLMs (prompting, RAG, tools and agents) rather than train models, the AI engineering roadmap is a better fit. If your interest is analysis, experiments and communicating findings more than modelling, look at Data science or Data analytics. Once you are deploying models regularly, MLOps and LLMOps takes the last stage much further.

Time and pace

This is one of Sefism's longer roadmaps, because it includes the maths and Python foundations:

• General focus: 7 stages, about 425 hours.
• Computer vision or NLP focus: 8 stages, about 495 hours.

The first two stages (Python for data, and The maths you need) are foundations, together about 130 hours. If you are already comfortable with them, your plan lets you skim or skip them, which shortens the path considerably.

A student's week (10 to 12 hours): two sessions on the stage's course or book, two sessions in a notebook applying it to a real dataset, and one on maths or statistics. The roadmap advises running the maths stage side by side with classical ML rather than finishing it first.

Someone working full time (about 5 hours): one evening of theory and a weekend notebook session. Skim the foundation stages you already know, and pick a focus early so the projects you build add up to one story.

Stage by stage

1. Python for data (about 60 hours, foundation)

Core Python for data work, Jupyter notebooks, NumPy arrays, shapes and indexing, vectorisation and broadcasting, pandas DataFrames for selecting, filtering and cleaning, groupby, merge and reshaping, charts that make one clear point, and CSV, JSON and Parquet. Vision learners add images as NumPy arrays; NLP learners add working with text in Python.

Milestone: answer five questions about a public CSV of at least 50,000 rows using pandas with no loops over rows, and produce three charts that each make one clear point.

2. The maths you need (about 70 hours, foundation)

Vectors, matrices and the dot product, matrix multiplication and shapes, derivatives and gradients, the chain rule, gradient descent, probability and Bayes' theorem, distributions, descriptive statistics, and hypothesis tests, p-values and confidence intervals. Aimed at intuition and code, not proofs.

Milestone: implement linear regression with gradient descent in NumPy from scratch, and explain in plain words what a p-value and a confidence interval do and do not tell you.

3. Data cleaning and feature engineering (about 35 hours)

Exploratory data analysis, missing values, outliers, encoding categorical data, scaling and normalisation, pipelines and ColumnTransformer, feature creation, feature selection, and optionally class imbalance.

Milestone: an exploratory analysis of a messy real dataset documenting every cleaning decision, and one engineered feature that measurably improves a simple baseline.

4. Classical machine learning (about 80 hours)

Supervised versus unsupervised learning, linear and logistic regression, bias, variance and overfitting, L1 and L2 regularisation, decision trees, random forests, gradient boosting, k-means clustering and principal component analysis, with scikit-learn as the working tool.

Milestone: a scikit-learn pipeline on tabular data that beats a dumb baseline with logistic regression, a random forest and a gradient boosted model, with an explanation of why the winner won.

5. Evaluating and tuning models (about 30 hours)

Train, validation and test splits, baselines, precision, recall and F1, ROC-AUC and PR-AUC, regression metrics, cross-validation, data leakage, hyperparameter search, and feature importance and explanations.

Milestone: a Kaggle competition entry with a cross-validated pipeline whose local score tracks the leaderboard, and a short note on which metric you optimised and why.

6. Deep learning (about 90 hours)

Neural networks and activations, tensors and autograd in PyTorch, loss functions, backpropagation, SGD and AdamW, the PyTorch training loop, loss curves, overfitting and regularisation, convolutional neural networks, embeddings, attention and transformers, and transfer learning and fine-tuning.

Milestone: fine-tune a pretrained image or text model on a dataset you collected yourself, with your own training loop and a validation curve that shows where it starts to overfit.

7. Computer vision in practice (about 70 hours, vision focus only)

OpenCV and classic image processing, image datasets and labelling, data augmentation, object detection, image segmentation, evaluating with IoU and mAP, vision transformers, and optionally running vision models on devices.

Milestone: fine-tune a pretrained detector or segmenter on images you labelled, report per-class mAP on a held-out set split by video or location, and show the ten worst mistakes with what you changed because of them.

8. NLP in practice (about 70 hours, NLP focus only)

Tokenisation, a bag of words and TF-IDF baseline, text datasets and labelling, sentence embeddings and semantic search, Hugging Face models and pipelines, fine-tuning text classifiers, named entity recognition, evaluating NLP models, and optionally prompting an LLM or fine-tuning.

Milestone: a text classifier or entity extractor on data you labelled with a written guide, beating a TF-IDF baseline with a fine-tuned transformer on macro F1, with results by class and by language and ten real mistakes explained.

9. Production ML and MLOps (about 60 hours)

Experiment tracking, data and model versioning, batch versus real-time serving, training-serving skew, CI/CD for ML, data and concept drift, retraining strategy, and ML system design.

Milestone: a containerised model API with tracked experiments, a registered model version, a CI pipeline that retrains and tests it, and a monitoring report that flags input drift (for vision, shifts in brightness, size or camera; for NLP, shifts in text length, language mix or vocabulary).

Projects by focus

The roadmap suggests four or five projects from the Sefism library for each focus, for example:

• General: Retail demand forecasting with inventory recommendations and Early warning system for student disengagement.
• Computer vision: Crop disease detection from phone photographs and Fabric defect inspection for small textile units.
• NLP: Roman Urdu and English customer message triage and Reproducing a published machine learning result.

How it works on Sefism

The public page at /roadmaps/machine-learning/ shows the full stage list, with the first two stages open to read. Everything on Sefism is free to learn from. A few actions use credits, adding a roadmap to your dashboard among them; new members get starting credits, the first roadmap is free, and anyone learning seriously who runs out can ask for more.

Adding the roadmap opens every stage in your dashboard, where you choose your focus, set your weekly hours and familiarity, and track progress topic by topic. Each topic has a short explanation, a check question, and a prompt to paste into an AI assistant so it quizzes you rather than doing the work. Following the roadmap also unlocks its companion cheat sheet. You can send Tauseef a question about any stage or project, or book a 1:1 session.

The companion cheat sheet

The Machine learning cheat sheet follows the same path with General, Computer vision and NLP editions and NumPy, pandas, scikit-learn and PyTorch snippets. Use it to revise after each stage and before exams or technical conversations.

Related on Sefism

• What is the difference between a data analyst, a data scientist and a machine learning engineer?
• My machine learning projects never leave a Jupyter notebook. How do I learn to deploy a model?
• Can I have a career in software or AI if I am bad at maths?
• Should I learn AI and machine learning, or become a strong software engineer first?
• Practical Deep Learning for Coders and the scikit-learn user guide, both used in several stages.

What you can do afterwards

You will have a cleaned and documented analysis, a tuned classical model with an honest evaluation, a fine-tuned deep learning model on data you collected, and one model running as a monitored service. That set covers what junior ML engineer and applied ML roles ask about, and gives a final year project real depth.

Frequently asked questions

How long does it take to learn machine learning? About 425 hours on the general focus, or about 495 with a vision or NLP stage. At 12 hours a week that is roughly 35 to 41 weeks. Skipping foundation stages you already know shortens it a lot.

How much maths do I need for machine learning? The maths stage covers linear algebra, calculus, probability and statistics at the level of intuition and code. The roadmap's view is that statistics will save you from more bad conclusions than calculus.

Should I learn deep learning or classical machine learning first? Classical first. On tabular data gradient boosted trees are usually the model to beat, and the roadmap asks you to build the dumbest possible baseline before anything else.

Computer vision or NLP? Choose the one whose problems you want to work on. Both share the first six stages and the production stage; the focus adds one practical stage and changes the projects.