Machine Learning

This course involves real mathematics, taught from the ground up but unavoidable. Python fluency is a hard prerequisite. For many people Data Analytics is the better route, and that is not a lesser answer.

Taught, not assumed — but it is real and it is here for a reason.

• Vectors and matrices as data
  Why a dataset is a matrix and what an operation on it means.
• Derivatives, and what "learning" means mechanically hands on
  Gradient descent explained until it stops being a magic word.
• Probability and distributions
  Conditional probability, and why a 99% accurate test can still be usually wrong.
• Loss functions
  How a model is told it is wrong, and why the choice changes what it learns.

Most of the job is not the model.

• Framing a problem as a prediction task
  And recognising when it is not one — a great many business problems are answered better by a query.
• Features, and why they matter more than the algorithm hands on
  Encoding, scaling, and creating the variable that actually carries the signal.
• Train, validation, test — and leakage hands on
  The mistake that produces a brilliant model which fails on day one.
• Baselines hands on
  Always beat "predict the average" before claiming anything.

Regression and classification with scikit-learn.

• Linear and logistic regression hands on
  Starting with the models you can explain to a client.
• Decision trees and random forests hands on
  Where they win, and how they overfit.
• Gradient boosting hands on
  XGBoost and friends — what wins most tabular problems in practice.
• Evaluating honestly hands on
  Precision, recall, ROC — and why accuracy is the wrong measure when 1% of cases are positive.

When there are no labels.

• Clustering and segmentation hands on
  k-means, and how to tell whether the clusters mean anything.
• Dimensionality reduction
  PCA for compression and for seeing structure.
• Working with text hands on
  Tokenising, embeddings, and classifying Urdu and English messages.
• Anomaly detection
  Fraud and fault patterns, and the false-positive cost nobody budgets for.

Enough depth to be useful, honest about what needs a research team.

• From logistic regression to a network hands on
  Layers, activations and backpropagation, built up rather than dropped in.
• Training in PyTorch hands on
  Batches, epochs, overfitting and early stopping.
• Transfer learning hands on
  Fine-tuning an existing model — the only realistic route for most projects here.
• When not to use deep learning
  Small tabular data, and the honest observation that gradient boosting usually wins it.

The step that decides whether the work was worth anything.

• Serving a model as an API hands on
  Wrapping it so an application can call it.
• Drift, and the model that quietly stops working hands on
  Monitoring, retraining, and why yesterday's accuracy is not today's.
• Explaining a prediction
  Feature importance and SHAP — and the legal and ethical weight of "the model decided".
• Bias and harm
  How a model trained on past decisions reproduces past unfairness, and what to check before deploying one that affects people.

Before you start

• Python Programming, or equivalent fluency
  Not negotiable. This course does not teach programming.
• Comfort with mathematics
  Real mathematics is involved and is taught here, but a student who dislikes it will find this hard going. Data Analytics is the better route for many people, and that is not a lesser answer.

Where it leads

• AI Automation & n8n
  For applying models in business workflows rather than building them.

Roles this prepares you for

• Junior ML Engineer
  Uncommon as a first role in Pakistan; usually reached from a developer or analyst position.
• Data Scientist
  Typically expects a degree-level background as well as these skills.
• Analytics Engineer
  The more reachable route, combining this with the data analytics skill set.

Interested in this programme?

Try the subject free first — there is no reason to pay before you know you like it.