HOW TO MAKE ML

Personal learning repository | Machine learning fundamentals | 2026 (ongoing)

A collection of small, self-contained machine learning exercises, each in its own subdirectory and focused on a single model. Rather than relying on ready-made pipelines, the exercises are implemented from the ground up to build a real understanding of how each algorithm works, from a single neuron trained with gradient descent to clustering, matrix factorisation and dimensionality reduction. It is probably my most used repository and it keeps growing as I explore new models.

Key contributions:

Implemented a single artificial neuron (logistic regression) from scratch in NumPy, trained with gradient descent, with an animated visualisation of the decision boundary converging during training.

Built a K-Nearest Neighbours classifier on a three-feature binary dataset with an interactive 3D Plotly visualisation and a comparison of predictions across different values of k.

Applied K-Means clustering to around 32,000 Spotify tracks (12 standardised audio features) and around 100,000 US traffic accident records, choosing k with the elbow method and silhouette score and visualising clusters through a PCA projection.

Decomposed MNIST digits and Olivetti face images with Non-negative Matrix Factorisation and visualised the learned components and reconstructions as images and animations.

Projected 99-feature ANSUR body measurement records into 2D with t-SNE to explore clustering in body shapes, colouring the embedding by individual features, BMI class and height class.

Documented categorical data encoding techniques in a dedicated notebook and maintained a single structured requirements file covering every exercise.