AI & Machine Learning

NumPy: Whole Columns at Once

Machine learning is maths on thousands of numbers at once. NumPy arrays let you do it in one line — no loops — and they're what every ML library runs on underneath.

pandas: Tables You Can Question

pandas is the spreadsheet of data science. Load 80 farms, select, filter, group and clean them — the everyday moves before any model is trained.

Visualising Data

A chart shows in a second what a table hides. Learn to pick the right chart for the question and draw it with matplotlib — scatter, bar, line and histogram.

Exploratory Data Analysis

Before any model: get to know your data. Summaries, distributions, outliers and correlations — and the discipline to say what the data does and doesn't show.

Vectors & Matrices

A model's prediction is a weighted sum of its inputs — a dot product. Stack many examples into a matrix and one multiplication predicts them all.

Statistics That Matter

Mean, median, standard deviation, the normal distribution, and why a sample never tells you the exact truth — the statistics every model result leans on.

Probability & Bayes

Why a positive test can still be unlikely, how spam filters think, and how to reason about uncertainty with simulation and Bayes' rule.

Gradients & Optimisation

How does a model find its best weights? It measures how wrong it is, feels which way is downhill, and takes a step. That's gradient descent — the engine of almost all machine learning.

Rainfall & Crop Yield

Predict maize yield from rainfall on smallholder farms in Nakuru County — and see exactly what “training a model” means.

Gradient Descent from Scratch

Train a linear regression model the way neural networks are trained: compute the gradient, step, repeat — and watch the loss fall epoch by epoch.

Overfitting & Regularisation

A model that memorises its training data fails on new data. Learn to spot overfitting from train and test error, and to rein it in with regularisation.

Cross-Validation

One train/test split can flatter or punish a model by luck. Cross-validation tests on every slice of the data in turn — the standard way to compare models honestly.

Logistic Regression

Your first classifier: predict whether a maize field will get blight, as a probability — and learn why 77% accuracy can be worthless.

Measuring Classifiers

Accuracy hides the mistakes that matter. Confusion matrices, precision, recall, F1, ROC curves — and choosing a threshold for the job the model actually has to do.

k-Nearest Neighbours

Predict from the most similar past examples. k-NN is the simplest model there is — and the clearest demonstration of why feature scaling matters.

Decision Trees & Random Forests

Models that ask a series of yes/no questions — readable enough to print as rules. Then why a forest of trees beats any single one.

Gradient Boosting

Trees trained one after another, each fixing the last one's mistakes. Gradient boosting — and XGBoost — dominate tabular-data competitions. Learn how, and when it's worth it.

Blight Early Warning

A county agriculture office wants to know which maize fields to visit before blight spreads. Build, tune and defend an early-warning classifier end to end.

Encoding & Scaling

Models only understand numbers on sensible scales. Turn categories into numbers without inventing fake orders, scale numeric columns, and handle categories you've never seen.

Feature Engineering

Often the biggest gains don't come from a fancier model but from better inputs. Craft features that encode what you know about the problem — and measure whether they help.

Data Leakage & Pipelines

The bug that makes models look brilliant and fail in the field. Spot leakage, then build end-to-end pipelines that prepare raw data and predict in one honest step.

k-Means Clustering

No labels, just data: let the algorithm group 250 mobile-money customers by behaviour, choose how many groups there really are, and describe what each one means.

PCA: Seeing Many Columns at Once

Squeeze six customer features into two directions that keep most of the information — to plot high-dimensional data, compress it, and see the clusters for yourself.

Anomaly Detection

Fraud is rare and rarely labelled. Hunt for suspicious mobile-money transactions with rules, z-scores and Isolation Forests — and learn why "unusual" and "fraudulent" aren't the same thing.

Mobile-Money Fraud Watch

A mobile-money provider's fraud team can review 25 transactions a day. Build the detector that fills their queue — with no fraud labels to train on — and report what it catches and what it costs.

Trends & Seasonality

Eight years of maize prices. Pull the series apart into its long-term trend, its yearly rhythm and the shocks that fit neither — the first step of any forecast.

Forecasting

Predict next year's prices — honestly. Build baselines, a regression with trend and seasonal features, test on a held-out year, and deal with the shock in the training data.

Market Price Forecast

A traders' association in Nairobi wants next year's monthly maize prices at Gikomba market to plan storage and sales. Build a forecast that beats the baseline, and deliver it with honest caveats.

The Perceptron

Build the single artificial neuron that started it all, train it with the perceptron rule — and hit the wall that stalled AI for over a decade.

Backpropagation

How does a network with thousands of weights know which ones to change? The chain rule, applied backwards. Build it by hand — and train a two-layer network that solves XOR.

Multi-Layer Networks

Stack layers, pick activations, choose an optimiser — and train real neural networks with scikit-learn. Plus how to read a loss curve and stop a network overfitting.

Convolutions & Images

To a computer an image is a grid of numbers. Slide small filters over it to find edges and shapes — the idea behind every image-recognition system — using real handwritten digits.

Handwritten Digit Reader

A savings group wants to digitise years of handwritten ledger entries. Train a digit reader on real handwriting, find where it fails, and design a workflow where people check what the model isn't sure about.

Text as Data

Models only understand numbers. Take 400 real-sounding customer reviews in English and Swahili, break them into tokens, and count what people talk about.

Bag of Words & TF-IDF

Turn every review into a row of numbers by counting its words — then weight the counts so the words that actually distinguish one text from another stand out.

Sentiment Classification

Train a model that reads a review and says whether the customer is happy — then open it up, see the words it relies on, and find where word-counting breaks down.

Word Embeddings

Give every word a list of numbers so that words used alike sit close together. Build embeddings from scratch from the reviews — and discover what they capture and what they miss.

Attention & Transformers

The idea behind every modern language model: let each word look at the others and decide which ones matter. Build scaled dot-product attention in NumPy, line by line.

How LLMs Work

Build a tiny language model that writes reviews, see why it confidently makes things up, and fix that the way real AI assistants do: by retrieving facts and putting them in the prompt.

Customer Feedback Assistant

An agri-supplies company gets hundreds of reviews a month in English and Swahili. Build a tool that flags unhappy customers, finds what they're unhappy about, and drafts a grounded reply.

Bias & Fairness

A digital lender's model never sees where applicants live — yet it turns down rural customers who repay just as reliably. Measure the bias, find where it comes from, and reduce it.

Explainability

An applicant was turned down and wants to know why. Explain the model globally — which features matter — and locally — why this person — and tell them what would have changed the answer.

Privacy & Data Ethics

Removing names doesn't make data anonymous. Re-identify people in a "de-identified" dataset, then decide what can safely be shared, and in what form.

Deploying & Monitoring

A model in a notebook helps nobody. Save it, wrap it in a function an app can call safely, and set up the monitoring that catches the day the world changes under it.

Responsible Lending Audit

The lender wants to launch its credit model nationwide. Audit it end to end — accuracy, fairness, explanations — choose a policy you can defend, and publish a model card.

Hugging Face: Models & the Hub

Most AI products start from a pretrained model, not from scratch. Run a real Hugging Face model in your browser, explore the Hub's Swahili models like a practitioner — tasks, licences, sizes — and shortlist the right one for a job.

From NumPy to PyTorch

PyTorch is the framework behind most modern AI. You already know what it does — you built it by hand in NumPy. Map every line of a PyTorch training loop to your own code, build a mini PyTorch, and write the autograd engine at its heart.

Experiments, Notebooks & GPUs

Real ML work is dozens of experiments. Make yours reproducible, track every run so you can say which model was best and why, and learn how seeds and splits can fool you — the habits behind tools like MLflow and Weights & Biases.