Machine Learning Foundations for LLM Work
The classical machine learning, statistics and probability the LLM courses assume, learned by building a ticket-triage model you can defend with numbers.
3 modules · 10 lessons~26 hoursUpdated 2026-11-13
- Lessons10
- Quizzes10
- Sources67
- Of study~26h
About this course
About this course
Every later course assumes you can split data without leaking, read precision and recall, say what a confidence interval means and follow a gradient. This course teaches exactly that, with the LLM-era reader in mind: baselines and leakage-free splits, linear and logistic regression fitted by gradient descent in numpy, metrics, thresholds and calibration, overfitting and cross-validation, the probability and statistics behind every eval number, and vectors, embeddings and a small neural network on CPU. Each lesson works on the same synthetic support-ticket dataset, pairs a rigorous explanation with a short video, and ends with a quiz and exercises you do with Claude. The capstone builds a triage model end to end and writes the one-page memo that says whether to ship it.
About the running example
The examples follow Hollowbrook Systems, a fictional mid-size software company whose operations team runs the support inbox. The course ships a generator for about 6,000 synthetic tickets (subject, body, product area, plan, category, an urgent flag, hours to resolve) plus a small usage table, with a known data-generating process, so you can check calibration against the true probabilities and watch leakage happen on purpose. Support tickets exist in every company; nothing here needs domain knowledge, and every ticket, customer and name is invented.
What you'll learn.
- 1Frame a prediction task as inputs, labels, metric and decision, build majority-class and heuristic baselines, and say what a model must beat
- 2Split data without leakage (random, stratified, time-based), explain why validation and test sets have different jobs, and find the leaks in a pipeline
- 3Fit linear and logistic regression by gradient descent in numpy, then choose a classification threshold from precision, recall, ROC and PR curves and a calibration check
- 4Diagnose overfitting from learning curves, apply regularisation and cross-validation, and explain bias and variance and where double descent breaks the classical picture
- 5Compute proportions, expected values, conditional probabilities, standard errors and confidence intervals, and run a paired comparison that says whether a 5-point difference on 30 cases means anything
- 6Compute dot products, cosine similarity and matrix products by hand, use embeddings for nearest-neighbour search, and train a small MLP on CPU in PyTorch while explaining its forward pass and backpropagation
3 modules, 10 lessons.
12h 50m of lessons
Module 1 · 3 lessons · 3h 25m
Learning from data
What a learning problem is, the baselines a model must beat, splits that don't lie, and fitting a line by gradient descent.
Module 2 · 3 lessons · 4h
Classifiers and generalisation
Logistic regression, the metrics and calibration that turn scores into decisions, and why models overfit and what to do about it.
Module 3 · 4 lessons · 5h 25m
The maths the later courses assume
Probability for evaluation, uncertainty and paired comparisons, vectors and embeddings, and a small neural network from scratch on CPU.
Ticket triage you can defend
Build, evaluate and defend an urgency-and-category triage model for Hollowbrook's support inbox end to end, from a leakage-free split to calibrated probabilities, confidence intervals and a one-page memo that says whether to ship it.