ML Foundations

Turn a raw event log into a model you can defend — and learn when a pre-trained tabular model beats the one you built.

10 lessons 3 phases beginner

Get started Signed in as your account.

what you'll build

what you walk away with.

01 A target you defined yourself, and a customer-by-cutoff panel built from a raw 1M-line event log.
02 A measured feature ladder — from a chance-level naive model to 0.84 AUC, knowing where it plateaus.
03 A from-scratch intuition for boosting: a booster recovering a function you defined, one residual round at a time.
04 Capacity knobs you set yourself, against a train-vs-validation curve that explains why.
05 A like-for-like NEXUS vs XGBoost comparison with native categorical handling on both sides.
06 Validation you can trust: grouped and temporal splits with a measured resolution floor.
07 A correct as-of join — and the measured 0.985-vs-0.740 lie the incorrect one tells.
08 Calibrated probabilities and an operating threshold you chose with cost knobs.
09 A measured sample-size curve showing where the pre-trained model structurally wins.
10 A champion-challenger decision rule for which model to run when.

curriculum

10 lessons across 3 phases.

phase A · Foundations — from events to a model you understand 4 lessons

01 Framing the problem
02 From chance to a model
03 How boosting actually works
04 Overfitting and capacity

phase B · Comparison and proof 4 lessons

05 The fair head-to-head
06 Validation you can trust
07 Point-in-time correctness
08 Imbalance and calibration

phase C · Choosing 2 lessons

09 When the pre-trained model wins
10 Using both

prerequisites

what you need before you start.

Python + pandas: Comfortable loading DataFrames, groupby, and column operations. No ML background assumed.
NEXUS API key: Provided for the workshop (format ak_...) — needed from lesson 5 onward.
Cloudsmith token: Required at setup time to install the private fundamental-client SDK — even for the lessons that make no API calls.
No toolchain setup: Nothing to install by hand — setup builds the Python environment for you. One system library (OpenMP) is checked with an exact fix-it command if missing.

start

how to start.

Once Claude Code is set up, tell it “set up the ml foundations workshop” and it clones the project and queues the first lesson.

New here? The getting-started guide walks you through installing the lwc CLI and the Claude Code plugin once — everything you need before that line works.