Module 3 of 13 in the AI Track, a free, self-directed curriculum for building AI that moves revenue.
The idea
In applied ML, 80% of the work and most of the failures live in the data. The model is the easy part. What matters is feature engineering, which is turning raw fields into something with signal. Encoding categories. Scaling numbers. And above all, not leaking. Data leakage is the number one way a model lies to you, by catching a glimpse of the answer while it trains. It looks brilliant in testing and falls over the week you ship it.
Conversion data is always lopsided. Most leads never convert. So you need to know why plain accuracy is useless on data like that, which is the next module, and how to deal with missing values without poisoning the whole thing.
Why it matters
Every CRM export is a mess. Taking a dirty prospect list and turning it into clean features with no leaks is the difference between a system that works and one that demos well and dies in production. This is the module that decides whether anything you build later is real.
Build this
A documented pass over a real, messy CRM export. Clean it, engineer the features, write down every decision you made and why. It’s the unglamorous pipeline everything else sits on top of.
How to work through it
- Géron chapter 2, the end-to-end project
- The data units of Google’s crash course
- Engineer features from a real messy export and document each call
- Find one example of leakage hiding in your own data
Resources
- Hands-On Machine Learning by Géron, chapter 2
- Google’s Machine Learning Crash Course, the data units