CLASSIC ML TRACK · 23 LESSONS

Real ML problems as they reach an engineer

Every lesson is a problem that came from the business: why profit falls while revenue grows, why a 99%-accurate model is useless, how to set a threshold from the cost of an error. The walkthrough with the numbers is open; the practice and the quiz live in the simulator.

01 Half the “ML tasks” you are handed are solved by an if-else 02 Revenue grew, profit did not. We found out why without a model 03 The discount lifted conversion — and brought in less money 04 The first model saved millions — and it was a linear regression 05 The insight turned out to be worth more than the model 06 99.2% accurate — and it caught zero fraudsters 07 A weak model on strong features beats a strong model on weak ones 08 On tabular data, boosting beats neural networks 09 0.97 on the test set, 0.61 in production 10 A gain of 0.005 against a spread of 0.004 is noise 11 93% of the alerts are false — and the economics still work 12 A cluster without a name is not a result 13 The model learned that phones go with cases, knowing nothing about products 14 A naive forecast gets you 80% of the way to a complex model 15 The bigger model is more accurate — and does not pay for itself 16 If you have never overfitted a network, you cannot train one 17 Four thousand photos. Training from scratch will not fly 18 What separates a mid-level engineer is designing the whole system 19 A model without monitoring is an incident that has not happened yet 20 Business problem first, then the data, and only then the model 21 The best project is the one you decided not to build 22 Half an hour spent reading errors beats a week of tuning 23 Who fixes the model at three in the morning