Sentiment Analysis on Movie Reviews
The problem
Build something that reads a movie review and correctly calls it positive or negative — simple in theory, full of edge cases in practice.
What I did
Cleaned and prepped the text, trained three models, and used 5-fold cross-validation to make sure the result wasn't a lucky split. Then read through every misclassified review by hand.
What came of it
The real failure mode wasn't simple negation — it was reviews praising a different movie while criticizing the one being reviewed, and sarcasm. About 90% accuracy, but the real value was knowing exactly why the other 10% failed.