(Please note: Since I am based in Spain, the video demo and the application's interface are currently in Spanish!)
I’ve been working on a recommendation and sentiment analysis system for Steam reviews for a few days now, and today I finally got the AI pipeline up and running and connected to the backend!
The development process presented me with a classic but challenging technical hurdle: severe class imbalance in the dataset (nearly 10 positive reviews for every 1 negative). The baseline model became lazy, shortcutting its way by predicting almost everything as positive.
Here is how I solved it:
Precompiled regular expressions in Python to clean raw, free-form user text in milliseconds (goodbye bottlenecks!).
Applied Random Undersampling to level the playing field.
Trained a Complement Naive Bayes (ComplementNB) classifier, specifically designed to handle skewed text distributions.
The results? The F1-Score for the negative class skyrocketed from a poor 0.22 to a solid 0.77, delivering a well-balanced confusion matrix that accurately catches real bugs and user complaints.
The production pipeline is now fully operational: the API is powered by #FastAPI (running incredibly light at just ~100MB of RAM) and the first frontend components are responding smoothly in #NextJS. Still a long way to go, but the app’s "brain" is officially alive.
https://www.youtube.com/watch?v=BtnJWd4yZfo
Any suggestions for improvement? I'd love to hear them!