14 · Personal & Academic Project
Amazon Review Sentiment Analyzer
Converting thousands of product reviews into clear quality signals
Business value
The system reached 91% classification accuracy and surfaced durability complaints across two product subcategories as a concrete quality-improvement signal.
01 · Business problem
More than 17,000 customer reviews contained repeated product complaints, but manual reading could not reliably identify the categories and themes driving negative sentiment.
02 · My approach
I cleaned and tokenized the review text, trained a BERT-based classifier, and connected sentiment predictions to product-level interpretation.
03 · What I built
A TensorFlow BERT pipeline, preprocessing and evaluation workflow, subcategory analysis, and a Gradio interface for interactive review classification.
What changed my thinking
The corrections, tradeoffsand decisions behind it.
- 01
A sentiment score was not the final insight
The useful question was which product areas repeatedly generated negative experiences.
- 02
Class performance mattered more than one headline metric
I reviewed where the classifier confused neutral and negative language instead of relying only on overall accuracy.
- 03
The interface changed the project
Gradio turned the model from a notebook result into something a non-technical user could test directly.