← AI Engineering / Case study 06
Customer Feedback Intelligence.
"Because reading 34,000 reviews manually is not data science. It's punishment." An NLP pipeline that turns raw Amazon reviews into sentiment, product groups and summaries a business can act on.
- Team
- Pavitra & Rima
- Context
- Ironhack AI Engineering
- Year
- 2026
- Data
- 34,620 Amazon reviews
Project presentation · use ← → or swipe
What it does
- Classify sentiment. Star ratings mapped to three classes and balanced (812 per class). Text prepared with stemming, negation handling ("not_good") and TF-IDF bigrams, then labelled with VADER and a RoBERTa sentiment model (cardiffnlp twitter-roberta-base).
- Group products. Sentence-Transformer embeddings (all-MiniLM-L6-v2) clustered with K-Means. The elbow method chose k = 5, shown with PCA, and each cluster got a business-friendly name like "Top-Rated & Premium Picks".
- Summarise. An LLM step writes a short summary per cluster, so a manager can read 5 paragraphs instead of 34K reviews.
- Show it. Streamlit + Plotly and Gradio dashboards with cluster selector, top products and raw data view.