Turn a brand's product reviews (Amazon, Shopify, Trustpilot, MercadoLibre) into a decision-ready report — sentiment breakdown, the top recurring complaints and praised features, and 3–5 concrete, quantified recommendations.

Now available in English and Spanish with optimized sentiment models for each language.

Sentiment breakdown — overall, by product, and over time

Top 5 recurring complaints and Top 5 praised features (NMF topic modeling)

3–5 recommendations — e.g. "40% of negative reviews center on packaging/staleness — fix the most-affected SKUs."

A shareable client-style HTML report plus supporting CSVs

|------|----------|-----------------|-------------| | ReviewSignal.ipynb | English | VADER | End-to-end analysis notebook (English reviews) | | es/ReviewSignal-ES.ipynb | Spanish | Multilingual BERT | End-to-end analysis notebook (Spanish reviews) | | outputs/ | — | — | English reports (HTML, Markdown, CSVs) | | outputs-es/ | — | — | Spanish reports (HTML, Markdown, CSVs) |

reviewsignal_report.html — Client-facing report (executive summary + embedded charts + recommendations)

reviewsignal_report.md — Markdown summary

recommendations.csv — Recommendations with focus SKUs and affected product count

top_complaints.csv — Complaint themes + prevalence

top_praise.csv — Praise themes + prevalencecus SKUs | | outputs/top_complaints.csv | Complaint themes + prevalence | | outputs/top_praise.csv | Praise themes + prevalence |

Demo data is the Amazon Fine Food Reviews dataset (~568k reviews), pulled via the Kaggle API (kagglehub):

The loader normalizes everything to a platform-agnostic schema (platform, product_id, user_id, rating, timestamp, title, text), so swapping in a Shopify / Trustpilot / MercadoLibre export is a one-line mapping change. kaggle auth login # Configure Kaggle API credentials

Then open ReviewSignal.ipynb and run all cells. Set SAMPLE_SIZE = None in the config cell to run on the full dataset (the default samples 80,000 reviews for responsiveness).

Sentiment — star ratings (1–2 negative, 3 neutral, 4–5 positive) plus a VADER compound score on the review text for time trends.

Topics — TF-IDF + NMF run separately on negative and positive reviews; each review is assigned its dominant theme and the theme's share is its % of that class.

Product names — the raw data only has ASINs, so a readable name is derived from the distinctive TF-IDF terms in each product's own reviews.