SimretAbebe/Fintech-Review-Analytics

NLP sentiment & thematic analysis of Ethiopian mobile banking app reviews (CBE, BOA, Dashen) using spaCy/transformers, with PostgreSQL storage # Ethiopian Bank Customer Review Platform An explainable AI system that predicts customer risk from Google Play Store reviews of Ethiopian banks (CBE, Bank of Abyssinia, Dashen Bank) — with native support for reviews written in English, Amharic, or Afaan Oromo. 🔗 **Live Demo** ## Business Problem Ethiopian banks receive thousands of customer reviews every month on the Google Play Store, but no team has the capacity to read them all manually. As a result, serious problems — failed transfers, login errors — often go unnoticed until many customers are already frustrated or have left. A second, less visible problem compounds this: a meaningful share of these reviews are written in Amharic or Afaan Oromo, and the original English-only analysis pipeline had no way to understand them, silently misreading real customer feedback. ## Solution Overview This project upgrades an earlier Week 2 review-analytics project from purely descriptive sentiment/theme analysis into a predictive, explainable, multilingual risk-intelligence system: 1. Engineers a proxy target label (`is_high_risk`) from star rating and complaint theme, since no direct churn label exists in the raw data — with rating deliberately excluded from the model's inputs to avoid data leakage 2. Trains and compares classification models, selecting the one that catches every actual high-risk review in testing 3. Adds SHAP explainability so every prediction can be traced back to specific contributing factors 4. Detects and translates Amharic and Afaan Oromo reviews, correcting a measured bias in the original pipeline 5. Serves everything through a live, interactive dashboard, containerized with Docker and deployed publicly ## Key Results - **Model recall: 100%** — the final model (Random Forest) caught every genuinely high-risk review in the held-out test set - **ROC-AUC: 0.982**, precision: 76.2% - **98 non-English reviews found** (68 Amharic, 30 Afaan Oromo) that the original pipeline had no way to understand …