CivicMind

Municipal complaint systems receive free-text submissions — often informal and code-mixed (Hinglish) — that have to be manually read, categorized, and prioritized before anyone can act on them.

Built an NLP classification pipeline that automatically tags incoming complaints by category and severity across both English and Hinglish input, exposing the result as a simple REST endpoint that any front-end or ops tool can call.

1. 01Text preprocessing: lowercasing, punctuation removal, and whitespace normalization
2. 02Feature engineering: combined word-level (1–3 n-gram) and character-level (3–5 n-gram) TF-IDF vectors
3. 03Logistic Regression classifier with class balancing, validated via stratified 5-fold cross-validation
4. 04Fixed severity mapping applied per predicted category
5. 05Flask REST API (POST /predict) deployed on Render

• ▸Manually built and labeled a custom dataset spanning 5 categories in English and Hinglish
• ▸Engineered combined word + character TF-IDF features to handle informal, code-mixed phrasing
• ▸Trained and validated a class-balanced Logistic Regression model with stratified k-fold cross-validation
• ▸Wrapped the trained model in a Flask REST API and deployed it to Render

• ▸Handling Hinglish (code-mixed English/Hindi) text without a large pretrained multilingual model
• ▸Choosing character-level features specifically to stay robust to spelling variation in informal complaints

A deployed REST API achieving ~73% test accuracy (~67% mean cross-validation accuracy) under stratified 5-fold cross-validation, classifying bilingual civic complaints by category and severity.