toxicity-detection

Local AI text classifier for Vietnamese: scam/phishing detection, emotion and sentiment analysis, toxicity and spam checks. Django + Hugging Face, runs on CPU, no API key.

Detecting Toxic comments using machine learning

Production-grade AI safety middleware — prompt injection detection, PII masking (Indian Aadhaar/PAN), toxicity filtering, structured outputs via FastAPI

“Built a multi-label toxic comment classification system using TF-IDF features and a One-Vs-Rest Logistic Regression model, evaluated with F1-score and ROC-AUC to handle class imbalance.”

Russian text toxicity detection.

Toxicity detection in a conversation or phases.

A multilingual text analysis system that performs sentiment analysis and toxicity detection with detoxified text generation.

A GitHub Action that hides abusive issue, pull request and discussion comments with the Unprofane profanity and toxicity API

Stop prompt injections in 20ms. The safety toolkit every LLM app needs. No API keys, no complex setup, just `pip install llm-guard` and you're protected.

ToxiGuard AI is a browser extension that detects and censors toxic language in real-time using TensorFlow.js. It offers fine-grained controls, visual feedback, auto-censor, adjustable sensitivity, and respects user privacy.

A novel approach to news summarization that promotes positive and constructive narratives by integrating sentiment and toxicity control into the summarisation process.

A comprehensive evaluation of LLM trustworthiness (toxicity, bias, and privacy) using the DecodingTrust framework on Microsoft Phi-2.

Toxic Tweets Classification. A machine learning project that classifies tweets as toxic or non-toxic using NLP techniques and multiple classification algorithms (Decision Trees, Random Forest, Naive Bayes, K-NN, SVM). Implements Bag of Words and TF-IDF feature engineering, achieving F1-scores above 0.93. Implemented entirely in Jupyter Notebook

Production-ready LLM safety layer: prompt injection detection, PII redaction, toxicity filtering, topic guardrails via FastAPI + Streamlit

Real-time NLP toxicity classification microservice using TF-IDF & Logistic Regression (90.72% accuracy on 77.8k Turkish comments).

A multilingual text analysis system that performs sentiment analysis and toxicity detection with detoxified text generation.

Fine-tuning a compact transformer for robust detection of evasive toxic language across decentralized social media platforms

Zero-shot LLM toxicity classification on Civil Comments. Compares Integrated Gradients vs. attention attribution consistency, evaluates demographic fairness (SPD, EOpp), and provides an interactive Streamlit explorer with local and Gemini API backends.

Real-time cyberbullying detection system using a fine-tuned BERT model to classify toxic messages. Integrates a Discord bot with a FastAPI backend deployed on AWS via containers, automatically flagging and removing harmful content while notifying admins with confidence scores.