detoxify

Comments & Twitter accounts gRPC classification service.

A reproducible evaluation pipeline for assessing toxic content in language model outputs using Detoxify, developed as part of a 10-week research project.

Measures Reddit toxic-comment prevalence and evaluates the 2020 Great Ban with weighted DiD on 45.6M comments (out-only universe). Result: −0.89 pp (95% CI [−1.16, −0.62]); parallel trends not rejected; label-noise swing up to 2.9 pp.

Self-hosted FastAPI wrapper around the Detoxify multilingual toxicity model. Optional moderation backend for Quelora.

Reproduction study of the paper "Detoxifying Text with MaRCo: Controllable Revision with Experts and Anti-Experts"

Two-layer, fail-closed content moderation in Python using Detoxify signals, local Ollama review, and deterministic policy gates.

A simple FastAPI wrapper around the Detoxify library

Model-agnostic fast api service to evaluate large language models for bias, toxicity and hallucinations.

AI-based video moderation system that detects violence and toxic speech using multimodal analysis and applies selective content filtering with policy-based reporting.

Bachelor's thesis on removing hate from online comments using paraphrasing: algorithm DPhate