BiasTrace

ML and data science students routinely train models on datasets with undocumented demographic imbalance, sampling bias, or geographic skew — and only discover it during peer review, a thesis defense, or after deployment, when it's too late to fix cheaply. BiasTrace catches this before training starts. Users paste a dataset's README, description text, or a Kaggle/HuggingFace URL. GPT-4o analyzes the methodology for known bias red flags — small or single-source samples, missing demographic documentation, no external validation set — and returns a 0-100 bias-risk score with severity-rated, evidence-backed flags. What makes BiasTrace different from a simple AI classifier: when a URL is provided, it uses Bright Data's live web integration to pull the dataset's actual public discussion or issues page and check whether bias has already been reported by other researchers. This grounds the analysis in real evidence instead of pure model judgment. Every audit closes with a concrete verification checklist and is saved to the user's private history for future reference. Built entirely in native.builder — the frontend, Supabase-backed auth and database, and Edge Functions were all generated and iterated on inside the platform, from PRD through deployment, with GPT-4o and Bright Data wired in as external integrations. Target users: ML/data science students, thesis writers, and early-career researchers who need a fast, evidence-based check before they train on or cite a dataset.