B
Bias AI System
The Bias AI System is an enterprise-grade AI Audit System built to detect, trace, and mitigate algorithmic bias in machine learning models and datasets. Combining deterministic statistical engines with Google Gemini LLM reasoning, it generates compliance-grade narrative audits, legal exposure risk mappings, and simulated bias mitigations via an interactive dashboard interface.
PythonFastAPIMLGenAILLM
About this project
Upload & Suggestion: The user uploads a CSV file through FileUpload.tsx. The backend automatically infers suitable target, sensitive, time, and text columns using suggest_column_roles in preprocessing.py.
Preprocessing: The dataset is parsed, missing values are imputed, and target columns are dynamically binarized. Text columns are encoded using sentence embeddings via text_encoder.py.
Core Fairness Core Metrics: Deterministic checks compute Demographic Parity, Equalized Odds, Adverse Impact Ratios, and the compound Bias Severity Index (BSI).
Proxy Tracing: Features correlated with the sensitive attribute are evaluated using Pearson and point-biserial correlations.
Temporal & Text Analyses: Temporal drift scans check for worsening or improving bias, while TF-IDF and VADER scan unstructured text for language bias.
Gemini Consolidator: Results are packed into a unified prompt for Gemini 1.5 Flash in llm_service.py to generate plain-English explanations, audit narratives, legal risk mappings, and 4-step rectification plans in a single pass.
Mitigation Simulation: Standard baseline model predictions are contrasted against a mitigated pipeline applying group reweighting.
Dashboard Render: The final enriched JSON payload (matching analyze-response.json) is displayed in the React dashboard results/page.tsx.