FairSense-AgentiX: An Agentic Platform for Multi-Modal Fairness and AI Risk Analysis

FairSense-AgentiX is an open-source Python platform for automated detection of bias in text and images and for assessing AI deployment scenarios against established risk frameworks.Existing fairness tools such as AI Fairness 360 (Bellamy et al., 2019), Fairlearn (Weerts et al., 2023), and LangFair (Bouchard et al., 2025) address structured datasets or large language model (LLM) output metrics but cannot audit unstructured content across text and images, explain their reasoning transparently, or connect findings to deployment-level risk frameworks in a single workflow.To address this gap, it implements a multi-node agentic workflow using LangGraph (LangChain AI, 2024), unlike static classifiers or rule-based checkers, in which an orchestrator routes each request by declared modality, executes a specialised subgraph, and iteratively refines the output through a self-critique loop.The platform supports three analysis modes (text bias detection, image bias detection, and AI risk assessment) and exposes both a Python library and a FastAPI (Ramírez, 2019) REST/WebSocket server with an accompanying React web interface.FairSense-AgentiX is designed to be LLM-provider-agnostic and toolswappable, making it applicable in a wide range of research and production contexts.It builds on and substantially extends a prior static pipeline, FairSense-AI (Raza et al., 2025), by replacing single-pass LLM inference with an orchestrated, self-correcting agentic loop and adding a retrieval-based risk assessment module aligned with established governance frameworks.