FairSense-AgentiX¶
An open-source platform from the Vector Institute for multi-modal bias detection, explainable fairness auditing, and retrieval-based mapping of deployment scenarios to the MIT AI Risk Repository and NIST AI RMF.
Peer-reviewed open-source software
FairSense-AgentiX was published in the Journal of Open Source Software on 30 October 2026 (DOI: 10.21105/joss.11060). Software archive: Zenodo 10.5281/zenodo.23106857 (release v0.2.2). If you use this work in research, please cite the JOSS paper.
What is FairSense-AgentiX?¶
FairSense-AgentiX detects bias in text and images and assesses AI deployment scenarios against established risk frameworks. It uses a LangGraph orchestrator with a critique-and-refine loop—not a single-pass black-box classifier—so results include spans, explanations, scores, and step-by-step telemetry when you use the API or UI.
An orchestrator routes each request by declared modality (text, image, or scenario), runs the matching workflow, evaluates output quality with an LLM critic, and optionally refines before finalizing. Tools (OCR, vision models, embeddings, FAISS, LLMs) are injected via a registry so the same graphs run with real providers in production or lightweight fakes in CI.
Analysis modes¶
-
Text bias
Structured instances (type, severity, span, explanation) and HTML highlights.
-
Image bias
OCR/captioning or a vision-language path, then the same bias pipeline as text.
-
Risk assessment
Semantic search over 1,538 MIT AI Risk Repository (V4, Dec 2025) risks plus NIST AI RMF–aligned actions.
Key capabilities¶
Agentic orchestration¶
- Modality-based routing into specialised LangGraph subgraphs
- Post-hoc LLM evaluation (0–100) with configurable refinement
- WebSocket telemetry for live agent timelines in the React UI
Production APIs¶
- Python
FairSenseAPI · FastAPI REST · batch jobs · full API reference
Configurable stack¶
- LLM providers (OpenAI, Anthropic, OpenAI-compatible endpoints) or
fakefor tests - Swappable OCR/caption/embedder tools via environment variables
How it works¶
Every request follows the same orchestrator loop: route → execute → evaluate → deliver (with an optional refine pass when the critic score is low).
-
1 · Route
Choose text, image, or risk workflow from the request and run preflight checks.
-
2 · Execute
Run the LangGraph subgraph: bias pipeline, OCR/VLM path, or FAISS over 1,538 MIT risks.
-
3 · Evaluate
LLM critic scores output (0–100). Fail → refine; pass → finalize.
-
4 · Deliver
Spans, explanations, highlights, risk matches, and agent telemetry.
Quick start¶
=== "PyPI"
```bash
pip install fairsense-agentix
export FAIRSENSE_LLM_PROVIDER=openai
export FAIRSENSE_LLM_API_KEY=sk-...
```
=== "Source"
```bash
git clone https://github.com/VectorInstitute/fairsense-agentix.git
cd fairsense-agentix
uv sync && source .venv/bin/activate
```
Configure a real LLM provider first
The default FAIRSENSE_LLM_PROVIDER is fake (synthetic placeholders for tests). See
Getting Started before running real analyses.
from fairsense_agentix import FairSense
engine = FairSense()
result = engine.analyze_text(
"We're looking for a young, energetic developer to join our startup team."
)
print(result.summary)
Documentation map¶
| Guide | Description |
|---|---|
| Getting Started | Install, .env, first analysis |
| User Guide | Text, image, and risk workflows |
| Server Guide | FastAPI + React UI |
| Architecture | Graphs, tools, orchestrator |
| API Reference | Python and REST |
| Developer Guide | Contributing and extending |
| MIT risk index | Bundled V4 corpus and RMF semantics |
| Changelog | Release notes (e.g. v0.3.0) |
Use cases¶
-
Content & marketing
Moderation pipelines, ad copy, and inclusivity reviews before publish.
-
HR & policy
Job posts, internal communications, and DEI alignment checks.
-
ML governance
Pre-deployment risk mapping, audit trails, and fairness research with agent metadata.
Citation¶
If you use FairSense-AgentiX in academic work, please cite the JOSS article:
@article{Narayanan2026,
doi = {10.21105/joss.11060},
url = {https://doi.org/10.21105/joss.11060},
year = {2026},
publisher = {The Open Journal},
volume = {11},
number = {126},
pages = {11060},
author = {Narayanan, Aravind and Sekhon, Karanpal and Alinoori, Mahshid and Raza, Shaina},
title = {FairSense-AgentiX: An Agentic Platform for Multi-Modal Fairness and AI Risk Analysis},
journal = {Journal of Open Source Software}
}
License & support¶
MIT License · Issues · Contributing
Built with by the Vector Institute AI Engineering team.