Team
SkywardAI Labs · Fairness Intelligence Labs
Developing causally grounded, interpretable debiasing methods across black-box, open-weight, and agentic large language models.
Who we are
Fairness Intelligence is a research group at SkywardAI Labs dedicated to understanding and mitigating bias in large language models using causal methods. We bridge explainable AI and causal inference to build more trustworthy, accountable AI systems.
Our research spans three tiers of model access — black-box APIs, open-weight architectures, and multi-agent agentic systems — developing unified causal principles that apply across all levels of complexity.
We publish at top-tier venues (CORE A*, A, IEEE Q1) and are committed to open, transparent research that promotes safer AI deployment in high-stakes domains such as healthcare, hiring, education, and public information systems.
Publications
Each project below is a self-contained, static HTML page that walks through the framework pipeline — one is a live playground, the rest are read-only worked examples using illustrative data. Where a project has an accompanying paper, its venue and tier are shown alongside.
Generates multiple reasoning paths, estimates each path's causal effect via front-door adjustment, and casts a causal-weighted vote for the final answer. Run it directly.
EACL 2026 FindingsPrompt-injection guarding (GenTel-Shield) and confidence calibration wrapped around a five-stage RAG pipeline for resource-constrained small businesses.
PreprintDetects when demographic-substitution variance signals an unfair toxicity prediction, then selectively applies prompt-guided fairness correction.
WWW 2026 · Web4GoodCounterfactual interventions and panel re-scoring combine into a single trust score that decides whether a knowledge-graph RAG answer should be trusted.
WWW 2026A k3s + Coder adapter layer between raw provisioned compute and interactive research workspaces, with a CI/CD pipeline that ships projects from a git push to a running workspace in under 5 minutes.
arXiv Preprint · arXiv:2603.23942Applies causal prompting to identify and correct implicit sentiment bias in large language model outputs.
IEEE Trans. Computational Social SystemsTeam
