Julia Toolkit with fairness metrics and bias mitigation algorithms

Streamlit + Python pipeline that benchmarks gym workout efficiency (kcal/min) using present sessions only. Generates sortable workout-type benchmarks, distribution plots, fairness-aware gap analysis with uncertainty/low-sample flags, and a data-quality report to prevent misleading comparisons.

Bias-audit artifacts for automated employment decision tools: NYC LL144 selection/scoring rates and impact ratios, EEOC four-fifths adverse-impact flags, and a score-traceability schema.

Open-source AI ATS — LLM resume screening, CV parsing, JD↔candidate matching & bias-audited ranking. Self-hostable, privacy-first, FastAPI.

On-device AI medical triage for community health workers — offline-first PWA with Gemma 4, QR sharing, async telemedicine, AI explainability, 200+ tests

Computational audit of U.S. immigration enforcement data: spatial analysis and pipeline bias estimation on ICE arrest records.

ML for emergency department triage: MIMIC-IV-ED evaluation, SHAP, threshold optimisation, bias audit

Reproducibility code for the PhenotypeCatalog dataset — Wikipedia-sourced per-image phenotype observations across 239 ethnic groups

Ten bias-audit instruments over ten frontier models: detection is robust, model rankings are not. Replication materials.

Forensic audit of persona-driven bias in RAG-based LLM agents — 1,120 evaluations across 4 agents, 4 ablation conditions, and 70 tax-compliance scenarios. Shows persona prompts override injected rules (κ = −0.156). Mistral-7B · FAISS · SentenceBERT.

End-to-end responsible-AI audit of a DistilBERT toxicity classifier: identity-group bias audit, adversarial attacks, mitigation techniques, and a production guardrail pipeline. Built on the Jigsaw Unintended Bias in Toxicity Classification dataset.

Reproducible fairness audit for clinical LLMs — the Geographic Disparity Index (GDI). Deterministic perturbation, multi-provider harness, Wilcoxon+BCa stats, FastAPI + Streamlit.

Student-dropout prediction with a fairness audit and a GenAI explanation layer. A mistake-driven ML capstone on the UCI Portalegre dataset.

NYC Local Law 144 bias audit + EU AI Act bias-check. Part of the CSOAI Layer-0: 8 protocols · 100/100 A+++++ · world-leading.

Bias-audited age estimation with presentation attack detection — PyTorch + ResNet-50 + ordinal regression, served as INT8 ONNX behind FastAPI, deployed on Hetzner with Docker Compose and Nginx

Visual Archetype Analysis of Geodemographic Clusters using Vision Language Models

AI-agnostic research workflow for steelmanning opposing views, auditing sources, and reaching evidence-grounded conclusions

LLM-augmented geodemographic classification of Indian districts using NFHS-5 dataset

Computational audit of U.S. immigration enforcement data: spatial analysis and pipeline bias estimation on ICE arrest records.