Independent research and evaluation frameworks for responsible AI deployment.
Built by Preethi Raghuveeran — AI governance practitioner focused on operationalising fairness, safety, and compliance in production AI systems.
Research into how physically proximate adversaries defeat AI-enabled women's safety systems — and which governance frameworks fail to prevent this.
Produces a 4-layer threat taxonomy × 5 attack methods × 6 adversary access levels, a scenario library of 13 grounded Indian-context scenarios, governance gap analysis across EU AI Act / NIST AI RMF / ISO 42001 / India DPDP, and a novel evaluation metric: Zero-Interaction Detection Rate (ZIDR).
Finding: No existing governance framework requires passive-detection-only robustness testing. Every framework evaluates AI systems against cooperative users. ZIDR names and measures the gap.
Zenodo DOI: 10.5281/zenodo.20028247
Investigates whether model quantization disproportionately harms already-disadvantaged groups in high-stakes prediction.
Trained RecidivismNet (740 parameters) on the COMPAS dataset and measured racial fairness at FP32, INT8, and INT4 precision. Conducted neuron-level attribution to identify the mechanism.
Finding: The 14.4 percentage point false positive rate disparity between Black and Other defendants is invariant to quantization — it persists at all three precision levels. Bias is encoded in high-magnitude neurons, which are inherently more quantization-stable. Making a model smaller does not make it fairer.
A safety-constrained routing system that selects the smallest model capable of handling each query — reducing carbon cost without degrading output quality below threshold.
Eval results (50 test cases, 2026-05-02): 45.5% carbon reduction vs always-large baseline · routing precision 100%/90%/100% across complexity tiers · P95 latency 0.09ms · 100% fallback reliability.
Zenodo preprint DOI: 10.5281/zenodo.19934621
Real-time monitoring system for autonomous agent goal drift. Scores each agent step against the original task intent using semantic similarity; flags DRIFTING (score 0.62–0.77) and ROGUE (score < 0.62) states with human-in-the-loop escalation.
Eval results: All 3 evaluation scenarios passing — safe task, goal-drifting task, and rogue exfiltration task correctly classified.
Demonstration of adversarial goal hijacking in a multi-step autonomous agent pipeline. Shows how an injected secondary objective (exfiltrate records, write cover report) can be introduced mid-task without triggering standard output filters.
Includes Phase 2 detection layer and human-in-the-loop review workflow.
A structured deployment evaluation framework for the COMPAS recidivism risk scoring tool, used in bail, sentencing, and parole decisions across the US.
Finding: COMPAS is deployment-blocked under EU AI Act, NIST AI RMF, and US 4/5ths rule — 4 critical threshold breaches including a Disparate Impact Ratio of 1.92× against a 1.25× legal limit. Replicates and extends ProPublica's 2016 investigation within 2–3% of original findings.
Applies the same structured evaluation framework to AI-based credit scoring systems. Covers fairness metrics, regulatory pass/fail thresholds (EU AI Act, ECOA), escalation paths, and monitoring cadence for high-risk consumer lending decisions.
- EU AI Act (2024) — Annex III high-risk classification, Art. 10, Art. 12
- NIST AI Risk Management Framework — GOVERN 1.1, MAP 1.5, MEASURE 2.5
- ISO 42001 — AI management system requirements
- US EEOC 4/5ths Rule — 29 CFR Part 1607
- India Digital Personal Data Protection Act 2023
- Algorithmic Fairness Auditor — interactive tool: quantization bias + COMPAS + credit scoring audit
- LLM Safety Evaluation Framework — multi-vendor safety benchmark across 5 harm categories
- Carbon-Fairness Efficiency Frontier — quantifies the carbon cost of fairness decisions
- AI Governance Tracker — EU AI Act, NIST RMF, ISO 42001 clause-level compliance tracker