Aarushi Jaitly
Red-teaming frontier models
I work at the intersection of AI safety, policy, and product, building at the frontier of what actually matters.
What I do
Three ways I keep AI honest.
Red-Teaming & Evals
I break frontier models on purpose: adversarial testing across jailbreaks, bias, and safety to surface how they fail before anyone else does.
AI Policy & Governance
I translate technical risk into governance, mapping frameworks like the NIST AI RMF and EU AI Act into real recommendations for regulators and enterprises.
Building Safety Tools
I ship the tooling that makes AI accountable, from open-source red-team harnesses to a solo-built regulatory intelligence platform.
About
Hi, I'm Aarushi.
I work in AI safety.
Which mostly means I spend my days trying to break frontier models on purpose, red-teaming them for the ways they fail, then writing the policy and building the tools that make them safer before they reach the rest of us.
Researcher by training, founder by instinct, policy nerd by conviction. I live where AI capability meets human consequence, and the stakes are real.
First-Author Paper, IEEE ICHI 2026
Multi-agent clinical AI safety (submitted)
NIST AI Red-Teaming
Assessed 60+ vulnerabilities in production LLMs
Founder, Regulus AI
Solo-built regulatory intelligence platform
Cooper Fellow & Merit Scholar
Awards that funded my education at Carnegie Mellon
GAO Policy Researcher
5 congressional recommendations on data privacy
BlueDot Technical AI Safety
Alignment, interpretability & evals curriculum
Recognition
Things I'm proud of.
First-author paper submitted to IEEE ICHI 2026
Multi-agent AI frameworks for clinical diagnosis, funded by NIST (Federal Award 60NANB24D231) and CMU AIMSEC.
Co-authored a NIST standards proposal (IR 8214C)
Worked directly with Google, Coinbase, and Stanford on MPC cryptography standards.
Founded Regulus AI
A solo-built, AI-powered regulatory intelligence platform.
Cooper Fellow & Merit Scholar, Carnegie Mellon
Competitive awards that fully funded my graduate education.
Secured NIST-backed research funding with UPMC
To build DEDICATE, an 8-agent clinical diagnostic system.
Career
Experience
TrustModel.ai
CurrentAI Safety Researcher
Advise on go-to-market and sector strategy for an AI assurance platform, engaging federal, state, and healthcare stakeholders. Translate a 10-dimension, ~6-7M-prompt AI evaluation methodology (safety, bias, robustness, compliance) into board and product-ready deliverables that certify enterprise AI systems for risk and compliance.
Regulus AI
CurrentFounder
Founded a solo-built, AI-powered regulatory intelligence platform that monitors legislation, agency rulemaking, and enforcement guidance across jurisdictions and scores each signal by risk. Architected a RAG reasoning pipeline (Claude) generating audience-tailored executive briefs in seconds, replacing a 72-hour research lag and $150K+/yr in outside counsel.
Carnegie Mellon University
CurrentGraduate Researcher, Multi-Agent Clinical AI (DEDICATE)
Secured NIST-backed funding to build DEDICATE, an 8-agent LLM diagnostic system routing sleep-disorder cases. Architected a five-layer LangChain + ChromaDB RAG pipeline with a 224-persona generator, achieving 87.8% faithfulness, 94% retrieval accuracy, and a 100% safety pass rate on a physician-validated rubric across 1,000+ patient cases.
Government Accountability Office (GAO)
Privacy & AI Governance Researcher (Contract)
Led a CMU capstone research contract with GAO analyzing 10+ privacy and AI governance frameworks (HIPAA, GDPR, NIST AI RMF, FTC Act, COPPA, CCPA) and 16+ manufacturer policies. Synthesized 30+ sources and 15+ stakeholder interviews into a risk-assessment framework, authoring 5 congressional policy recommendations delivered to GAO leadership.
Fireblocks
Product Policy Intern
Benchmarked 6+ competitor custody platforms into a competitive brief presented to C-suite, shaping a $20B+ market-entry strategy. Mapped 4 regulatory frameworks into a risk/opportunity matrix, and worked directly with Google, Coinbase, and Stanford on a NIST standards proposal (IR 8214C).
KKR (via TresVista)
Investment Banking Analyst
Sourced and evaluated 100+ alternative-asset opportunities supporting $500M in deal flow. Built DCF, comparable-company, and underwriting models improving valuation accuracy by 7-15%.
KPMG
Junior Consultant
Built 50+ market assessments across geospatial, sustainable finance, and FMCG, identifying $30M+ in growth opportunities via TAM analysis and competitive positioning.
Grant Thornton
Consulting Intern
Designed a digital-transformation roadmap for 3 government services impacting 10K+ citizens, grounded in stakeholder interviews with program leads.
Academic
Education
Carnegie Mellon University
CurrentM.S. in Data Science & Public Policy Management · Concentration: Artificial Intelligence
Aug 2024 – May 2026
A Cooper Fellowship and a Merit Scholarship from Carnegie Mellon fully funded my graduate education, backing my move into AI safety research.
Symbiosis International University
B.Sc. in Economics (Hons.)
Jul 2020 – May 2023
Research
AI safety, on the record.
Peer-reviewed work and applied research across model evaluation, red-teaming, and AI governance.
Multi-Agent AI Frameworks for Clinical Diagnosis Support: Benchmarking LLM Reasoning with Sleep Disorders as a Testbed
Aarushi Jaitly, Helom Berhane, Deepa Burman MD, Anand Rao, Ramayya Krishnan, Rema Padman. Carnegie Mellon University & UPMC
This paper presents a multi-agent AI framework for clinical decision support using sleep disorder management as a controlled testbed. The framework introduces three components: (1) a combinatorial synthetic patient persona corpus of 224 profiles spanning clinically realistic comorbidity and etiology combinations; (2) 24-month longitudinal care pathway modeling across six clinically spaced episodes; and (3) a knowledge retrieval pipeline restricted to nine pre-approved medical sources, supporting a TEVV methodology. A pilot evaluation benchmarks five frontier LLMs (GPT-4o, Claude 3.5 Sonnet, Gemini 1.5 Pro, DeepSeek-V2, Llama 3 70B) in the Doctor Agent role. All five converged on plausible diagnoses, yet none replicated the differential diagnostic reasoning used by the physician benchmark, a gap that directly motivates the multi-agent architecture. Working paper submitted to IEEE ICHI 2026; funded by NIST (Federal Award ID 60NANB24D231) & CMU AIMSEC.
Read the paperApplied AI Safety & Policy
NIST AI Red-Teaming & Governance
AI Safety Researcher · Aug 2025 – Dec 2025
Assessed 60+ vulnerabilities in production LLMs through adversarial testing; identified policy non-compliances across bias mitigation, fairness, trust, safety, and explainability, translating findings into actionable NIST AI RMF recommendations.
GAO, Consumer Health Data Privacy
Privacy & AI Governance Researcher · Jan 2026 – May 2026
A CMU capstone contract for the U.S. Government Accountability Office: a multi-regime gap analysis across 10+ frameworks (HIPAA, GDPR, NIST AI RMF, CCPA) examining the biometric-data lifecycle in consumer wearables, producing 5 congressional recommendations on consent reform.
BlueDot Impact, Technical AI Safety
Fellow · 2025
Completed BlueDot Impact's 6-unit technical AI safety curriculum covering the alignment problem, training techniques for safer models, evaluations used by frontier labs (Anthropic, OpenAI, Google DeepMind, Meta), mechanistic interpretability, and harm minimization.
Work
Things I've built
Featured
Regulus AI
Solo-built regulatory intelligence platform monitoring legislation, agency rulemaking, and enforcement across jurisdictions, scoring each signal by risk. A RAG reasoning pipeline (Claude) turns a 72-hour research lag into executive briefs in seconds.
DEDICATE, Multi-Agent Clinical AI
NIST-funded, 8-agent LLM diagnostic system for sleep disorders, built with UPMC. A five-layer LangChain + ChromaDB RAG pipeline and a 224-persona generator reach 87.8% faithfulness, 94% retrieval, and a 100% safety pass rate across 1,000+ cases.
Open Source & Safety Tools
Agent Alignment Evaluation Workbench
activeFully local tool (Streamlit + Ollama/Qwen3) evaluating AI-agent alignment inside simulated org hierarchies, scoring responses on the RICE framework (Robustness, Interpretability, Controllability, Ethicality). Behavioral flagging auto-detects deception, power-seeking, and self-preservation.
LLM Red-Team Harness
activeOpen-source, provider-agnostic framework (OpenAI, Anthropic, HuggingFace) for adversarial LLM evaluation across jailbreak, prompt injection, privacy leakage, harmful instructions, misinformation, and bias, with a modular, severity-tiered scoring architecture.
Red-Team Policy Analyzer
activePython/SQL pipeline generating adversarial prompt variations across 5 jailbreak framing patterns and 15+ attack vectors spanning a 5-domain harm taxonomy, surfacing policy coverage gaps where over 30% of variations return unsafe.
ML Fairness Auditor
activePython/SQLite pipeline training and auditing three classifiers on the UCI Adult Income dataset (48,842 samples) for bias across sex, race, and age, operationalizing demographic parity, disparate impact, equalized odds, and calibration error with significance testing.
Skills & Training
Always leveling up.
Training modules completed
Tools & frameworks
Unit AI-safety fellowship