Projects
AI Safety
Published Research
▶Multi-Turn Jailbreaks
Research project investigating adversarial attacks on large language models through multi-turn conversations. As Research Lead of a 6-person team at AI Safety Camp, I explored vulnerabilities in conversational AI systems and developed methods to understand and mitigate jailbreak techniques.
Published at COLM SoLaR Workshop, 2025
LLM Unlearning Evaluation Framework
Comprehensive framework for evaluating machine unlearning in Large Language Models. As mentor of a 6-person team at SPAR, I established standardized metrics and methodologies for assessing the effectiveness of unlearning techniques in removing specific knowledge from trained models.
Published at COLM SoLaR Workshop, 2025
Chess Neural Networks Analysis (Leela Interpretability)
Research project analyzing learned look-ahead behavior in chess neural networks, based on "Evidence of Learned Look-Ahead in a Chess-Playing Neural Network". Investigating how neural networks develop and utilize predictive capabilities in strategic gameplay.
Published in TMLR, 2025
Reinforcement Learning Inductive Biases
Research investigating how RLHF training creates bias in language models toward learning "easy to extract" features. This work demonstrates that while reward models can easily identify certain features, this constrains the sophistication of the model's reasoning capabilities.
Last updated: November 2023
Long-Context Agent Behavior (Algoverse 2025)
Research investigating how extended context windows affect LLM agent safety behaviors. Led team of 6 mentees exploring whether long contexts lead to emergent safety issues or capability changes in autonomous agents. Published as AAAI 2026 TrustAgent workshop paper.
Published at AAAI 2026 TrustAgent Workshop
Robust Probe Training with Limited Labeled Samples
Investigated methods to train more robust and performant probes when the amount of labeled samples is limited, particularly for hard-to-classify or superhuman tasks. CHAI internship work at UC Berkeley. The project produced negative results; write-up forthcoming.
CHAI internship completed October 2025; write-up of negative results forthcoming
Goal Drift in Long-Horizon Tasks
Investigated how LLM agents gradually abandon or modify their original goals during extended tasks such as stock trading and hospital triage simulations. Key findings: model family matters significantly—OpenAI models follow instruction hierarchy well, reasoning-enabled models stay on track better, and models can self-cause drift in agentic coding setups without external pressure. Two ICLR workshop papers accepted, with arXiv submissions and a third paper in progress.
2 ICLR 2026 workshop papers accepted; third paper in progress
Tribalism in Fine-tuned Models
Investigated whether fine-tuning can accidentally induce tribal behavior in language models. Originally framed as studying anti-sycophancy, the results revealed that models develop selective preferences that generalize beyond their training context, a potentially concerning pattern of group-based selective behavior.
LessWrong blog post published, 2026
Ongoing Projects
▶Tool Failure and Deceptive Behavior (SPAR Fall 2025)
Investigating how LLM agents behave when tools become unreliable, testing whether this triggers deception, reward hacking, or unauthorized actions. Key finding: GPT-5 models showed drastically increased reward hacking when encountering tool failures, even when deferring to a human was an option.
Write-up phase; COLM main conference submission in progress, March 2026
Agentic Situational Awareness (SPAR Spring 2026)
Continuation of the Algoverse project on operationalizing situational awareness in LLM agents. Found serious quality issues in DeepMind's agentic situational awareness benchmark—tasks where humans wouldn't score well either, yet the benchmark appeared in the Gemini 3 Pro system card and International AI Safety Report. Building improved benchmarks.
In progress, SPAR Spring 2026
Efficient Benchmarking for Safety Evaluations (SPAR Spring 2026)
Applying Fluid Benchmarking (IRT + adaptive testing) to safety evaluations. The original approach showed 50x efficiency gains on MMLU; targeting expensive agent benchmarks like OS-HARM ($53 and 5 hours per run). Can also help identify uninformative and broken benchmark tasks.
In progress, SPAR Spring 2026
Alloy Agents
Solo independent project combining high-capability/low-propensity models with low-capability/high-propensity models on Control Arena. Exploring whether mixing models from different providers can achieve both strong capability and strong safety properties.
In progress, March 2026
Tools & Utilities
▶AI Safety Scraper
Web scraping tool specifically designed for collecting and organizing AI safety research papers and resources. Automates the process of gathering relevant publications from various academic sources and repositories.
Last updated: January 2025
Random
▶Summarizer
A Python-based text summarization tool designed for processing and condensing large documents. Features automatic text extraction, intelligent summarization algorithms, and customizable output formats for research and academic use.
Last updated: June 2025
OpenRouter Trends
Analysis tool for tracking and visualizing trends in AI model usage and performance across different platforms. Provides insights into model adoption patterns and comparative performance metrics.
Last updated: June 2025
Alignment Faking MVP
Minimum viable product exploring alignment faking behaviors in AI systems. Research prototype investigating how models might appear aligned while pursuing different objectives internally.
Last updated: June 2025
Literature Review System
Automated system for conducting systematic literature reviews in AI and machine learning research. Features paper discovery, categorization, and analysis tools to streamline the research process.
Last updated: January 2025
Quantum Computing
PhD Projects
▶QGRAND
Quantum computing project focused on advanced quantum algorithms and implementations. Explores novel approaches to quantum computation with applications in optimization and machine learning.
Last updated: July 2024
Hackathons
▶Quantinuum iQuHACK 2024
Participation in the Quantinuum iQuHACK 2024 in-person challenge, focusing on quantum computing applications and algorithm development. Collaborative project exploring cutting-edge quantum technologies.
Last updated: February 2024
QCHACK 2022
Participation in the Quantum Coalition Hackathon 2022, developing innovative quantum computing solutions. This hackathon project focuses on practical applications of quantum algorithms and technologies.
Last updated: April 2022
Microsoft x IonQ Challenge 2022
Participation in the MIT iQuHACK 2022 Microsoft x IonQ Challenge, focusing on quantum computing applications using IonQ's trapped-ion quantum computers. This project explores quantum algorithms and their practical implementations.
Last updated: January 2022
QuGreen - Quantum Spectral Functions
Repository for the Spectral Function Team in Qiskit Hackathon Europe, focusing on quantum algorithms for computing Green's functions and spectral properties of quantum systems.
Last updated: June 2021
Q-CTRL QCHACK Atomistic
Repository for the QCHACK 2021 Hackathon focusing on quantum control applications in atomistic systems. This project explores quantum control techniques for manipulating atomic and molecular systems.
Last updated: April 2021
patoLogic iQuHACK 2021
Team patoLogic project from iQuHACK 2021, focusing on quantum computing applications and algorithm development. This collaborative hackathon project explores innovative quantum solutions.
Last updated: January 2021
Miscellaneous
Course Projects
▶CS231n Course Implementation
Complete implementation of Stanford's CS231n: Convolutional Neural Networks for Visual Recognition course assignments. Covers fundamental deep learning concepts including CNNs, RNNs, and advanced optimization techniques.
Last updated: May 2023
Learning from Data Course
Implementation of assignments from Caltech's "Learning from Data" course, covering fundamental machine learning theory and applications. Includes practical implementations of various learning algorithms and statistical methods.
Last updated: May 2023