Current

Founding Engineer

Robocurve, Inc. San Francisco, California | September 2026 - Present

  • First author on Robocurve’s reports evaluating GPT-6 Astra and Fable 5.1 on real robot-arm manipulation tasks.
  • Co-author on the RoboHarm study. Across seven frontier models, requests refused in text were mostly carried out once the model controlled a robot arm.
  • Co-author on the Inspect Robots paper. Gave feedback on library changes and helped develop the paper’s main arguments. The library has had nearly 100k downloads since its July release.
  • Run Robocurve’s open-source benchmarking program. Currently working with eight robotics teams on 50+ new benchmarks, to be released as a joint paper in Q1 2027.

Research Experience

Research Fellow (David Lindner stream)

MATS Berkeley, California | June 2026 - September 2026

  • Led a project on how agent harnesses affect safety: 5 frontier models, 5 production harnesses, 5 behavioral domains.
  • Found that the same model’s misalignment rate shifted by up to 67 percentage points depending on the harness.
  • Ablated system prompts, tool schemas, and loop structure to find which parts of the scaffold drive the effect. Paper under review.

Research Fellow (Santiago Aranguri stream)

Supervised Program for Alignment Research (SPAR) Remote | February 2026 - May 2026

  • Stress-tested the realism win rate, a metric for how realistic an evaluation looks. Transcript position alone moved results by up to 43 points.
  • Found that win rates and absolute realism scores from the same judge often disagree.

Independent Researcher

AI Safety Research Collaboration Remote | December 2025 - February 2026

  • Built an evaluation harness using OpenAI and Anthropic APIs to simulate reinforcement learning training.
  • Evaluated frontier models on their propensity to resist RL training under misaligned goals. This work became Exploration Hacking, published at ICML 2026.

Research Fellow (Diogo Cruz and Eyon Jang stream)

Supervised Program for Alignment Research (SPAR) Remote | September 2025 - December 2025

  • Developed evaluation pipelines to quantify goal drift in language model agents on long-horizon tasks.
  • Designed experiments to study how conditioning stronger models on weaker models’ context affects goal drift.
  • Led the writing and submission of my first-author paper, Inherited Goal Drift, to ICLR 2026 workshops.

Research Assistant

Trustworthy ML Lab La Jolla, California | June 2025 - August 2025

  • Investigated chain-of-thought (CoT) reasoning and faithfulness in large language models.
  • Developed and tested steering-vector methods for improving CoT faithfulness, achieving promising preliminary results on the “hint” setup from Chen et al. (2025).

Reviewing & Mentorship

  • Reviewer, ICLR 2027.
  • Mentor, SPAR (Fall 2026): five mentees working as technical program managers on Robocurve’s benchmarking program.

Industry Experience

Associate Data Scientist, Refinance Team

Jerry Services Inc. Chicago, Illinois | December 2023 - July 2024

  • Led the analysis of A/B tests and development of ML models for the refinancing product which reduced credit pull costs by 70% with no negative impact on loan volume.