ML Researcher — AI Safety, LLM Reliability & Interpretability
Berkeley, California
Builder–researcher focused on the reliability, robustness, and safety of large language models. Currently an Anthropic Fellow researching when activation-probe safety monitors fail and how to predict failure before deployment. Background spans statistical guarantees for LLM inference, large-scale data curation, and production ML reliability.
AI Safety Fellow, Anthropic Fellows Program
Aug 2026–present · Berkeley, CA
Safeguards-mentored research project studying when activation-probe safety monitors fail — mapping failure modes of linear probes for policy enforcement and predicting probe performance from the written policy itself.
ACL 2025 Findings · First author
Refined the 1.3T-token FineWeb-Edu corpus into a 10B-token astronomy subset; fine-tuning LLaMA-3-8B on the curated data improved MMLU astronomy from 69% to 76%. The method was also validated on law and medicine. Datasets, code, and model are open-sourced.
Working paper · 2025
A sequential LLM inference system that adaptively allocates compute based on confidence: calibrated uncertainty estimates, statistical early stopping, and methods for knowing when models should abstain.
Paper in preparation · SPAR, Spring 2026
Methods for increasing human–AI complementarity on harm-identification tasks: a learned (LoRA-based) confidence signal and learned routing for combining human and AI judgments. Designed the team's calibration protocol.
CORDA Democracy Fellowship · 2026
An open evaluation suite, built on inspect-ai, measuring how reliably LLMs answer civic questions — 1,214 graded answers across 8 evals and 9 scorers. Fellows built the persona-drift taxonomy, sycophancy and demographic-bias analyses, and a cross-model bias study on byte-identical candidate profiles differing only in policy direction.
See Now.
Mentor: Rishub Jain, Google DeepMind.
civic-evals, with BlueDot Impact.
Built a 135M-parameter Transformer reranker competitive with multi-billion-parameter baselines; contributed to work awarded SIGIR 2025 Spotlight; first-author publications in ACL Findings and SIGIR.
Trained domain-specific LLMs using selective pretraining and noise-filtered corpora; HPC-scale workflows; parameter-efficient tuning.