AI Safety Researcher · Data Scientist · Physics PhD
I do empirical AI-safety research on reasoning models and build machine-learning and document-intelligence systems, with a research scientist's habit of asking: What can fail, how do we know, and how do we measure it? My work connects physics, careful evaluation, and practical AI tools.
Focus
My background moves from physics and numerical simulation into applied machine learning and, most recently, empirical AI safety. The connecting idea is the same: build models, test their limits, and explain when they should be trusted.
Chain-of-thought faithfulness, answer commitment, pre-registered causal experiments on reasoning traces, prefix interventions, resampling studies, and evidence about what visible reasoning actually reveals.
Classifier ambiguity, calibration, threshold-dependent behavior, finite-horizon learning-curve prediction, conservative extrapolation, experiment tracking, and reproducible benchmarks.
PDF parsing, structured extraction, native text-layer workflows, retrieval-augmented question answering, hallucination reduction, and evaluation of extraction quality.
Selected work
These projects are designed to be more than portfolio demos. They test reliability, expose failure modes, and connect mathematical ideas to usable AI systems.
A pre-registered causal study of chain-of-thought faithfulness in a reasoning model: prefix interventions and 500 resampled continuations test whether visible reasoning reveals when the final answer becomes determined.
View projectA diagnostic toolkit for evaluating binary classifiers beyond a single threshold, with ambiguity-focused metrics, calibration analysis, confidence intervals, and statistical testing.
View projectA reproducible benchmark for finite-horizon gradient-boosting learning-curve prediction across many convergence regimes, noise levels, and extrapolation methods.
View projectA pharmaceutical document-intelligence and RAG project that combines PDF extraction, structured information retrieval, document question answering, and evaluation workflows.
View projectA lumber product-name normalization engine that maps inconsistent seller strings to canonical structured outputs using fuzzy matching, dimension extraction, treatment detection, confidence scoring, a Streamlit interface, and chatbot-style explanations.
View projectAn end-to-end MLOps pipeline with DVC, MLflow, pytest, GitHub Actions, model-performance checks, and drift monitoring with Evidently.
View projectResearch
My research background spans quantum materials, crystal-field theory, noncollinear magnons, molecular and defect-mediated transport, dark matter phenomenology, modern ML diagnostics, and, most recently, empirical AI safety on reasoning models.
Experience
My path combines academic research, teaching, scientific software, and project-based machine-learning training.
Developing Decided Mid-Thought, a pre-registered study of chain-of-thought faithfulness and answer commitment in reasoning models.
Taught undergraduate physics, developed optics and electronics lab materials, supervised teaching assistants, and integrated AI tools into scientific workflows.
Completed project-based training across supervised learning, unsupervised learning, deep learning, NLP, computer vision, SQL, Git, and end-to-end ML workflows.
Built retrieval-augmented and document-intelligence workflows for healthcare documents using native PDF extraction, LLMs, and RAG architectures.
Developed high-dimensional numerical models, optimized scientific simulations, analyzed noisy data, and contributed to peer-reviewed work in condensed matter physics and quantum materials.
Built simulation-based models using Python and Fortran, including N-body simulation, molecular dynamics, and parallelized scientific computing workflows.
Tools
Scikit-Learn, XGBoost, PyTorch, Transformers, LLM APIs, reasoning-trace analysis, calibration, model evaluation, NLP, RAG, MLflow.
Python, SQL, FastAPI, Pydantic, Docker, Linux/Bash, Git, GitHub Actions, Jenkins, CI/CD, pytest.
NumPy, Pandas, Polars, Numba, Fortran, HPC, MPI, numerical optimization, Monte Carlo simulation.
PDF parsing, information extraction, OCR-aware workflows, structured extraction, evaluation of extraction quality.