Sitemap

About me

Deep research tools from OpenAI, Google, and Perplexity promise source-grounded synthesis, but their reliability depends heavily on what they are searching. Professional tools connected to peer-reviewed databases like PubMed have a built-in quality gate: the barrier to publication is literally peer review. But for code repositories, blog posts, and other sources with no editorial barrier, deep research inherits every error, exaggeration, and fabrication in the source material. Digging into some of the recent literature myself, studies often seem to find citation accuracy as low as 40% and fabricated references. I tested all three tools on the same library-evaluation question as a small experiment to show how they can fail (à la proof by contradiction). One tool recommended an abandoned library as its top pick with an apparently fabricated release date. The issue of AI-generated content infiltrating peer review itself is a separate and important problem, but not within the scope of this post.

After reading Ang’s ‘Asset Management: A Systematic Approach to Factor Investing’ and rereading the newest 50th anniversary edition of Malkiel’s ‘A Random Walk Down Wall Street,’ I wanted to explore how funds like Dimensional’s DFUS address indexing inefficiencies without fully committing to factor tilts. The goal is to capture market beta while addressing index inefficiencies such as adverse selection. 5-factor regressions indicate DFUS is a VTI equivalent for US equities, but a VXUS equivalent remains unsolved. This is not financial advice and I am just your average retail investor.

Based on some great discussions with my previous Pfizer colleagues Brett South, Ajit Jadhav, Jay Ronquillo, Jon Mauer, and Stephen Watt, this project tests whether large language models (LLMs) can reproduce established, human-built oncology value frameworks such as the ISPOR Scorecard and the ASCO Value Framework. I compare three LLM approaches (a multi-agent system, a single-LLM pipeline, and retrieval-augmented generation) on data from ClinicalTrials.gov, PubMed, and OpenFDA, and check their output against the published reference values in Langdon et al. (2016). Later on I added a fourth, MOA-based multi-agent version to improve the synthesis step and make it easier to trace where each number in the final scorecard came from.

An R package and Shiny dashboard that bring Baltimore datasets into DataTrail, the Johns Hopkins Bloomberg School of Public Health’s data science training program, so students can practice on local data. This was my ScM thesis project.

Together with Tiffany Hsieh and Bowen Chen, I built this R Shiny app to give some intuition for survival analysis, from both the theory side and the applied side. We’d found a publicly available Moderna vaccine survival dataset on GitHub that we wanted to dig into, so the app first walks through the basic survival analysis ideas without much math and then applies them to that dataset.

This pediatric TBI prediction web-application is part of my honors thesis on imbalanced outcome pediatric patient mortality classification, where the best-performing C5.0 decision tree classifier trained on Synthetic Minority Over-sampling TEchnique (SMOTE) subsampled data is the predictive model implemented here. The clinicians we spoke with suggested two uses: cases where a clinician can’t make a diagnosis, and as an indirect aid for giving parents or guardians a rough survival estimate. I kept the app simple, fast, and portable on purpose.

My first aim in this stochastic models and simulation project was to understand the Metropolis-Hastings algorithm and a few Markov chain Monte Carlo (MCMC) diagnostics at a more intuitive, visual level, using plots that are both animated and interactive. The second was to present that compactly to people who are new to MCMC and to R. Markov chains lend themselves well to animation, since each step depends on the previous one and there are plenty of parameters worth watching evolve.

This project simulates multivariate data from three underlying linear models with different amounts of correlation among the predictors, then checks how well several variable selection methods do on each dataset when a linear regression is fit with the variables they pick. We simulate 50 datasets for each of the underlying models in section 7 of Tibshirani’s (1996) paper “Regression Shrinkage and Selection via the Lasso”, fit models using several variable selection and shrinkage methods, and compare them on the Akaike Information Criterion (AIC), the Bayesian Information Criterion (BIC), and adjusted R-squared.

This was my previous personal website, which I built mostly to get better at R Shiny and HTML. Shiny is really built for interactive visualizations and statistical analyses, and a mostly static site didn’t need any of that, so I eventually retired it in favor of this one.