Thirty years of FOMC transcripts, a million protein decoys, a stack of payer policy PDFs. Different domains, same job: figure out how to measure the thing before you model it.
I finished a BA in Mathematics and Economics at Hamilton College in May 2026 (Honors in Economics, Soper Research Prize) and started an MS in Statistics at Rutgers this fall. I am open to leaving the program for the right full-time research or analyst role, and I am equally happy to talk about Summer 2027.
My main work so far is a paper with Prof. Ann Owen, a former Fed economist, that uses LLMs to read every FOMC participant's policy stance out of the transcripts and then asks who actually moves the room. I also built a graph transformer for protein structure scoring, shipped an iOS app, and ported a Stata econometrics command to Python because I needed it and it did not exist.
When I am not at a keyboard I am usually at a piano or a basketball hoop.
Formal FOMC votes almost never dissent, so the votes hide most of the disagreement in the room. We read the transcripts instead. A multi-model LLM pipeline (Claude and GPT-4o) scores every participant's economic assessment and policy preference at every meeting from 1990 to 2020, giving a stance measure for the seven non-voting Reserve Bank presidents as well as the voters.
Two findings. Regional economic conditions shape how presidents talk about policy, not just how they vote. And non-voters matter: their discussion is associated with voters revising their views within the meeting, with recorded dissents, and with the position the Chair carries into the next meeting's proposal.
Before running any regression I treated the measurement as its own problem. Early scoring versions were not accurate enough to trust, so the pipeline was rebuilt and validated against two independent benchmarks before touching the panel.
A 12-layer gated graph transformer that predicts per-residue lDDT for protein decoys, combining ESM-C sequence embeddings with Rosetta energy terms as node and edge features. Trained on more than a million decoy structures on Hamilton's HPC cluster (multi-GPU, NVIDIA A100) and submitted to the CASP17 benchmark.
My part was the experiment design and the training infrastructure: data loaders that did not choke on a million graphs, checkpointing, and keeping runs reproducible across nodes.
Python port of Stata's xtscc: Driscoll-Kraay standard errors for panels with cross-sectional and serial dependence (Hoechle 2007). Written because I needed it for the FOMC paper and could not find a faithful implementation. Validated against Stata output and against Bobrov et al. (2025).
iOS app that finds books through real endorsements rather than ratings. SwiftUI front end on a FastAPI backend running a multi-stage GPT-4.1 pipeline: parallel agents search for endorsements, a verifier agent plus deterministic filters reject false positives. Persistent caching cut repeat-query cost to near zero; adds cost telemetry, rate limiting, and server-side auth. Monetized with AdMob and Amazon affiliates.
Pulls CPT and ICD-10 procedure and diagnosis pairs, with their medical-necessity status, out of semi-structured payer policy PDFs. LangChain, GPT-4, and Pydantic schemas for typed output; deployed as a Gradio app on Hugging Face. The kind of document a health insurer's analytics team spends a lot of hours reading by hand.
RAG course recommender for Hamilton. An ETL pipeline harvests course APIs, department pages, and PDF syllabi (session-authenticated) into one dataset; a GPT-4 Turbo recommender sits on top with rule-based PDF chunking, custom vector search, and conversation memory. Demoed live to faculty, 3.8/5 satisfaction.
Solution to "Pent-Up Frustration 3." Posted the code, not just the answer.
Bi-LSTM sentiment model with a capacity-scaling study (256 to 2048 dims), 91% on IMDB 50K. UNet++ segmentation with deep supervision, a hybrid loss, and CBAM attention, 80.9% mIoU on Oxford-IIIT Pet. Whisper fine-tune on Hugging Face.
Thirty-odd repos in total, including a job-hunting agent built on the Anthropic SDK, healthcare provider (NPI/PECOS) queries, and an NHS symptom scraper. All of it is on GitHub.
Started at five, and studied with Rita Shklar from 2014 through high school. If I get one piece, it is Ginastera's Danzas Argentinas. In 2020 I placed 3rd at the ENKOR International Competition and took 1st and 2nd places in the Great Composer Competition series, and I have played in master classes with Natasha Paremski and Daniel Epstein.
I painted through high school, mostly oil and gouache, mostly ordinary scenes: breakfast, a barbershop in Princeton, my dad reading at night. Fish Market won a Scholastic Art & Writing Gold Key. I have not picked up a brush since, but a few of the pieces still hang around, so they are on the right.
Co-president of the Hamilton College Basketball Club, 2025 to 2026. Not a competitive player; I just like the repetition of shooting, and running a club taught me more about scheduling people than any course did.
English, Mandarin (fluent speaking, still working on reading and writing), and intermediate Spanish, maintained mostly through a Spanish playlist I started during lockdown.
If you are hiring for a data or research analyst role and any of the above looks useful, I would like to hear from you. I am a US citizen and do not need sponsorship. I reply to email.