Projects
A collection of things I've built: 10 projects, 2 awards, and 1 paper since 2025.
2nd place, OpenAI-sponsored Finance Agent track, AgentX–AgentBeats (Berkeley RDI)
Data Matchmaker Benchmark AI Agent
Dec 2025 – Jan 2026
Evaluator agent that scores other AI agents' accuracy on data integration tasks. Team project; I helped deploy and test it.
3rd of 50+ teams, UC Berkeley 7th Annual Datathon for Social Good
EquiMatch
Nov 2025
College recommender for underserved students across 6,000+ institutions, matching on 10 equity factors like child care, with an AI counselor that explains each match.
Accepted to the NeurIPS 2026 GenAI4Health Workshop
CohortFlow
May 2026 – Sep 2026
Co-authored with UCSF's Yang Lab. An AI agent that finds trial-eligible patients and lets clinicians change criteria mid-run without restarting, cutting wait time up to 64%.
ByT5 Akkadian OCR Correction
Jan 2026 – Present
Fine-tuned a byte-level model to fix OCR errors in scanned Akkadian transliterations, cutting character errors 74%. Now defining requirements to make it a scalable autocorrect tool for researchers, through Berkeley's Data Discovery program.
Acne Detection
Apr 2026 – May 2026
My classifier hit 98.7% on its own photos but lost to a coin flip on new ones, fooled by watermarks and backgrounds. Just 20 new photos got it above chance.
CohortEx
Jun 2026 – Jul 2026
Automated pipeline that reads biomedical papers and extracts who was in the study: conditions, age range, sex, and eligibility rules, as structured data.
Efficient Frontier
Dec 2025
Stock Time-Series Analysis & Portfolio Modeling. Data science and machine learning applications to financial market analysis using stock data from Yahoo Finance.
CapsuleDx
Dec 2025
Capsule Network project using chest X-ray image dataset to develop a classification model that can accurately distinguish between COVID-19 positive and negative cases.
LyricCal
Sep 2025 – Dec 2025
Personalized song recommender that analyzes song mood and acoustic features to find mathematically most similar tracks.
