I'm a rising senior at the University of Maryland studying Computer Science with a machine learning concentration and a business minor. I'm interested in applying AI to solve real-world problems in the software, business, and financial sectors. I enjoy learning about new topics and technologies in the machine learning and scientific fields, and I leverage that curiosity to build things that work.
Incoming internship.
Engineered a soft bounce email management system designed to protect email credibility, including internal SQL databases and Python scripts that synchronized systems through SendGrid API calls.
Mentored students in the Climate Computing stream of FIRE, supporting research development, data analysis, modeling, and presentation of undergraduate research posters.
Researched relationships between wave activity flux and Arctic weather patterns, writing Python scripts for simulations and analysis on Derecho, the NCAR supercomputer.
Interactive 3D globe mapping the global AI supply chain as a directed, layered graph — power grids, fabs, advanced packaging, data centers, and model labs. Constraints like CoWoS capacity and grid-equipment lead times are first-class entities: select one and everything downstream lights up. Every claim traces to a public, dated source.
An AI-powered operating manual for an HVAC parts distributor. Ingests messy real-world documents — manufacturer bulletins, regulatory filings, supplier emails, SOPs — into a citation-grounded knowledge base queryable in natural language. Measured on a 30-question golden set: 0.97 recall@5 hybrid retrieval and 0.98 citation recall from the agent loop.
A financial intelligence platform pairing a citation-validated RAG system over SEC filings (10-K/10-Q/8-K) with a real-time implied-volatility workstation — surfaces, skew, term structure, and expected moves from live option chains. Retrieval and answer quality are measured with a real eval harness (Recall@k, MRR, NDCG), and every value carries explicit provenance.
A Chrome extension that turns technical YouTube videos into a just-in-time glossary. Structured LLM extraction over transcripts, strict JSON schema validation, quote-based timestamp resolution, and cache-first storage render playback-synced concept cards inside the watch page — with evaluation tooling for precision, recall, and timestamp coverage.
End-to-end ML platform predicting professor ratings from PlanetTerp review data. Extracts 60+ text, sentiment, grade, and course features; benchmarks linear, tree, and boosting models with tuning (best holdout R² 0.81); and serves saved experiment runs, metrics, and predictions through a FastAPI backend and React dashboard.