At CompQSoft, I built a semantic search system over government datasets using Sentence-BERT, Hugging Face, and nearest-neighbor vector search. I fine-tuned embedding models and evaluated retrieval relevance against similarity metrics and benchmarks.
I also built Python pipelines to clean, transform, validate, and index database records, and wrote SQL for extraction, aggregation, and data quality checks.
In my market-data project, I built an incremental lakehouse for 20 securities and trained a LightGBM volatility forecaster that beat the HAR-RV benchmark in all five out-of-sample years, with 15.5% lower QLIKE.
At Texas A&M University, I’m developing an evaluation framework for an LLM agent that updates probabilistic wildfire-ignition risk beliefs from spatiotemporal data. My work examines calibration, uncertainty quantification, and fuzzy-logic evidence fusion against ground-truth outcomes.
