MACHINE LEARNING / DATA ANALYSIS / STATISTICAL MODELING
Understanding problems with data,
putting answers into practice through analysis and engineering.
My background is in data science and biostatistics. I work on machine learning, language model applications and statistical modeling. I use Python, PyTorch and SQL to prepare data and build models, then evaluate them through controlled comparisons and error analysis. My projects span clinical risk prediction, marketing strategy and financial data analysis.
PROJECTS & RESEARCH
I fine-tuned Qwen3-4B with QLoRA and combined structural supervision, constrained decoding and one repair attempt. On 100 BIRD development questions, execution accuracy rose from 36% to 44% and the share of executable queries from 71% to 77%.
2026.09 — 2026.10 · Qwen3-4B / BIRD
I estimated shared and lead-specific risk components from 12-lead ECGs and traced their contributions over time. Across roughly 155,000 recordings, Directional reduced the mean component count from 72 to 36.5, with a mean C-index of 0.7460.
2026.05 — 2026.07 · CODE-15 twelve-lead ECGs
I combined follow-up records from 5,368 breast cancer patients, using only events observed by each prediction time. A Transformer encodes variable-length event histories and a Cox loss trains the model to estimate subsequent risk.
2025.04 — 2025.08 · Longitudinal event modeling · Patient-level isolation
Combine SLIC, ResNet18 and training-fold PCA/GMM to build a 102-dimensional patient representation with 50 spatial and longitudinal features; evaluate cross-center generalization in 827 patients.
2026.06 — Present · Exploratory research · No final model selected
Provide aggregate data for a five-page interactive dashboard, from incremental events and reconciliation to daily KPI updates.
2026.07 — 2026.09 · Synthetic data · 19 tests
Identify incremental purchase opportunities with a regularized DR-learner under a 30% sending budget; select C1 on validation data and compare it with random sending and response prediction.
2026.07 — 2026.09 · 200,039 customers · Retrospective exploration