Industry

Clinical software in use

AutoVS, a vital-sign collector for surgical wards at Queen Savang Vadhana Memorial Hospital. I built it, deployed it, and it has run on the wards since 2026.

• Headless hospital-page extraction, Google Sheets and Trello routing, and approval-gated writes: every comment needs an explicit review and a read-back check.
• Native Mac and iOS apps, a Windows build, and an install-free web app behind a passcode. Credentials stay on the device. This public bio is separate from that host.
• A signed LINE new-admission intake that routes exact attending matches to the attending's Trello board and holds unknown matches for review.
• Clinical Case Assistant: source-linked case review, grand-round drafts, and case chat.
• Stack: Rust and Node web services, Python (FastAPI, Playwright), Swift for iOS and macOS, launchd services on a Mac mini. Hand-built and self-hosted. Tooling: multi-agent LLM workflows for pipeline QA and code review with MLflow tracing, under one rule: no participant row leaves the cluster.
• See tools.dhupbh.com.

Industry experience

Data Scientist, The Gang Technology Co., Ltd., Bangkok, Thailand. October 2022 to December 2024. Data science consult for the Bureau of Epidemiology, Ministry of Public Health: the Thailand COVID-19 vaccine effectiveness project.

• Roles: part-time Data Scientist, October 2022 to December 2022. Data Scientist IV, December 2022 to September 2023. Part-time Data Scientist, September 2023 to December 2024, during medical study at Chulalongkorn University.
• Designed and ran a 1:4 matched case-control study (n=134,135; severe cases n=26,827) from a national COVID-19 dataset of more than 15 million records.
• Estimated vaccine efficacy with conditional logistic regression: univariate, multivariate, and stratified models. Tested more than 40 matched case-control prototypes with different match conditions, outcomes, and exposures.
• The odds of a severe outcome approximately halved with each added vaccine dose. Efficacy peaked 1 to 2 months after the latest dose. The Ministry of Public Health used these vaccine-effectiveness reports in its national booster-policy work.
• Built the statistical analysis pipeline in Python (Pandas, Statsmodels, Lifelines): case-control matching, model fits, stratification, and forest-plot reports.
• Ran a retrospective cohort study (n=138,264 of 156,574 screened) that compared a one-fifth-dose intradermal ChAdOx1 booster with intramuscular dosing (Kaplan-Meier and Cox models). The analysis did not detect a statistically significant difference between administration routes, adjusted HR 0.88, 95% CI 0.76 to 1.02. Last and corresponding author, International Journal of Infectious Diseases, 2024.
• Trained Ministry of Public Health staff on Python and vaccine-efficacy analysis.

Skills

Python. Unsupervised ML (Louvain, Leiden). Survival and longitudinal analysis (KM, Cox, CTMC). RFECV and SHAP. NumPyro. U-Net. ComBat and CovBat. scikit-learn. Optuna and MLflow. Streamlit. SLURM and Singularity on a university HPC cluster. Rust, FastAPI, Node, and Swift for clinical apps. Playwright. Cloudflare Workers. Multi-agent LLM workflows for pipeline QA with MLflow tracing.

Route

Academic readers, start at Publications. Industry roles: email me@dhupbh.com. One-page industry CV: /cv-industry.pdf.