# Seon Min Kim > Four years in drug regulatory affairs taught me to ask what evidence a claim needs. Now I build health AI and data pipelines with that question designed in. Open to roles in the US, Singapore and Korea. ## Tracks - Applied AI / ML: Building LLM and ML systems, deciding what the model may and may not decide, and testing whether results hold up. Résumé: https://estherkim97.github.io/resumes/seonmin-kim-applied-ai.pdf - Health Data Science & Bioinformatics: Cleaning large raw datasets end to end, integrating multi-omics data, and drawing statistically sound conclusions. Résumé: https://estherkim97.github.io/resumes/seonmin-kim-health-data-science.pdf ## Regulatory & safety intelligence - [GLP1-Sentinel](https://estherkim97.github.io/projects/glp1-sentinel): A signal detection pipeline over 20 years of raw FDA adverse event reports that reproduces known GLP-1 safety signals. Key result: 20.9M (FAERS reports, 2005 Q2 to 2024 Q3). - [FDA Pathway Intelligence](https://estherkim97.github.io/projects/fda-pathway-intelligence): Scores 510(k), De Novo and PMA pathways from public FDA data, ranks predicate devices, and explains its evidence and uncertainty. Key result: 9 (Example products with expected pathways defined in advance). - [SourceLock](https://estherkim97.github.io/projects/sourcelock): Sentence-level evidence checks for pharma medical, legal and regulatory (MLR) review. Rules keep final authority; the model only matches meaning and suggests edits. Key result: 5 (Verdict types for every sentence). ## Clinical AI - [Rare Disease Phenotype Agent](https://estherkim97.github.io/projects/rare-disease-agent): Only the first step uses an LLM, and it cannot name a disease. Everything after is lookup against curated data, so the ranking cannot be made up. Key result: 72.3% (Top-1 accuracy, 669 real patients, ranking stage alone). - [Clinical Trial Eligibility Extraction](https://estherkim97.github.io/projects/clinical-trial-eligibility): An NLP pipeline that pulls eligibility criteria from ClinicalTrials.gov so protocol benchmarking takes minutes instead of weeks. Key result: 200+ (Medical entities extracted from 40 trials). - [DigitalTwin](https://estherkim97.github.io/projects/digitaltwin): A web app that screens for respiratory disease from cough recordings. Built by a team at the Berkeley AI Hackathon. ## Biomedical research & discovery - [Kidney transplant multi-omics](https://estherkim97.github.io/projects/kidney-transplant-multiomics): Brought four to five institutions together and integrated CyTOF, metabolomics and clinical data from kidney transplant patients. Findings are confidential pending a patent application. Key result: 28 (Patients). - [Scleroderma lung disease scRNA-seq](https://estherkim97.github.io/projects/pulmonary-fibrosis-scrnaseq): Single-cell analysis of about 120,000 cells from 30 samples of scleroderma-associated interstitial lung disease, mapping cell–cell signaling with CellChat. Key result: ~120K (Cells across 30 samples). - [NVIDIA × scverse GPU omics](https://estherkim97.github.io/projects/nvidia-scverse-gpu-omics): Single-cell RNA and ATAC analysis on GPUs with RAPIDS, cuML and cuGraph. Invited participant, cuGraph lead. Key result: 13–26× (Faster than CPU, on up to 16,000 cells). - [Prompt2Pump](https://estherkim97.github.io/projects/prompt2pump): Compares candidate drugs to exercise-mimicking signals using molecular embeddings and LINCS gene-set matching. 1st place, Track 5. Key result: 1st (Track 5, Stanford Multi-omics Hackathon 2026). - [MaxBind-AI](https://estherkim97.github.io/projects/maxbind-ai): Built the training dataset from ChEMBL, PLINDER and PDBBind for a team project that placed 2nd at the UCSF QBI Hackathon and received follow-on funding. 2nd place. Key result: $15K (Follow-on funding). - [BioGraph Intelligence](https://estherkim97.github.io/projects/biograph-intelligence): From a PubMed search to extracted claims, support and conflict checks, research gaps and a knowledge graph, running on a local LLM. - [AuData](https://estherkim97.github.io/projects/audata): A research-integrity auditor that checks a single paper for statistical errors, numerical inconsistencies, figure manipulation and citation problems, with a human reviewing every flag. - [SeaTRAC TB Hack Day](https://estherkim97.github.io/projects/seatrac-tb-hackday): A one-day, multi-site hack day on tuberculosis clinical trial data, profiling how patients respond to drug treatment. ## Experimentation & analytics - [A/B testing studies](https://estherkim97.github.io/projects/ab-testing-studies): A simulation where the naive result reversed once measured on the full population, and a dashboard over more than 1.2 million sensor records. Key result: 6.4% → 14.4% (Naive comparison, which reversed to 19.1% → 11.6% on the full population). - [Medication adherence risk profiling](https://estherkim97.github.io/projects/patient-adherence): Finds the patient groups most likely to stop taking their medication, so adherence programs know where to start. Logistic regression plus an R Shiny dashboard. ## Experience - 2025.06 – 2026.03: AI health technology evaluation intern, UCSF Innovation Ventures (confidential) - 2025 summer: Catalyst intern, UCSF Innovation Ventures (MAC Copilot product profile) - 2025.06 – 2026.09: Researcher, Sarwal and Sirota labs, UCSF (kidney transplant multi-omics) - 2025.11 – 2026: Graduate researcher, Sirota and Yang labs, UCSF (scleroderma lung disease scRNA-seq) - 2025 summer: Educational assistant, DATASCI 202, UCSF (about 20 graduate students) - 2020.07 – 2024.08: Regulatory affairs specialist, Global RA team, Chong Kun Dang - 2019.01: Intern, SNP Genetics - 2018.01 – 2019.07: Undergraduate researcher, Sogang University ## Education - 2024.07 – 2026.09: MS Health Data Science, UCSF - 2021.02 – 2023.02: MS Pharmaceutical Sciences, Yonsei University - 2015.12 – 2020.02: BS Life Science, Sogang University Contact: https://www.linkedin.com/in/seon-min-kim