Health-data & AI policy · behavioral-science researcher · builder
I help build the data and AI backbone of a national health system — and write the research and the code behind it.
Junghwan Park directs the Division of Healthcare Data Development at Korea's Ministry of Health and Welfare, where he leads national strategy for health data and AI. He earned his PhD in public health at UC San Diego, where he built just-in-time adaptive interventions for physical activity using machine learning and control engineering. He also designs and ships AI and agent systems himself — from local-LLM research harnesses to deployed full-stack tools.
Advisor: Eric Hekler. Dissertation: Optimizing Adaptive Interventions for Physical Activity.
Planned the One Million Genomes Project, Korea's national bio big-data initiative to sequence and link genomic and clinical data at population scale. It launched in 2022 as a cornerstone of the country's precision-medicine strategy.
Built the backend, mobile app, and monitoring system for JustWalk, an NIH-funded system-identification experiment, then applied machine learning and Bayesian regression to identify the "just-in-time" states where an intervention is most likely to help someone move.
Ported Lunit's open L1-16B-A3B medical mixture-of-experts model to Apple MLX and ran a disciplined 2–8-bit quantization sweep, scoring accuracy, speed, and memory on MedQA, MedMCQA, and PubMedQA in English and Korean — a concrete study at the intersection of medical AI and on-device inference.
Led the planning team for APEC's 15th High-Level Meeting on Health and the Economy, hosted in Korea, and steered the negotiation of the Joint Ministerial Statement among member economies — advancing multilateral cooperation across the region.
A self-completing research system that snowball-searches the literature, deduplicates with embedding similarity, runs a frozen local note-generation pipeline, stores everything in a knowledge graph, and periodically synthesizes findings — aimed at a queryable corpus approaching a hundred thousand papers.
Doctoral research optimizing just-in-time adaptive interventions by incorporating idiographic, dynamic predictions of behavior. It combined system identification, machine learning, and control engineering to decide when a digital coach should act to support physical activity.
A hierarchical multi-agent "thinking system" in which ephemeral agents reason continuously on small local models, passing context across generations in a pass-the-baton loop. An exploration of sustained, low-cost machine reasoning rather than one-shot prompting.
Led a cross-ministerial team developing Korea's response to U.S. pharmaceutical trade actions and a national biosecurity strategy, integrating twelve government ministries into a single framework for addressing biological threats and securing biopharmaceutical supply chains.
A second brain, built from scratch. store is a Go service that gives a language model durable memory: a Postgres database extended with Apache AGE for graph edges and pgvector for embeddings, fronted by a BGE-M3 model so notes, entities, and conversations are written once and retrieved by meaning. It is the canonical-fact backbone that several of these other projects read from and write to — the quiet infrastructure that lets an agent remember across sessions.
Established the national platform linking KCDC, HIRA, and the Cancer Registry.
Drafted Korea's first health-information de-identification guideline.
Framework for universal access to AI-driven health, from R&D to industry.
Drafted a $1B R&D plan and led legislation for advanced regenerative therapies.
Ran a welfare fraud-prevention program that saved an estimated $2.5B.
An adaptive activity intervention optimized within a randomized controlled trial.
ML modeling of home energy-load growth and homeowner energy literacy.
Human exon database for alternative-splicing analysis — early bioinformatics work.
Can a swarm of sub-1B local models reason like a frontier model?
An LLM designs and weight-tunes a network by reasoning — no backpropagation.
Passage-level personal knowledge wiki with an auto-generated ontology.
Reference-crawling RAG that walks a seed paper's citation graph N levels deep.
Self-evolving multi-agent framework with pgvector long-term memory.
A deployed full-stack mindmap shared between agents and humans.
Autonomous data-science agent with a network-jailed, reproducible runner.
Conversational paper-research assistant running on a local Gemma model.
An MCP server that generates publication-quality TikZ figures.
An LLM handed real microcontrollers, deciding what to build (with rbp-agent on RP2040).
AI Capacity Building for the Health Workforce in Korea: From Medical AI to AI Medicine
Forum on Harnessing AI for Health Equity, Asian Development Bank (ADB), Manila
Structural Characteristics of the Korean Health System and Digital Health Innovation
Stanford Graduate School of Business, invited lecture, Seoul
AI Basic Health Framework
WHO Department of Digital Health and Innovation, online