- AI Research
- Healthcare
- Award Winner
“Get time back. Move care forward.”
M.A.R.S: AI Clinical Documentation
Hospital LLM Datathon
- Period
- September 2025- October 2025
- Role
- Team LeadSystem DesignPrompt EngineeringClinical Evaluation
- Team
- Team Lead (Me)2 LLM Engineers1 Clinical Advisor
- Org
- Seoul National University Hospital
What?
Built an LLM system that writes discharge summaries from real hospital records. Won 2nd place in Seoul National University Bundang Hospital's datathon.
Why?
Doctors spend up to 30% of their time turning messy notes into discharge summaries, and a national strike left them even more short-staffed.
How?
Cleaned up noisy, duplicated records, then designed department-specific, multi-step prompts with guardrails against hallucination, validated by clinicians.
I'm redesigning this case study as a web page. Every detail is in the PDF version below.
What is M.A.R.S, and what did I do here?
M.A.R.S
: Medical Auto-documentation with Real-world Structuring
A two-round, 2-month datathon hosted by Seoul National University Bundang Hospital.
Built an LLM system that turns raw clinical notes into discharge summaries, using MIMIC-IV dataset and real notes from 400 patients.
Team Lead
- Assembled the team and ran the project end to end
- Researched ~30 papers and clinical domain knowledge to define the approach
- Built the data cleansing, prompt design, and full system architecture
- Validated & bridged: ran clinician evaluations and translated between clinicians and engineers
- Presented at the final
Re-engineered Clinical Workflow with LLMs
🥇 1st Place @ Preliminary (1 of 100)
🥈 2nd Place @ Final (2 of 10)
Offered Research Role
Invited to co-publish and intern at
SNUBH Center for AI in Healthcare.
Built for Real Deployment
Our system was piloted in SNUBH’s
clinical documentation workflow.
Clinician-Validated
4 departments, 2 feedback rounds,
3.55 → 3.70 satisfaction.
Tap a page to open it at full resolution.
