01
Proven AI stalls
The top barrier to AI adoption is data quality (49%), not algorithms.
On-premise medical-data solution · MediPipe
Hospital documents, made AI‑ready
MediPipe organizes reports, readings and clinical notes into standard tables and asks clinicians only about suspect values. Patient data never leaves the hospital.
AI-ready data — clinician-verified data AI can use right away
01 · Input Any document In its own format
Reports
Readings
Sheets
02 · MEDIPIPE Organize · verify
- Map to clinical fields
- Check cross-field meaning
- Flag suspect values
Patient data never leaves the hospital
03 · Output
Verified data CSV · JSON export Confirmed
Date Standard
Item Mapped
Value Checked
Verify once, reuse everywhere
ResearchEMRAI
98.9% Less time per report (3 min → 2 s)
95.6% Faster data prep (3 months → 4 days)
3 sites Hospital deployments
4 LOIs From university hospitals
Why now
Hospitals have started using AI — and stalled on data
WHY NOW
Non-reimbursed market entry tripled hospital adoption of medical AI in about a year, and the government opened 62 hospital data-cleansing projects for 2026.
02
80% unstructured
Data prep takes 63–80% of a research project’s timeline.
03
By hand, every time
Values are copied into spreadsheets, and every project starts over.
MediPipe
Any hospital document, in a standard table
Clinicians don’t type — they only confirm.
- 01
Collect
Reports, readings, spreadsheets — as-is
- 02
Convert
Into the same clinical fields
- 03
Verify
Clinicians check only suspect values
- 04
Export
CSV · JSON (FHIR planned)
MediPipe review screen — each value is shown in place on the source, with low-confidence values flagged first.
Tap the image to view it full size.
The report and patient details shown are sample data for demonstration.
3-Layer Validation
Only values that pass all three checks are confirmed
In medicine, AI that invents plausible values is not acceptable.
01
Semantic Check
Catches values in the wrong field, like swapped systolic and diastolic.
02
Source Grounding
Every value links to its source; nothing is invented.
03
Human-in-the-Loop
Nothing is final until reviewed, with a full audit log.
Confirmations build a hospital-specific knowledge graph that becomes the next project’s rules.
Use cases
Three hospitals, three modules
Task-specific modules on MediPipe, so nothing is rebuilt per hospital.
Paid PoC underway
PFTPipe
Konyang Univ. Hospital · Pulmonary function
95.6%
Research-data prep 3 months → 4 days
- Joint research on a K-DATA-funded project
PoC complete
RxPipe
Ewha Mokdong Hospital · Patients’ own medications
94.1%
Automatic conversion coverage
- EMR integration targeted for end of 2026
PoC complete
MBPTPipe
NHIS Ilsan Hospital · Bronchial provocation
98.9%
3 min → 2 s per report
- Under review at JMIR Medical Informatics
How it scales
Finished modules install as-is at the next hospital — 30 person-days down to 2–3.
Paid contract K-DATA-funded Konyang project · phase 1 done
4 of 5 Specialists intending to buy after a demo
3 patents 1 filed · 2 joint filings in preparation
98.9% is based on 2,572 reports; 94.1% is conversion-path coverage, not recognition accuracy.
Awards
Encouragement Award
2026 SPRINT Next Unicorn's Day MediPipe pitch
Korea Entrepreneurship Foundation Chairman's Award · Nov 2026
Excellence Award
2026 SKKU Pre TeX-Corps business plan pitch
Sungkyunkwan University · Oct 2026
Top Excellence Award
2026 Lab Startup Competition IR pitch
Sungkyunkwan University · Sep 2026
Top Excellence Award
2026 Joint Lab-Startup Camp IR pitch
Sungkyunkwan University · Aug 2026
Research foundation
Multi-institution data-cleansing experience, turned into a product.
- Cleansed a 30-year, 990K-record pulmonary cohort
- Clinical validation with faculty at five hospitals
- 18 international journal papers · 8 first-author · h-index 9
- RCMLR workshop paper (NeurIPS 2026) · first author
- DeepBreath external AUROC 0.883 (research stage)
Selected publications
Deep learning models for acute kidney injury prediction: multi-center external validation and evaluation under simulated continuous monitoring conditions
Separate and joint associations of cardiorespiratory fitness and healthy vascular aging with subclinical atherosclerosis in men
Novel artificial intelligence-based technology to diagnose asthma using methacholine challenge tests
Median alert lead time tracks how long patients were observed, and runs opposite to specificity
Why MediPipe
The gap between generic OCR and the CDW
01
Verified to clinical meaning
Beyond characters — is each value in the right field?
02
Modules that compound
The next hospital installs with a tenth of the effort.
03
On-premise, inside the EMR
Runs on segregated networks; results show in the EMR.
Swipe sideways to view the full comparison.
| Comparison | MediPipe | Generic document AI | CDW · CDM | LLM structuring |
|---|---|---|---|---|
| Input | Hospital reports · clinical notes | General documents | Structured data | Text records |
| Clinical verification | ✓ Through clinician review | Characters only | — | Limited |
| Hallucination control | ✓ No value without a source | — | — | Needs separate checks |
| Deployment | On-premise · segregated | API · cloud | On-premise | Varies |
Growth strategy
Start with one department, grow to the hospital
Our first customers are department faculty running research projects. We expand from department to lab to the hospital’s AX team.
Market size (2026) A KRW 1.5T market, starting with pulmonology and allergy
TAMKorean medical data
KRW 1.48T 2026 estimate
SAMBuild · process · manage
KRW 762.6B KRW 3.76T projected by 2031
SOM394 hospitals
KRW 449.2B Departments at general hospitals and above
LAMInitial focus
KRW 23.6B Pulmonology and allergy · first 3 years
Roadmap From structuring to platform
- Q4 '26Stage 1 · Structure
- Report recognition · field extraction
- Target field accuracy 99.8%
- '27Stage 2 · Contextualize
- Clinical-note cleansing · time-series interpretation
- Dec '26 – '28Stage 3 · Verify · integrate
- Source-linked HITL · audit logs
- FHIR KR Core integration API
- '29 –Stage 4 · Platform
- Device integration · AI trained on in-hospital data
Long-term vision
Capture clinicians’ judgment as data, so every hospital delivers the same standard of care and research.
Team
Medical-AI regulatory experience and clinical-data expertise
Led by a CEO who drove MFDS SaMD approvals and a quality and regulatory lead from clinical practice.
Kyung Hyun Lee
이경현
Co-founder & CEO PhD student in Digital Health, SAIHST, SKKU · leads R&D
- 8 years in medical AI
- Led 2 MFDS SaMD approvals
- 18 international journal papers · 8 first-author
- NeurIPS 2026 workshop paper, first author
Dayoon Lee
이다윤
Co-founder & Chief Quality Officer Former research nurse · quality, regulatory and research data
- Research nurse 3 yrs · OR nurse 2 yrs
- Medical-device RA · QA
- MS student, Medical Device Industry, SKKU
Collaboration network
Advisory work has grown into deployments and purchase intent.
Contact
Make your hospital data AI-ready
Adoption, research collaboration and investor materials — we’d love to hear from you.
Email us directly lkh256@breathyou.care
- MediPipe adoption
- Research collaboration
- Investment / other
We review product adoption and clinical or research collaboration.