Longitudinal EMR AI

Long, sprawling EMRs — in 5 seconds ChartOneShot reads them

Time a physician needs to review a complex case: 25 min. ChartOneShot cuts that to 5 sec, protecting the golden hour.

Source: Nolan et al., Mayo Clinic multi-site survey study, Applied Clinical Informatics (2017)

Try It Now See How It Works

Patient-level longitudinal summaries · source citations

Longitudinal Chart

9 visits · 7 months

Patient: 52yo male (liver transplant outpatient)

[2026.01.23] Liver transplant follow-up — Tacrolimus 6.2, LFT normal

[2026.01.30] Same as 1.23. Stable.

[2026.02.10] Right pubic fracture from TA — ER

[2026.02.20] AST 58 ↑ ALT 64 ↑ ...

[2026.03.02] AST 26 ✓ ALT 29 ✓ back to normal

[2026.03.30] Stable, maintained...

[2026.04.10] Bee-sting anaphylaxis...

Document-level chunking · pgvector · EmbeddingGemma 300M

Multi-specialty consult5.2s

Primary condition

Living-donor liver transplant follow-up (months 12–17)

AST/ALT transiently elevated after trauma → back to normal within 2 weeks; Tacrolimus trough stays stable

Timeline (top: primary / bottom: secondary)

1.23Checkup

2.10Pelvic Fx

2.20AST↑

3.02Normal

4.10Anaphyl.

4.25Stable

Abbrev. / jargonf/u → follow-upBID → twice dailyHLD · jargon?

Source quotes: 9 attached — each item traceable to its source chunk ID.

PT-2026-LT000152yo maleLiving-donor liver transplantOutpatient 7 · ER 2● Reading chart…

[2026.01.23 Outpatient — Liver transplant routine follow-up (month 12)]

S: Stable, no specific complaints. No dyspnea on exertion.source

O Lab: AST 28, ALT 31, T-Bil 0.9, ALP 78, GGT 32

Tacrolimus trough 6.2 ng/mL (target 5-8)

PMHx: HBV-related LC s/p living-donor liver transplant, HTN, DM, HLD⚠ source-preserved

A: Liver graft stable. No drug-related adverse effects.

──────────────────────────────────────

[2026.02.20 Outpatient — Follow-up (10 days post-TA)]

Lab: AST 58 ↑, ALT 64 ↑ (above upper-normal — elevated)source

A: Likely post-trauma muscle injury. Rejection/infection unlikely.

──────────────────────────────────────

[2026.03.02 Outpatient — Follow-up (recheck)]

Lab: AST 26 ✓, ALT 29 ✓ (returned to normal range)source

A: Confirmed transient elevation. Not rejection.

↳ summary folds out

Post living-donor liver transplant follow-up (months 12–17)

AST/ALT: stable (1.23–1.30) → transient post-trauma rise (2.20, AST 58·ALT 64 ↑) → returned to normal within 2 weeks (3.02, AST 26·ALT 29) → remained stable (3.30–4.25)

9 citations · traceable

Scroll and the chart reads itself. 25 min → 5 sec

Longitudinal

Not one chart — across all 9 visits

9 events · longitudinal

2024.08

Living-donor liver transplant performed (HBV-LC, right hepatic lobe donation)baseline

source ✓

2026.01.23

Liver transplant routine follow-up — LFT normal, Tacrolimus 6.2primary

source ✓

2026.01.30

Liver transplant routine follow-up — remains stableprimary

source ✓

2026.02.10

Right pubic ramus fracture from TA — conservative management in ERincidental

source ✓

2026.02.20

LFT
trend

Transient rise to AST 58·ALT 64 ↑ (10 days after TA, presumed trauma)primary

source ✓

2026.03.02

Returned to normal AST 26·ALT 29 ✓ — rejection excludedprimary

source ✓

2026.03.30

Liver function stable + fracture clinically healedprimary

source ✓

2026.04.10

Grade 2 anaphylaxis after bee sting — EpiPen × 1incidental

source ✓

2026.04.25

Liver function stable (AST 27·ALT 30) + allergy recoveredprimary

source ✓

Pipeline

How It Works

LangGraph inference pipeline

Webhook In

Document Chunker

Document-level chunking

Abbreviation Lookup

Specialty dictionaryDeterministic

RAG Retriever

pgvectorEmbeddingGemma 300M

MedGemma 27B

vLLMQLoRA r=164-bit NF4

Citation Verifier

Source citation enforced

Webhook Response

JSON structured

Use baseline MedGemma 27B as-is

Korean clinical specialization via per-specialty QLoRA PEFT adapters

+ DPO applied next (3-stage roadmap)

LangGraph Inference Pipeline

Ready

RuntimevLLM

Base ModelMedGemma 27B

AdapterQLoRA (PEFT) r=16

Quantize4-bit NF4

Vector DBpgvector

EmbeddingEmbeddingGemma 300M

GPURTX A6000 48GB

NetworkClosed-network on-prem

Swap nodes to flexibly change the model, retrieval, or chunking strategy

Adaptation roadmap · RAG + Few-shot (1) → + QLoRA SFT (2) → + DPO (3, optional)

Step 01

Upload EMR

Upload the raw text EMR as-is. No preprocessing needed.

Step 02

AI Analysis

An LLM reads the text in clinical context, structuring vitals, chief complaint, allergies, and the patient timeline around the current condition.

Step 03

Results Dashboard

Structured JSON in under 30 seconds. Split View compares original and summary, abnormal values highlighted.

Webhook In

Document Chunker

Document-level chunking

Abbreviation Lookup

Specialty dictionaryDeterministic

RAG Retriever

pgvectorEmbeddingGemma 300M

MedGemma 27B

vLLMQLoRA r=164-bit NF4

Citation Verifier

Source citation enforced

Webhook Response

JSON structured

Step 01

Upload EMR

Upload the raw text EMR as-is. No preprocessing needed.

Step 02

AI Analysis

An LLM reads the text in clinical context, structuring vitals, chief complaint, allergies, and the patient timeline around the current condition.

Step 03

Results Dashboard

Structured JSON in under 30 seconds. Split View compares original and summary, abnormal values highlighted.

Use baseline MedGemma 27B as-is

Korean clinical specialization via per-specialty QLoRA PEFT adapters

+ DPO applied next (3-stage roadmap)

LangGraph Inference Pipeline

Ready

RuntimevLLM

Base ModelMedGemma 27B

AdapterQLoRA (PEFT) r=16

Quantize4-bit NF4

Vector DBpgvector

EmbeddingEmbeddingGemma 300M

GPURTX A6000 48GB

NetworkClosed-network on-prem

Swap nodes to flexibly change the model, retrieval, or chunking strategy

Adaptation roadmap · RAG + Few-shot (1) → + QLoRA SFT (2) → + DPO (3, optional)

Capabilities

Features

Accurate compression of long EMRs

Split View Comparison

Compare the original EMR and AI summary on one screen. Long, repetitive charts compress into key trends at a glance.

Original

Summary

Patient-level Longitudinal Summary

Extracts only the core conditions from multi-visit outpatient and inpatient records. Primary trends above the arrow, secondary issues below.

Deterministic Abbreviation & Jargon Resolution

Resolves terms via a per-specialty abbreviation/jargon dictionary. Items not in it stay verbatim and get flagged "presumed jargon" — no LLM guessing, no hallucination.

f/u→follow-up

BID→twice daily

HLD→Not registered · presumed jargon

Data Sovereignty

Per-specialty adapters trained on MedGemma 27B plus Yangsan Pusan National University Hospital's own data. Patient data and weights stay locked inside the closed network.

Closed network · LIVE

External transmissions: 0

Split View Comparison

Compare the original EMR and AI summary on one screen. Long, repetitive charts compress into key trends at a glance.

Original

Summary

Patient-level Longitudinal Summary

Extracts only the core conditions from multi-visit outpatient and inpatient records. Primary trends above the arrow, secondary issues below.

Deterministic Abbreviation & Jargon Resolution

Resolves terms via a per-specialty abbreviation/jargon dictionary. Items not in it stay verbatim and get flagged "presumed jargon" — no LLM guessing, no hallucination.

f/u→follow-up

BID→twice daily

HLD→Not registered · presumed jargon

Data Sovereignty

Per-specialty adapters trained on MedGemma 27B plus Yangsan Pusan National University Hospital's own data. Patient data and weights stay locked inside the closed network.

Closed network · LIVE

External transmissions: 0

Split View Comparison

Compare the original EMR and AI summary on one screen. Long, repetitive charts compress into key trends at a glance.

Original

Summary

Patient-level Longitudinal Summary

Extracts only the core conditions from multi-visit outpatient and inpatient records. Primary trends above the arrow, secondary issues below.

Deterministic Abbreviation & Jargon Resolution

Resolves terms via a per-specialty abbreviation/jargon dictionary. Items not in it stay verbatim and get flagged "presumed jargon" — no LLM guessing, no hallucination.

f/u→follow-up

BID→twice daily

HLD→Not registered · presumed jargon

Data Sovereignty

Per-specialty adapters trained on MedGemma 27B plus Yangsan Pusan National University Hospital's own data. Patient data and weights stay locked inside the closed network.

Closed network · LIVE

External transmissions: 0

Trust

Cited evidence

ClaimSource · chunk

AST/ALT transient rise, back to normal in 2 weeks — rejection ruled out

“AST 26 ✓, ALT 29 ✓ (back to normal range) … not rejection.”

2026.03.02 · chunk-pt01-04✓ verified

Tacrolimus trough consistently within target (5-8)

“Tacrolimus trough 6.2 ng/mL (target 5-8)”

2026.01.23 · chunk-pt01-00✓ verified

Triple immunosuppression maintained, no dose change

“Immunosuppression protocol: Tacrolimus + Mycophenolate + Prednisolone triple therapy”

2024.08 · chunk-pt01-bl✓ verified

“HLD” — abbreviation not in dictionary

Kept verbatim. Not expanded by guess.

2026.01.23 source⚠ kept verbatim

Unverified 0· Citations 9· Preserved 1

< 0s

EMR Analysis

Physician 25 min → 5 sec

0 → 1

Visit Consolidation

9 outpatient/inpatient visits into one timeline card

Closed-network Operation

0 external transmissions · on-prem

0B+α

MedGemma 27B Korean clinical specialization

QLoRA PEFT + DPO roadmap

Synthea™

Interactive Demo

Synthetic patient data generated by MITRE Synthea™. Not real patient information.

Challenge

If you were the doctor, could you beat the AI?

Read this long outpatient chart and identify the patient's primary concern. ChartOneShot finishes the same task in 5 seconds, so see how many seconds it takes you.