ClinicalBrief AI

AI-Assisted Clinical Decision Support Pipeline

Patient lab data → GPT-4o + RAG + live PubMed → confidence-gated clinical brief → clinician inbox in under 60 seconds

GPT-4o · RAG

SerpApi · NIH PubMed

Foxit eSign

Nutrient DWS

AI Builders Hackathon 2026 · SaaS Product Track · U. Bhowal · LearnHive Labs · learnhive.org

PROBLEM STATEMENT

Clinical Decision Support is broken for most clinics

$50K+

annual cost of enterprise

CDS tools per clinic

2+ hrs

per complex case spent

searching literature

40%

AI hallucination risk in

clinical scenarios

0

open inspectable CDS

pipelines available today

The Core Problems:

💸

Enterprise CDS tools (UpToDate, Epic) cost $500+/clinician/year — inaccessible to small and rural clinics

🧠

AI tools hallucinate clinical recommendations — a hallucinated drug dosage or contraindication can harm patients

⏱️

Clinicians spend 2+ hours per complex case searching literature and writing structured briefs manually

🔒

No open, inspectable, evidence-grounded CDS pipeline that any clinic can run without enterprise software

SOLUTION OVERVIEW

ClinicalBrief AI — Five-Step Agentic Pipeline

1

📋 Patient Intake

Google Sheets trigger — lab values, medications, diagnoses. No app, no login.

2

🔍 Live Evidence

SerpApi queries NIH PubMed in real time — 3 peer-reviewed papers injected before AI runs.

3

🧠 Clinical Agent

GPT-4o at temp 0.1 + RAG over 6 clinical guidelines → structured JSON brief with citations.

4

⚖️ Confidence Gate

Deterministic IF node — ≥80%: auto-sign via Foxit. <80% or extreme values: human review.

5

📧 Signed Delivery

Clinician receives HTML email + Foxit eSign request. Agent cannot sign — human boundary by design.

TARGET USERS

Who ClinicalBrief AI is built for

🏥

Small & Rural Clinics

Cannot afford $50K/year enterprise CDS. ClinicalBrief AI runs on commodity APIs and a Google Sheet. Zero infrastructure cost.

👩‍⚕️

Independent Physicians

Spend 2+ hours per complex case searching literature. ClinicalBrief AI delivers a source-cited brief in 60 seconds.

🏫

Medical Education

Residents and students learning evidence-based medicine. Every recommendation links to the exact PubMed paper it came from.

🔬

Healthcare Researchers

Need to run CDS pipelines at scale on real patient data. Fully inspectable n8n pipeline — every node auditable.

🌍

Global Health Orgs

Operating in low-resource settings where enterprise software is not an option. Open source, inspectable, replicable.

🛡️

Health Tech Builders

Developers building CDS products who need a reference architecture for RAG-grounded, confidence-gated AI in healthcare.

PRODUCT FEATURES

What makes ClinicalBrief AI different

Live PubMed Evidence

SerpApi retrieves current NIH publications before every brief — 2024/2025 papers, not frozen training data

RAG-Grounded Output

6 clinical guidelines embedded (ADA 2024, ACC/AHA 2023, JNC 8, WHO, NKF KDIGO, USPSTF 2024) at 1,536 dims

Deterministic Safety Gate

IF node — not ML — routes ≥80% confidence to auto-sign, <80% or extreme clinical values to human review

Fully Inspectable

14 nodes in n8n — every input/output visible. Trace any recommendation to the exact PubMed paper.

Human Signs — Agent Cannot

Foxit eSign boundary: agent generates, clinician must authorize. Responsible AI by architecture, not by claim.

Zero Enterprise Dependency

Google Sheets intake, Gmail delivery, public domain guidelines. Runs without any enterprise software license.

TECHNICAL ARCHITECTURE

How it works — 14 nodes, 3 chained APIs, 1 confidence gate

Google Sheets

Trigger

Nutrient DWS

Extract

SerpApi

PubMed

Build Agent

Prompt

GPT-4o

Agent + RAG

Confidence

Gate

≥ 80% confidence

< 80% or clinical flags

✅ Auto-Sign Path

Foxit eSign → signed PDF → Gmail delivery → Audit Log

⚠️ Human Review Path

Clinical flags displayed → human review queue → Audit Log

Confidence formula:

confidence_score = (agent_score + extraction_confidence) / 2

n8n

GPT-4o

text-embedding-3-small

SerpApi

Nutrient DWS

Foxit eSign

Google Sheets

Gmail API

AI TECHNOLOGIES USED

The AI stack — and why each component was chosen

GPT-4o

Clinical Reasoning Agent

Temperature 0.1 — near-deterministic output. Same patient data always produces same recommendation. Structured JSON schema enforced by output parser.

text-embedding-3-small

RAG Embeddings

1,536 dimensions, cosine similarity retrieval. 6 clinical guidelines (ADA, ACC/AHA, JNC 8, WHO, NKF KDIGO, USPSTF) embedded and retrieved per query.

SerpApi

Live Evidence Retrieval

Google Scholar API → NIH PubMed. 3 peer-reviewed publications retrieved per patient and injected into agent prompt before any generation. Live 2024-2025 evidence.

Nutrient DWS

Lab Data Extraction

Document Web Services for structured extraction from patient lab PDFs. Field-level confidence scoring. Extreme value detection (HbA1c >10%, LDL >190, BMI >40).

Foxit eSign

Document Signing

REST API for tamper-evident electronic signing. Agent generates the brief — clinician must sign. This boundary is the responsible AI design principle.

n8n

Workflow Orchestration

14-node visual pipeline. Every node inspectable. Confidence gate is a deterministic IF node — not ML. Full audit trail per execution.

IMPACT & VALUE PROPOSITION

Why ClinicalBrief AI matters

$50K

enterprise CDS cost

replaced per clinic

UpToDate · Epic · DynaMed

60s

sheet row to

clinician inbox

end-to-end runtime

80%

confidence threshold

for auto-sign

deterministic gate

6

open-access clinical

guidelines in RAG

ADA · AHA · JNC 8 · WHO

Open & Inspectable

14 nodes in n8n. Every input and output is visible. A clinician or regulator can trace any recommendation back to its exact source.

Responsible by Architecture

The confidence gate is an IF node — not a model. Extreme clinical values force human review. The agent cannot override this boundary.

Grounded in Evidence

SerpApi retrieves live 2024-2025 PubMed papers before every brief. RAG retrieves named clinical guidelines. Zero reliance on training data memory.

FUTURE ROADMAP

From open pipeline to production CDS platform

Phase 1

Q4 2026

Production Infrastructure

▸ Supabase pgvector — persistent RAG store, no reseed required

▸ Real Nutrient DWS PDF upload — replace simulated Google Sheet extraction

▸ Foxit eSign production endpoint — full tamper-evident signing

▸ n8n queue mode — parallel multi-patient processing

Phase 2

Q1 2027

EHR & IoT Integration

▸ HL7 FHIR webhook — replace Google Sheets with standards-compliant EHR feed

▸ IoT wearable integration — Apple Watch, Fitbit, CGM direct webhook

▸ Multi-specialty knowledge base — oncology, cardiology, nephrology tracks

▸ HIPAA-compliant data handling — BAA agreements, encryption at rest

Phase 3

Q2-Q3 2027

SaaS Platform

▸ Clinician dashboard — Looker Studio confidence trends, review rates, turnaround

▸ Multi-tenant architecture — clinic onboarding, credential management

▸ Mobile app — clinician brief review and signing on iOS and Android

▸ API product — embeddable CDS engine for health tech builders

The pipeline is live today. The path to production is clear. The architecture is inspectable and replicable by any clinic or health tech team.

ClinicalBrief AI

Open Clinical Decision Support — Built in n8n · Inspectable · Replicable

Google Sheet → SerpApi PubMed → GPT-4o + RAG → Confidence Gate → Foxit eSign → Clinician

GPT-4o

SerpApi

Nutrient DWS

Foxit eSign

n8n · 14 nodes

RAG · 1536 dims

Try the live pipeline by adding a new line on the shared document link and watch your email...

Add a row to the Google Sheet → get a clinical brief in your inbox within 60 seconds

n8n: udayshankar.app.n8n.cloud/workflow/sBl3HSAqigFBB2t1

Voluntarily contributed to the open innovation community · Published under LearnHive Labs · learnhive.org · U. Bhowal