CLARA: AI's Life-Saving Intervention

Clinical Logic Assessment & Reasoning Assistant

CLARA is a multimodal AI safety net dedicated to supporting clinical reasoning and helping identify diagnostic shadowing to optimize patient care.

The Problem: Cognitive Load & System Strain

• Diagnostic error impacts 12 million Americans annually [1].
• In women's health, the crisis is more severe: women are diagnosed an average 4 years later than men [2].
• A major factor is Diagnostic Shadowing [3]: the tendency to inadvertently attribute physical symptoms to psychiatric history.
• This is a recognized systemic issue-it's the result of Dual Process Theory being strained under immense clinical pressure.
  • Physicians operating under high cognitive load are forced to rely on System 1 (fast/heuristic thinking) just to keep up.
  • The analytical System 2 can be inadvertently bypassed, allowing necessary cognitive shortcuts to occasionally result in unintended clinical blind spots [4].

The Solution: A Cognitive Support Strategy

CLARA is built as an "External System 2"-a background AI ally powered by Gemini 3 Pro.

• Not just a scribe. Rather than merely transcribing conversations, CLARA supports the reasoning process behind the encounter.
• Functions as a safety net, continuously running in the background to reduce physician cognitive burden.
• Highlights potential diagnostic blind spots and offers gentle clinical nudges for review prior to patient discharge.
• Helps physicians maintain their highest standard of clinical accuracy, even when their mental bandwidth is exhausted.

Methodology & Impact

• Developed with Google AI Studio.
• Processes consented consultation audio and outputs standardized "Clinical Insights" (JSON).
• Converts subjective reports of being dismissed into objective, actionable data for the care team.
• Empowers equitable patient care while reducing cognitive fatigue and systemic risk for providers.

Getting Started

Prerequisites

• Node.js (latest LTS recommended)

Local Setup

1. Clone the repository and navigate to the project root.
2. Install dependencies:

   npm install

3. Create a .env file from the example and set your Gemini API key:

   cp .env.example .env

   GEMINI_API_KEY=your_key_here

4. Run the development server (Vite on :3000 + Express on :3001):

   npm run dev

5. Open http://localhost:3000 in your browser.

---

Deploying to Google Cloud Run

The app is packaged as a single Docker container: the Express server serves the built Vite frontend and proxies all Gemini calls server-side, so the API key never reaches the browser.

Prerequisites

• Docker installed locally
• Google Cloud CLI installed and authenticated
• A Google Cloud project with billing enabled

One-command deploy (source deploy — no local Docker needed)

gcloud run deploy clara \
  --source . \
  --region us-central1 \
  --platform managed \
  --allow-unauthenticated \
  --set-env-vars GEMINI_API_KEY=your_key_here

Cloud Run builds the container using the Dockerfile, then deploys it. The GEMINI_API_KEY is stored as a Cloud Run environment variable and is only accessible inside the container.

Optional: use Secret Manager instead of an env var

For production, store the key in Secret Manager so it never appears in deploy logs or the Cloud Console UI:

# 1. Create the secret (one-time)
echo -n "your_key_here" | gcloud secrets create GEMINI_API_KEY --data-file=-

# 2. Deploy referencing the secret
gcloud run deploy clara \
  --source . \
  --region us-central1 \
  --platform managed \
  --allow-unauthenticated \
  --set-secrets GEMINI_API_KEY=GEMINI_API_KEY:latest

Manual build & push (optional)

# Build
docker build -t gcr.io/YOUR_PROJECT_ID/clara .

# Push
docker push gcr.io/YOUR_PROJECT_ID/clara

# Deploy
gcloud run deploy clara \
  --image gcr.io/YOUR_PROJECT_ID/clara \
  --region us-central1 \
  --platform managed \
  --allow-unauthenticated \
  --set-env-vars GEMINI_API_KEY=your_key_here

---

References

1. Singh H et al. BMJ Qual Saf 2014.
2. Westergaard D et al. Nat Commun 2019.
3. Reiss S et al. Am J Ment Defic 1982.
4. Trimble M UMJ 2015.

---

For more information or contributions, see the codebase and issues on GitHub.