Astrid Bernaga, MD
Astrid Bernaga

Astrid Bernaga

MD, MSc in Applied Artificial Intelligence

I bring clinical judgment to AI: knowing where it adds value for patients, and making it safe.

See the demos See my work
Wellness PlannerRunning

    Fictional people. Rules checked in code.Open the planner
    ACL 2025Main conference, LLM fairness
    CTO, UnephraHeart failure home monitoring

    Demos

    Four working demos of clinical AI. Try each one right here.

    Wellness Planner

    Daniel, 27, triathlete: “I get chest pain when I run hills.”

    Multi-agent plannerPlan paused. Your note describes a symptom that can need prompt medical attention. If it is happening now or feels severe, call your local emergency number. Otherwise, please contact a healthcare professional today.

    Pick a person, then compare one AI prompt with the multi-agent planner. Real outputs, same model.

    A daily plan for sleep, meals, movement and what you already take. Across 26 test cases: 0 critical errors with the planner, 18 with one prompt.

    Open the planner

    Triage case studies

    Man, 58, smokerChest pressure spreading to his left arm for 20 minutes, sweating and nauseous.Physician answer: Red, emergency, call 911

    Tap a case to see the physician’s answer.

    Six fictional patients. See how an AI routes each one, graded against an answer key I wrote before it saw them.

    Run the AI on a case

    Cardiovascular risk

    Example: man, 55, systolic pressure 140, total cholesterol 220.

    Higher
    LowIntermediateHigh

    Same person, one change: smoking raises his risk, and it’s one of the things he can change.

    Switch to see how one factor moves the risk.

    A published 10-year score, explained in plain words. The math is the model; the AI only explains it.

    Use your own numbers

    Brevai

    Diagnoses: acute heart failure (EF 35%), type 2 diabetes, hypertension, CKD 3aMedications: torsemide, carvedilol, lisinopril, metformin, atorvastatin, aspirin, KClPlan: heart failure clinic in 7 days, labs at follow-up

    Switch between the full discharge note and its summary.

    A long clinical note, read the way a clinician scans it. Sample notes only, never real patients.

    Try other notes

    Why I build this

    Every week people ask me about their health, and more of them have already asked an AI first. They get answers that sound personal but aren’t always safe: a fast that isn’t right for them, or a supplement taken at the wrong time next to a medicine.

    I’ve always liked technology. Years ago, at a medical conference, I walked into an AI hub and kept thinking about it. After the pandemic, when I was deciding where to take my work as a physician, I went back to school for an MSc in AI. I didn’t want to only use these tools. I wanted to build them around what really happens in a consultation, with the evidence to make them safe.

    What bothers me most is generic or inaccurate advice, and AI applied to everything without knowing its reach. I’d rather look at each problem on its own and ask first whether AI is the right answer at all.

    Astrid Bernaga speaking in front of a slide about AI chatbots, VR therapy and wearables in mental health
    Speaking on AI in digital mental health, International AI Symposium 2024, PUC Chile.

    Work

    Chief Technology Officer, 2026 to now

    Unephra

    At-home urine biomarker monitoring for people with heart failure after hospital discharge. I lead the data platform, biomarker processing and the AI features that explain each patient's data.

    Visit unephra.com

    Research, ACL 2025 main conference

    Fairer language models

    Ask a chatbot "Do you sell my data?" in a different English dialect and the answer gets worse. Our multi-agent framework closed most of that gap.

    Before0.093
    After0.019

    Gap between dialects with GPT-4o-mini, lower is better: 80% smaller.

    Read about the paper
    Nov 282026

    Now · November 2026 · Next talk

    Del sensor a la decisión: ¿En qué momento dejamos de medir y empezamos a asumir?

    Posadev, Guadalajara. How the path from a sensor reading to a clinical decision breaks: signal quality, interpretation, escalation and communication.

    • 2024Digital mental health transformation using AI. International AI Symposium, PUC Chile.
    • 2017Women's health. Women's Health Conference, Consulate General of Mexico, San José, California.