---
title: README
emoji: 🪶
colorFrom: blue
colorTo: indigo
sdk: static
pinned: false
---

<p align="center">
  <img src="./assets/hero-meddies-org.png" alt="Meddies — verifiable clinical intelligence for real-world care" width="100%" />
</p>

# Meddies

> Meddies delivers verifiable clinical intelligence for real-world care.

## Who are we?

Meddies is a clinical-AI team building an evidence-led layer on top of hospital EMRs. We work alongside clinicians on the workflows that consume the most of their time — note writing, safety screening, decision support, patient search — and ship the datasets and models needed to make those workflows verifiable rather than plausible.

We are based in Vietnam, aligned with Ministry of Health guidelines, and deploy with on-premises de-identification so identifiable patient data never leaves the hospital network.

## What we ship

Open releases live on this org. Each release ships with a research-grade card stating its scope, license, limits, and intended use.

**Datasets**

- [`meddies-pii`](https://huggingface.co/datasets/Meddies/meddies-pii) — synthetic PII across 17 languages and 7 entity families for multilingual clinical and administrative documents.
- [`meddies-persona-vie`](https://huggingface.co/datasets/Meddies/meddies-persona-vie) — Vietnamese patient personas for synthetic healthcare generation, triage simulation, and downstream clinical workflows.
- [`meddies-consultant`](https://huggingface.co/datasets/Meddies/meddies-consultant) — synthetic Vietnamese consultations covering doctor–patient exchanges across clinical scenarios.
- [`meddies-patient-safety`](https://huggingface.co/datasets/Meddies/meddies-patient-safety) — Vietnamese clinical red-team prompts, judged on safety, quality, and skill.

**Models**

- [`meddies-pii`](https://huggingface.co/Meddies/meddies-pii) — token-classification model for clinical PII extraction, trained on the matching dataset.

Internal artifacts — production models, hospital-tuned weights, and patient-derived data — stay on-premises and are not published here.

## How we work

- **Privacy through de-identification.** Patient PII is stripped on-premises by Meddies utility models before any clinical reasoning runs. The main reasoning model only sees de-identified inputs, so identifiable patient data never leaves the hospital network.
- **Verifiable, not plausible.** Outputs are graded against MOH guidelines and clinician review, not against vibes. We publish the evaluation harness alongside the model.
- **Synthetic-first for open release.** Public datasets are synthetic by construction so the open work cannot leak patient information. Real-world evaluation happens inside partner hospitals under access control.
- **One source of truth per fact.** Every claim on a card maps to a checkable artifact — schema, eval row, generator script, or clinician sign-off.

## Vision

A care system where every clinical decision is traceable to evidence, every minute saved on typing returns to the patient, and every AI output a clinician relies on can be verified — not trusted on faith. Open datasets and open models for the verifiable parts; on-premises de-identification for the parts that touch identifiable patient data.

## Get in touch

- Website — [meddies-ai.com](https://meddies-ai.com)
- Demo — [app.meddies-ai.com](https://app.meddies-ai.com)
- Contact — [contact@meddies-ai.com](mailto:contact@meddies-ai.com)

If you want to use a Meddies dataset or model in commercial work, please reach out at [contact@meddies-ai.com](mailto:contact@meddies-ai.com). If you test a release on real workflows, share failures, not just wins — that is how the next version gets safer.