Medas — Mattia Smacchia

The problem

A medical assistant that answers from the model's memory is dangerous. Answers have to come from the patient's own record, cite their sources, convey the right urgency, and ignore documents that do not belong to the patient. All data in the project is synthetic.

The solution

A chat assistant that answers only from the patient's own record.

Retrieval with citations. The record and its PDFs are indexed in pgvector, and every answer cites the passages it used.
Document checks. Documents that do not belong to the patient are refused before indexing, and personal data is masked before anything is embedded.
Triage. Every answer carries a severity from S1 to S5, shown as a badge.
Secure sign-in. The Next.js chat uses Clerk, and the API trusts only the verified session, never the request.

Demo

Two questions on synthetic data. The second, about naproxen, is checked against the patient's allergy list and answered with an S1 emergency flag and the supporting record entries.

The problem

The solution

A chat assistant that answers only from the patient's own record.

Retrieval with citations. The record and its PDFs are indexed in pgvector, and every answer cites the passages it used.
Document checks. Documents that do not belong to the patient are refused before indexing, and personal data is masked before anything is embedded.
Triage. Every answer carries a severity from S1 to S5, shown as a badge.
Secure sign-in. The Next.js chat uses Clerk, and the API trusts only the verified session, never the request.

Demo

Two questions on synthetic data. The second, about naproxen, is checked against the patient's allergy list and answered with an S1 emergency flag and the supporting record entries.