My first public repo - Why I built MedSynth

A few days ago I came across a tweet from a woman named Miriam. She’s 35, has rare metastatic breast cancer, less than 1% of cases, almost no documentation. She was asking the internet to help her find researchers who would take her case seriously.

Javi López (@javilop in Twitter) is a developer and responded by spending a week applying AI to her complete medical history. He unified hundreds of documents, ran two frontier models independently, then had each challenge the other’s findings in rounds until both stopped finding new things. He shared the methodology openly so anyone could replicate it.

I am finishing a master’s in AI and thinking about where to specialize. I replied to Miriam’s tweet and told her that her case was pushing me toward healthcare AI. I meant it. So I built MedSynth.

The core problem: a patient with a rare condition accumulates years of documents that no single doctor ever sees together. Patterns go unnoticed. Contradictions stay buried. Standard AI tools retrieve fragments on demand but build nothing permanent.

MedSynth builds a persistent structured wiki from a patient’s documents, then runs two frontier models against each other in an adversarial loop until they converge. The output is a structured clinical report: suggested tests, rare disease differentials, medication flags, clinical trial eligibility, and specific questions to bring to a doctor.

The clinical reasoning protocols are Javi’s work, adapted from what he built for Miriam. The codebase is open source: github.com/daniszwarc/medsynth

If you are interested in collaborating, I’d genuinely like to hear from you.

7 comments