Asclepius-HI
VillageHacks 2025 Supermemory Track winner. We built a patient/clinician healthcare demo with memory-aware chat, citation-backed retrieval, and a basic escalation path for risky messages.
“Most medical chat demos forget the previous visit. I wanted our hackathon project to remember context, cite what it used, and avoid answering when a message looked risky.”
In a real clinic, something like this would only be useful if it routed low-risk questions safely and escalated anything concerning. Our version was a hackathon proof-of-concept for that workflow, not medical software.
I was worried about two common failure modes in healthcare AI demos: answers without sources, and answers that should have been escalated instead. For VillageHacks 2025, we tried to handle both inside the hackathon window: retrieve cited context before answering, and send higher-risk messages toward clinician review.
I built the RAG backend with ChromaDB and a simple ingestion/retrieval flow, then connected it to separate patient and clinician screens. The patient side handled chat and symptom context; the clinician side showed summaries, citations, and escalated messages. We kept the database simple for the demo, but structured it so it would not be painful to move beyond SQLite later.
- 01Patient + clinician screens — separate views so the demo could show both sides of the workflow.
- 02FastAPI service — chat, summary, and escalation endpoints; chat calls retrieval before generation.
- 03RAG layer — prior notes and demo documents are embedded into ChromaDB, then returned chunks are passed into the answer with citation metadata.
- 04Escalation check — simple high-risk message detection before letting the model answer. This needs much better evals before any real use.
- 05SQLite demo storage — chosen for hackathon speed, with the schema kept simple enough to migrate later.
Citations that actually trace
The first naive version could have just asked the model to cite sources, but that is not trustworthy. I made the citations come from retrieved ChromaDB chunks instead, so the UI only shows citations tied to context we actually passed into that answer.
Conservative without being useless
For the demo, we needed the assistant to answer normal questions but stop on risky ones. I used a simple escalation check, but this is one of the parts I would not trust without clinician-reviewed examples and much more testing.
Hackathon scope pressure
We had to cut anything that looked like real compliance, auth, or production medical workflow. The project is strongest as a demo of the interaction pattern: memory, citations, and escalation. It is not a deployable health product.
- →Won the Supermemory Track at VillageHacks 2025.
- →Demoed memory-aware Q&A, citation-grounded answers, and risk-escalation handoff in a working end-to-end flow.
- →The schema was written so SQLite could be swapped later, but the project has not been tested in a real clinical setting.
- ·This is a hackathon prototype, not medical software. It has no clinical validation, no real patient data, and no production privacy/security layer.
- ·If I rebuilt it, I would start with clinician-reviewed escalation examples and retrieval evals before adding more features.
- ·The demo would also need real auth, audit logging, and safer data handling before it could leave a hackathon context.