MedDroid SDK (Python) — meddroid-sdk

Fine-tune medical models, build RAGs, run local inference, and publish to the MedDroid Marketplace — from one CLI.

Status: early scaffold (v0.1.0). The training / RAG / serving helpers wrap standard open-source tooling and work today; submit posts to the live Marketplace intake (the medicalandroid-chat Worker → Supabase) and falls back to writing submission.json only if that's unreachable. Not yet published to PyPI — install from source for now.

Your files, data and compute stay on your own machine/GPU; we only store the submission's link + metadata.

PyPI distribution name is meddroid-sdk; the import module and CLI are meddroid.

Install

# once published:  pip install meddroid-sdk

# from source (this folder) for now
pip install -e .

# with training / serving / rag extras
pip install -e ".[train,serve,rag]"

Use

meddroid login
meddroid train --base phi3-mini --data data.jsonl --out ./my-model
meddroid rag build ./my-files --out ./my-rag
meddroid rag query ./my-rag "What is the sepsis bundle timeline?"
# many questions? load once, answer instantly (REPL, or --port for a local HTTP API):
meddroid rag serve ./my-rag
meddroid serve ./my-model --port 8000
meddroid submit --name "Sepsis EWS" --brand "My Clinic" \
  --link https://huggingface.co/your-org/sepsis-ews --public --price 49

Base models: medgemma-4b, openbiollm-8b, meditron-7b, phi3-mini (or any Hugging Face repo id). Each keeps its own licence — review before commercial use.

We store the link, model card and brand — never your weights. Host the files on Hugging Face / Kaggle / Drive. Public listings can be monetised; private ones are invite-only.

Guided, step-by-step version: https://medicalandroid.com/#/model-studio