Nabd — Post-Visit Clinical Scribe

A post-visit clinical assistant. A doctor speaks a short summary after a patient visit; Nabd turns that into a structured SOAP note grounded in the patient's real history, lets the doctor ask follow-up questions on demand, and surfaces cross-specialty connections via Nabd Mind — all with a human approval step before anything is finalized.

Team

Setup

pip install -r requirements.txt

Validate the locked schemas

Each schema file has a runnable self-check:

python schemas/stt_schema.py
python schemas/rag_schema.py
python schemas/soap_schema.py

Each prints a valid example payload if the schema is well-formed. To check the actual mock data in /mock_data against the schemas, load it directly:

import json
from schemas import STTOutput, RAGOutput, SOAPNote

STTOutput.model_validate(json.load(open("mock_data/sample_stt_output.json")))
RAGOutput.model_validate(json.load(open("mock_data/sample_rag_output.json")))
SOAPNote.model_validate(json.load(open("mock_data/sample_soap_output.json")))

If any of these raise a validation error, something in /mock_data no longer matches /schemas — fix before building anything downstream.

Repo Structure

/schemas          ← locked Pydantic models (source of truth for interfaces)
/mock_data         ← valid sample payloads — build against these in parallel
/stt               /rag              /llm_soap
/ask_nabd          /nabd_mind        /orchestration
CLAUDE.md          ← shared architecture context for Claude Code

Each component folder has a README.md (the assignment) and a RESULTS.md (the actual research: what was tested, real numbers, the decision and why) — see CLAUDE.md's "General Workflow" section.

Scope This Phase

• STT: English only (language_detected field stays for future Arabic/mixed support)
• OCR: frozen, out of scope
• 100% open-source stack — no proprietary APIs in the core pipeline