GOAT Scribe: Emergency Medicine Edition

Production-ready Emergency Medicine scribe with built-in safety guardrails.

Designed for ED teams and NVIDIA healthcare AI startups that need a curated, bleeding-edge reference stack.

---

NVIDIA Healthcare AI Showcase · November 2025

This repository now doubles as a living showcase of NVIDIA’s latest open healthcare AI tooling. The H100 playbook demonstrated here spans clinical text, imaging, multimodal, and deployment workflows:

• CUDA 13.0 Update 2 with driver 580.105.08 (Blackwell-ready) and Nsight Compute/Systems 2025.4.
• CUTLASS 4.3.0 kernels + FlashAttention 3.0.1 for high-throughput Hopper compute.
• NVIDIA Nemotron Nano / Super Models served through NIM (text-only llama‑3.1‑nemotron‑nano‑8b‑v1 + VL llama‑3.1‑nemotron‑nano‑vl‑8b‑v1 examples).
• NVIDIA NeMo 25.09.1 + ModelOpt 0.21 for speech/Large Language Model fine-tuning.
• MONAI 1.5.1 (weekly) pipelines for medical imaging triage and FHIR export glue.
• Clara Holoscan 3.0.1 operators for edge inference + DALI (cuda130) preprocessing.
• TensorRT‑LLM 0.17.1 & Triton 3.2.1 for low-latency deployment.
• Presidio 2.2.360 + GOATnote guardrails for HIPAA compliance.

📁 See docs/NVIDIA-Healthcare-AI-Showcase.md for setup commands, architecture diagrams, and links to each toolkit.

---

Features Built for ED Teams

Safety guardrails for high-stakes clinical documentation:

• ✅ Vital signs validation: HR, BP, RR, Temp, SpO2, GCS within valid ranges
• ✅ Medication limits: 11 common ED medications with max safe doses
• ✅ Protocol warnings: ACLS, ATLS, sepsis, stroke, STEMI protocols
• ✅ ED documentation structure: 12-section format with Medical Decision Making
• ✅ High-risk awareness: Prompts for critical rule-outs
• ✅ Medical-legal compliance: Return precautions, informed consent

---

Features

Emergency Medicine Safety (CRITICAL)

Built-in Guardrails (cannot be disabled in production):

• ✅ Vital Signs Validation: HR, BP, RR, Temp, SpO2, GCS within valid ranges
• ✅ Medication Limits: 11 common ED medications with max safe doses
• ✅ Protocol Warnings: ACLS, ATLS, sepsis, stroke, STEMI protocols
• ✅ ED Note Structure: All 12 required sections enforced

Example: Catches "HR 350 BPM" → ⚠️ Critical violation (valid range: 20-250)
Example: Catches "Morphine 50mg IV" → ⚠️ Exceeds max safe dose (15mg)

Core Capabilities

• Guardrails validated: ./deploy/verify.sh and CLI tests flag unsafe vitals and medication doses (e.g. morphine 50 mg → critical violation).
• HIPAA PHI detection: Microsoft Presidio 2.2.360 removes identifiers; smoke tests surface redaction counts in CLI output.
• FHIR R4 export: FHIRExporter integration for GCP Healthcare API (see Python API example).
• Measured MONAI speedup: scripts/monai_h100_benchmark.py on H100 (CUDA 13.0.2) recorded 3.27 ms baseline vs 1.57 ms optimized (≈2.1×).
• Reproducible profiling: Nsight Compute artefacts live in docs/benchmarks/ncu/latest/ for transparent kernel analysis.

Technical Stack

• Model: nvidia/llama-3.1-nemotron-nano-8b-v1 served via NVIDIA NIM (public model as of Nov 2025)
• PHI Detection: Microsoft Presidio 2.2.360 (HIPAA 18-identifier coverage)
• FHIR: GCP Healthcare API (R4 dataset/store helpers)
• Infrastructure: H100 PCIe, CUDA 13.0.2, CUTLASS 4.3.0, FlashAttention 3.0.1

---

🚀 Deploy in 5 Minutes

For startups evaluating NVIDIA's healthcare AI stack:

# One-command deploy to H100
git clone https://github.com/GOATnote-Inc/scribegoat.git
cd scribegoat
export NGC_API_KEY="nvapi-YOUR-KEY-HERE"
./deploy/quick_start.sh

What you get:

• ✅ H100 deployment validated via ./deploy/h100_auto_deploy.sh (Nov 12 2025 smoke test)
• ✅ Safety guardrails (vitals, meds, protocols)
• ✅ HIPAA-compliant PHI detection
• ✅ FHIR R4 export ready
• ✅ Gradio UI on port 7860

Full deployment guide: deploy/DEPLOY.md

---

📊 Profile & Optimize

See NVIDIA tech in action:

# Run Nsight profiling (exports to docs/benchmarks/ncu/)
./deploy/profile.sh

Technologies showcased:

• CUDA 13.0 (Blackwell support, tile programming)
• CUTLASS 4.3 (high-performance primitives)
• FlashAttention-3 (H100-optimized attention)
• cuDNN 9.15, TensorRT 10.14

Profiling guide: deploy/PROFILE.md

Latest Nsight Dataset (Nov 12 2025)

• Report: docs/benchmarks/ncu/latest/ncu_monai_opt_latest.ncu-rep (MONAI UNet, torch.compile, AMP)
• Summary CSV + plots: docs/benchmarks/ncu/latest/
• Top kernels: vol2col (39% time) + vol2im (19% time) dominate; CUTLASS GEMM/GEMV kernels peak around 53% SM occupancy in this run.
• Full-metric replay inflates wall-clock time; for quick spot checks run sudo ncu --set roofline -c 1.
• Recreate sanitized artefacts with python scripts/process_ncu_report.py --raw /tmp/ncu_raw.csv --outdir docs/benchmarks/ncu/<tag>.

Usage

Web UI (Gradio)

python app.py
# Visit: http://localhost:7860
# Or get Brev public URL: brev urls

Command Line

# Generate ED note
python -m goatnote_scribe.cli "35M with chest pain, 2h duration..."

# From file
cat encounter.txt | python -m goatnote_scribe.cli

# With FHIR export
python -m goatnote_scribe.cli --export-fhir "Patient presents..."

Python API

from goatnote_scribe import GOATScribe

scribe = GOATScribe()

result = scribe(
    "45M with fever, cough, dyspnea. BP 128/82, HR 92, RR 18, SpO2 96%."
)

print(result['note'])               # Complete ED note (12 sections)
print(result['guardrail_safe'])     # Safety validation
print(result['fhir_bundle'])        # FHIR R4 bundle

scribe.wipe()  # HIPAA cleanup

FHIR Export to GCP

from goatnote_scribe import GOATScribe, FHIRExporter

scribe = GOATScribe()
exporter = FHIRExporter()

# Generate note
result = scribe("Patient encounter text...")

# Upload to GCP Healthcare API
response = exporter.upload_bundle(result['fhir_bundle'])
print(f"Uploaded: {response['id']}")

---

Configuration

Environment Variables

# Required
NGC_API_KEY=nvapi-YOUR-KEY-HERE

# Optional (defaults shown)
NIM_URL=https://integrate.api.nvidia.com/v1
MODEL_NAME=nvidia/llama-3.1-nemotron-nano-8b-v1
TEMPERATURE=0.1
MAX_TOKENS=512

# GCP FHIR (optional)
GCP_PROJECT_ID=your-project-id
GCP_LOCATION=us-central1
GCP_DATASET=scribe-dataset
GCP_FHIR_STORE=scribe-store

Programmatic Configuration

from goatnote_scribe import GOATScribe, ScribeConfig

config = ScribeConfig(
    nim_api_key="nvapi-...",
    model_name="nvidia/nemotron-nano-9b-v2",
    temperature=0.1,
    max_tokens=512,
    enable_reasoning=True  # Use /think for audit trails
)

scribe = GOATScribe(config=config)

---

Architecture

2-Pass Generation Pipeline

Clinical Text
    ↓
PHI Detection (Presidio)
    ↓
De-identified Text
    ↓
Draft Generation (/think enabled)
    ↓
Self-Critique Pass
    ↓
Final SOAP Note
    ↓
FHIR R4 Bundle
    ↓
GCP Healthcare API (optional)

Key Design Decisions

Model Choice: Nemotron Nano 9B v2

• 6x throughput vs Nemotron Super 49B
• 128K context vs 32K (full encounter support)
• Toggleable reasoning for HIPAA audit transparency
• Well-documented API, proven production use

PHI Detection: Microsoft Presidio

• 18 HIPAA identifier detection
• Anonymization before model inference
• No PHI sent to API endpoints

Memory Safety

• Zero-residue GPU cleanup via wipe()
• Ephemeral processing (no disk persistence)
• CUDA cache clearing after each session

---

HIPAA Compliance

PHI Protection

✅ 18 HIPAA Identifiers Detected:

1. Names
2. Geographic subdivisions (ZIP, city, state)
3. Dates (birth, admission, discharge, death)
4. Phone numbers
5. Fax numbers
6. Email addresses
7. Social Security numbers
8. Medical record numbers
9. Health plan numbers
10. Account numbers
11. Certificate/license numbers
12. Vehicle identifiers
13. Device identifiers
14. URLs
15. IP addresses
16. Biometric identifiers
17. Full-face photos
18. Any unique identifying numbers

Audit Trail Features

• Toggleable Reasoning: /think token preserves model's reasoning process
• Redaction Map: Returns [(entity_type, start, end)] for each PHI detection
• FHIR Bundles: Structured, auditable output format
• Memory Wiping: Zero-residue cleanup via wipe() method

Security Best Practices

# Always wipe after processing
try:
    result = scribe(clinical_text)
    process_result(result)
finally:
    scribe.wipe()  # Ensures cleanup even if error occurs

---

Performance

Throughput (H100 GPU)

Model	Tokens/sec	First Token	Context
Nemotron Super 49B	~250	150ms	32K
Nemotron Nano 9B v2	~1500	<50ms	128K

6x throughput improvement enables real-time clinical documentation

Latency Targets

• Draft Generation: <2 seconds
• Self-Critique: <2 seconds
• Total Pipeline: <5 seconds (including PHI detection)

---

Development

Running Tests

# Install dev dependencies
pip install -e ".[dev]"

# Optional: NVIDIA healthcare add-ons (requires https://pypi.nvidia.com)
pip install --upgrade -r requirements-nvidia-healthcare.txt \
    --extra-index-url https://pypi.nvidia.com

# Run tests
pytest

# With coverage
pytest --cov=src/goatnote_scribe --cov-report=html

Code Quality

# Linting
ruff check src/
black --check src/

# Type checking
mypy src/

Pre-commit Hooks

# Install pre-commit
pip install pre-commit

# Setup hooks
pre-commit install

# Run manually
pre-commit run --all-files

---

API Reference

GOATScribe

class GOATScribe:
    def __init__(self, config: Optional[ScribeConfig] = None)
    
    def __call__(
        self,
        prompt: str,
        audio: Optional[bytes] = None,
        patient_id: str = "anon-001"
    ) -> Dict[str, any]
    
    def wipe(self) -> None

FHIRExporter

class FHIRExporter:
    def __init__(self, config: Optional[ScribeConfig] = None)
    
    def upload_bundle(self, bundle: Dict) -> Dict
    
    def get_patient(self, patient_id: str) -> Dict
    
    def search_documents(
        self,
        patient_id: Optional[str] = None,
        limit: int = 10
    ) -> Dict

---

Roadmap

Version 1.0 ✅ COMPLETE

• ✅ Nemotron Nano 9B v2 integration
• ✅ Emergency Medicine guardrails (vitals, meds, protocols)
• ✅ HIPAA-compliant PHI detection (18 identifiers)
• ✅ FHIR R4 export to GCP Healthcare API
• ✅ One-command deployment (< 5 min)
• ✅ H100 profiling with NCU/Nsight
• ✅ Gradio UI and CLI

Version 1.1 (In Progress)

• NCU CI/CD integration (performance regression detection)
• Comprehensive test suite (pytest + guardrails)
• Example notebooks (chest pain, stroke, trauma)
• Audio transcription (Whisper large-v3-turbo)

Version 2.0 (Planned)

• LoRA fine-tuning for ED specialty (MIMIC-IV-ED)
• Batch processing API
• Multi-language PHI detection (Spanish, Chinese)
• FHIR R5 support
• Clara integration for medical imaging

---

Contributing

See CONTRIBUTING.md for development guidelines.

Reporting Issues

• Security vulnerabilities: Email b@thegoatnote.com (do not open public issues)
• Bugs: Open an issue with reproduction steps
• Feature requests: Open an issue with use case description

---

License

Apache License 2.0 - see LICENSE for details.

---

Citation

@software{goat_scribe_2025,
  title = {GOAT Scribe: H100-Optimized HIPAA-Compliant Medical Scribe},
  author = {Dent, Brandon},
  year = {2025},
  url = {https://github.com/GOATnote-Inc/scribegoat},
  note = {NVIDIA Nemotron Nano 9B v2}
}

---

Contact

Email: b@thegoatnote.com
Organization: GOATnote Inc.
Repository: https://github.com/GOATnote-Inc/scribegoat

---

Acknowledgments

• NVIDIA: Nemotron Nano 9B v2 model and NIM platform
• Microsoft: Presidio PHI detection framework
• Google Cloud: Healthcare API and FHIR infrastructure
• HL7: FHIR R4 healthcare interoperability standards