Sara

A clinical workflow agent for physicians

Sara is a 4-billion parameter clinical workflow agent capable of orchestrating end-to-end digital clinical tasks. Built on MedGemma and fine-tuned on just 284 examples, Sara outperforms models up to 100x its size on the MedAgentBench clinical benchmark.

What Sara Can Do

Sara executes multi-step clinical workflows against a FHIR R4 server through autonomous GET/POST operations:

• Patient lookup — search by name, DOB, MRN
• Lab result retrieval — magnesium, potassium, HbA1c, glucose
• Vital signs recording — blood pressure, CBG e.t.c
• Medication ordering — replacements with dosing calculations
• Referrals & service requests — orthopedic surgery, follow-up labs
• Care plan management — conditional check-and-order workflows

Each task runs as a multi-turn agent loop: Sara reasons about what FHIR call to make, executes it, reads the result, and decides the next step — up to 8 rounds per task.

Architecture

Platform

Three serverless services on Modal:

Service	Compute	Role
Sara Model	A100 GPU	Serves Sara 1.5 4B via an API
Sara Agent	CPU	Orchestrates agent loop
FHIR Server	CPU	HAPI FHIR R4

The frontend is a Next.js app then deployed on Vercel.

Agent Workflow

The agent receives a clinical task, builds a prompt with FHIR function definitions, and enters a loop: call the model, parse the response (GET / POST / FINISH), execute against the FHIR server, feed the result back, and repeat until the task is complete. All steps are streamed to the frontend in real time via Server-Sent Events.

Fine-Tuning

Base model	google/medgemma-1.5-4b-it
Dataset	Nadhari/MedToolCalling (284 samples)
Method	QLoRA — 4-bit NF4, LoRA r=16 α=32
Trainer	SFTTrainer (TRL) with custom Gemma 3 collator for loss masking
Hardware	NVIDIA H100 80GB, Flash Attention 2
Output	Nadhari/Sara-1.5-4B-it

See Notebooks/ for the full fine-tuning and inference notebooks.

Benchmarking

Evaluated on 300 clinical tasks across 10 task types using the MedAgentBench protocol (pass@1, 8 rounds max, 15 models).

Sara achieves state-of-the-art on 4 tasks (Procedure History at 96.7%, and perfect scores on Patient Search, Allergy Information, and Immunization Records) and has zero invalid actions across all 300 tasks.

See MedAgentBench/ for full results, per-task breakdowns, and the benchmarking script.

Repository Structure

Sara/
├── README.md
├── requirements.txt
├── Sara_platform.png
├── Sara_agent_workflow.png
│
├── MedAgentBench/                  # Benchmarking
│   ├── benchmark_models.py         # Run benchmarks on any OpenAI-compatible API
│   └── outputs/benchmarks/         # Results for 15 models, CSVs, plots
│
├── Notebooks/                      # Fine-tuning & inference
│   ├── Sara.ipynb                  # QLoRA fine-tuning notebook
│   └── Sara_Inference.ipynb        # Inference & demo notebook
│
└── src/
    ├── backend/                    # Modal-deployed Python services
    │   ├── sara_agent.py           # Agent API (FastAPI + SSE)
    │   ├── sara_model.py           # LLM serving (A100 GPU)
    │   ├── fhir_server.py          # HAPI FHIR R4 server
    │   ├── agent.py                # Agent core logic
    │   ├── Dockerfile.fhir         # FHIR server Docker image
    │   └── utils/                  # Action parser, FHIR client
    │
    └── frontend/                   # Next.js clinical UI (Vercel)
        ├── src/app/                # Pages (landing, chat)
        ├── src/components/         # Chat, artifacts, landing, UI
        ├── src/hooks/              # useChat, useStreaming
        └── src/lib/                # API client, task definitions

Getting Started

Prerequisites

• Python 3.9+
• Modal account (backend deployment)
• Docker (local FHIR server)
• Node.js 18+ (frontend)

Backend

git clone https://github.com/Alfaxad/Sara.git
cd Sara
pip install -r requirements.txt

# Deploy all three services to Modal
modal deploy src/backend/sara_model.py
modal deploy src/backend/fhir_server.py
modal deploy src/backend/sara_agent.py

Frontend

cd src/frontend
npm install
cp .env.example .env.local
# Edit .env.local with your Modal endpoint URL
npm run dev

Environment Variables

Variable	Where	Description
NEXT_PUBLIC_MODAL_URL	Frontend	Sara Agent API URL
SARA_URL	Modal (sara_agent)	Sara Model endpoint
FHIR_URL	Modal (sara_agent)	FHIR Server endpoint

Disclaimer

This project is for illustrative purposes only and does not represent a finished or approved product. It is not representative of compliance to any regulations or standards for quality, safety or efficacy. Any real-world application would require additional development, training, and adaptation. The experience highlighted in this demo shows Sara's capability for the displayed task and is intended to help developers and users explore possible applications and inspire further development. Demo data obtained from MedAgentBench.

Citation

@misc{sara-clinical-workflow-agent,
    author = {Alfaxad Eyembe},
    title = {Introducing Sara: A Clinical Workflow Agent},
    year = {2026},
    howpublished = {\url{https://www.kaggle.com/competitions/med-gemma-impact-challenge/writeups/sara}},
    note = {Kaggle}
}