Spark Triage: Building an AI Assistant for Emergency Care

This is a submission for the Hacktoberfest Open-Source AI Challenge Week 2: Side Hustle.

Spark Triage: Building an Open-Weight AI Assistant for Emergency Care

In an emergency department, every minute matters. Healthcare professionals often have to assess several patients under pressure, organize clinical information, and determine who needs attention first.

I built Spark Triage to explore how AI agents and deterministic clinical rules can support this process while keeping healthcare professionals in control.

My goal is to build something that can eventually be adapted to healthcare environments in Nigeria, including facilities where internet connectivity and resources may be limited.

What I Built

Spark Triage is an AI-powered emergency triage desk assistant that helps organize patient information and generate proposed urgency levels for clinical review.

The system explores how AI can help process presenting complaints, identify supporting evidence, and organize triage workflows. It combines an AI agent with deterministic rules based on the adult South African Triage Scale (SATS) ruleset.

Rather than allowing a language model to make the final decision on its own, the system separates AI assistance from rule-based scoring. A nurse or clinician can review, confirm, or override a proposed result.

I see potential applications in hospitals, emergency departments, and healthcare organizations looking to improve the organization of their triage workflows. However, Spark Triage is still a prototype and requires further clinical validation before it can be used for real patient care.

Try the Demo

Live demo: Open Spark Triage

Please use synthetic patient information only. The prototype is intended for demonstration and evaluation, not for making real clinical decisions.

Source Code

GitHub repository: github.com/onloopdev/spark-triage

The repository contains the project implementation and documentation for its triage workflow, AI-agent integration, and deployment approach.

How I Built It

I built Spark Triage using Python, an AI-agent workflow, open-weight language models, Ollama, and the Model Context Protocol (MCP).

The architecture separates the main responsibilities:

• AI agent: Helps process presenting complaints and identify supporting evidence.
• Deterministic triage rules: Calculate proposed urgency using coded rules instead of relying solely on the language model.
• MCP integration: Provides a structured way for the agent to interact with application tools and operations.
• Human oversight: Keeps the clinician involved in reviewing and confirming or overriding proposed results.
• Audit and evaluation: Supports the inspection of workflow actions and assessment of agent behaviour.

The development configuration defaults to Qwen2.5:3B, while the Docker configuration specifies Gemma 4:31B through an Ollama-hosted endpoint.

Online Demo, Open-Weight Models, and Hospital Deployment

The current public demo uses an online Ollama inference endpoint. For hospital deployment, my intended approach is to run a compatible open-weight model on infrastructure controlled by the hospital, following the deployment options documented in the GitHub repository.

This distinction is important. I want hospitals to have the option of controlling where AI inference takes place and how patient information is stored, rather than depending entirely on an external AI service.

A hospital-controlled deployment could reduce dependence on external inference services and support greater control over sensitive data. However, local deployment, offline operation, secure storage, and recovery during power or network interruptions must be implemented and properly tested before those capabilities can be promised.

My longer-term goal is to adapt Spark Triage to the realities of Nigerian healthcare facilities, where affordability, privacy, infrastructure, and reliable access to care all matter.

Why Open Innovation Matters

For me, open innovation means having the freedom to adapt technology to the needs of the people it is intended to serve.

Open-weight models allow developers to experiment with different models, explore local inference, and reduce dependence on proprietary AI APIs. In healthcare, these choices matter because privacy, infrastructure costs, data control, and accountability are just as important as model performance.

I am particularly interested in the potential for this approach in Nigeria, where some healthcare facilities face unreliable internet connectivity and limited resources. A properly configured local deployment could give hospitals more control over their AI infrastructure and reduce reliance on external services.

Open-source development also creates opportunities for developers and qualified healthcare professionals to inspect the implementation, identify weaknesses, and contribute improvements.

I want Spark Triage to grow through collaboration and careful evaluation, with clinicians remaining responsible for clinical decisions.

What Comes Next?

My next priorities are to improve the agent workflow, strengthen testing and evaluation, document deployment requirements, and explore how the system could be adapted to local healthcare needs.

Before real-world clinical use, the system will need appropriate clinical review, validation, privacy safeguards, and testing in the environments where it is intended to operate.

I see Spark Triage as a starting point: an opportunity to explore how open-weight AI and transparent rules can support healthcare workflows without removing human judgment from the process.

Explore the Project

• Live demo: Spark Triage
• Source code: GitHub — onloopdev/spark-triage

Thanks for checking out Spark Triage. Feedback, technical suggestions, and contributions are welcome.