PedsCore AI — Pediatric Clinical Learning with Alexa+

An agentic pediatric training platform using Alexa+ and MCP to navigate clinical tools, calculate scores and run interactive emergency and mass-casualty simulations.

Comment

Inspiration

As a pediatrician, I have experienced firsthand how difficult it can be to access the right clinical information quickly, especially in time-sensitive environments.

Artificial intelligence offers remarkable possibilities, but there is a fundamental challenge: a language model should not invent clinical calculations or replace professional judgment.

That challenge inspired PedsCore AI.

I wanted to explore a different approach: allowing conversational AI to understand clinical questions and connect professionals with reliable software, while keeping clinical calculations and simulation assessment within deterministic, testable systems.

AI understands the question. PedsCore provides verifiable clinical knowledge.

What it does

PedsCore AI is an open-source pediatric clinical learning platform connected to conversational interfaces through a self-hosted Model Context Protocol (MCP) server.

It provides six integration tools that enable:

  • Discovery of pediatric clinical assessment tools.
  • Retrieval of structured clinical inputs, definitions, and references.
  • Deterministic calculation of supported pediatric scores.
  • Launching synthetic pediatric mass-casualty simulations.
  • Retrieval of individual simulated patient findings.
  • Submission and evaluation of triage decisions.

In the live demonstration, PedsCore calculates an Apgar score of 9/10 using its deterministic clinical engine.

The experience then connects to SIM IMV, a pediatric mass-casualty simulation platform, to launch a school bus accident involving 12 synthetic patients.

For patient 01, the learner submits a GREEN triage category. SIM IMV evaluates that decision using a deterministic implementation of JumpSTART, an internationally recognized pediatric mass-casualty triage algorithm, and provides educational feedback.

The AI orchestrates the process. It does not independently calculate the clinical score or determine the correct triage category.

How I built it

I developed the integration using TypeScript, Node.js, React, and the Model Context Protocol.

The public MCP server implements protocol version 2025-11-25 over Streamable HTTP, making its tools accessible through a standards-based interface.

The system integrates two complementary components:

PedsCore: an open-source pediatric platform containing clinical discovery metadata, structured references, and deterministic calculation engines.

SIM IMV: a pediatric emergency simulation platform that evaluates decisions through defined triage algorithms.

The MCP integration exposes six structured tools, connecting conversational interfaces with these existing systems.

I deployed the public MCP server on Vercel and built a live Judge Console that demonstrates real tool calls and structured responses.

I also implemented protocol tests, deterministic calculation checks, and end-to-end verification of the simulation workflow.

The project follows a deliberate architectural boundary: conversational AI handles intent and orchestration, while clinical and educational logic remains in ordinary, reproducible software.

Challenges I ran into

The most important challenge was establishing a safe boundary between generative AI and clinical logic.

A conversational interface must be able to find and execute a tool without silently reconstructing its underlying medical formula or inventing a result.

Integrating PedsCore with SIM IMV introduced additional challenges involving separate deployments, structured patient findings, protocol compatibility, error handling, and reproducible simulation outcomes.

Deployment reliability was another challenge. During development, I migrated the MCP service to Vercel and verified the complete remote workflow again.

Finally, I needed to create a demonstration that made the technical architecture understandable without losing sight of its educational purpose.

Accomplishments that I'm proud of

I am particularly proud of delivering a working, publicly accessible MCP server rather than only a conceptual AI prototype.

The project includes:

  • Six functioning MCP integration tools.
  • A public interactive demonstration.
  • Reproducible deterministic clinical calculations.
  • An end-to-end pediatric mass-casualty simulation workflow.
  • Automated tests and documented technical evidence.
  • A public, MIT-licensed open-source contribution.

The live integration successfully demonstrates an Apgar score of 9/10 and the correct evaluation of a simulated JumpSTART triage decision.

I am also proud that the project is designed around a clinically important principle: AI should support healthcare professionals, not replace their judgment.

What I learned

This project reinforced a central lesson: conversational intelligence does not need to perform clinical reasoning or arithmetic internally to be useful.

Its value can come from making structured knowledge and established software easier to discover and use.

I also learned how important open standards, reproducible tests, transparent tool contracts, and clear failure handling are when integrating AI into sensitive professional domains.

Building with MCP demonstrated the potential of connecting conversational interfaces to specialized software without surrendering control over the underlying clinical logic.

What's next for PedsCore AI

My next goal is to expand the clinical workflows accessible through MCP while continuing to improve evidence traceability, safety boundaries, and independent clinical review.

I also want to enrich the educational feedback available through SIM IMV, support additional simulation scenarios, and explore new standards-based conversational experiences.

Ultimately, I hope PedsCore AI can contribute to a future in which healthcare professionals benefit from more accessible digital knowledge without sacrificing transparency, reproducibility, or professional responsibility.

PedsCore AI is intended for professional education and informational support. It does not provide autonomous diagnosis, treatment decisions, or a substitute for clinical judgment.

Live demonstration: https://pedscore-ai-mcp.vercel.app/judge-demo

Open-source repository: https://github.com/sferurek/PedsCore

Built With

Share this project:

Updates