MediSync AI

MediSync AI

Comment

MediSync AI: A Strategic Human-AI Collaboration Framework

High-Performance Innovation for the Build with MeDo Hackathon

1. Executive Summary

MediSync AI represents a paradigm shift in healthcare orchestration. It is not merely a dashboard; it is a Strategic Human-AI Collaboration Framework designed to bridge the "Last Mile" of clinical decision-making. By integrating the Gemini 3.1 Pro cognitive engine with a modular Agent-to-Agent (A2A) and Model Context Protocol (MCP) architecture, MediSync AI transforms fragmented patient data into actionable, high-fidelity clinical insights.

2. Inspiration: The "Last Mile" Challenge

The inspiration for MediSync AI stems from a critical observation in modern healthcare: Data abundance vs. Insight scarcity.

Clinicians today are overwhelmed by a "data tsunami" from Electronic Health Records (EHR), wearable devices, and laboratory results. However, the time required to synthesize this data into a coherent clinical strategy is a major bottleneck, leading to clinician burnout and delayed care.

I was inspired to build a solution that treats AI not as a passive search tool, but as an active co-pilot. The "MeDo" platform’s ability to turn natural language descriptions into functional full-stack applications provided the perfect "foundry" to forge this vision. I wanted to create a system that mirrors the collaborative intelligence of a high-performance surgical team, where specialized agents (Triage, Pharmacy, Research) work in synergy under human strategic leadership.

3. Problem Statement: The Fragmentation of Care

The Analytical Perspective

In current clinical workflows, the information entropy $H(X)$ of patient data is high. Data is distributed across disparate silos, making the mutual information $I(X; Y)$ between raw data $X$ and clinical outcome $Y$ difficult to extract.

The problem can be modeled as a optimization challenge: $$\min_{t} \text{Time to Decision}(t) \quad \text{subject to} \quad \text{Accuracy} \geq 99.9\%$$

Key Pain Points:

  1. Cognitive Overload: Sifting through hundreds of data points for a single patient.
  2. Interoperability Gaps: Difficulty in moving context between different specialized domains (e.g., from Triage to Pharmacy).
  3. Static Records: EHRs are often "digital graveyards" rather than living, breathing intelligence systems.

4. The Solution: MediSync AI

MediSync AI solves these problems through a Tri-Layer Intelligence Architecture:

A. The Cognitive Layer (Gemini 3.1 Pro)

At the core is the Gemini 3.1 Pro model, which performs Multimodal Clinical Synthesis. It doesn't just read numbers; it understands the context of those numbers.

B. The Orchestration Layer (MCP/A2A)

We implement a modular agent framework. Each agent is a specialized "micro-service" of intelligence:

  • Triage Orchestrator: Uses real-time vitals to calculate a dynamic Clinical Risk Index (CRI).
  • Pharmacy Liaison: Analyzes drug-drug interactions (DDI) using a knowledge graph approach.
  • Research Agent: Scans global clinical trials to find matches for complex cases.

C. The Human-Strategic Interface

A high-performance dashboard designed for "Mission Control" style oversight. It prioritizes information density without sacrificing clarity, using the Technical Dashboard Recipe for maximum professional utility.

5. How the Project Works: Technical Deep Dive

The Clinical Risk Index (CRI) Logic

MediSync AI calculates a weighted risk score based on vital sign deviations. Let $V$ be the set of vitals ${HR, BP, O_2, Temp}$. The risk $R$ is calculated as: $$R = \sum_{i=1}^{n} w_i \cdot \frac{|v_i - \mu_i|}{\sigma_i}$$ Where:

  • $w_i$ is the clinical weight of the vital sign.
  • $v_i$ is the current value.
  • $\mu_i$ and $\sigma_i$ are the normal mean and standard deviation for the patient's demographic.

Data Flow

  1. Ingestion: Patient data (simulating FHIR standards) is loaded into the state.
  2. Synthesis: The analyzePatientData service sends a structured prompt to Gemini, requesting a JSON-formatted clinical summary.
  3. Context Propagation: The resulting analysis is fed into the Agent Orchestrator, allowing specialized agents to provide "second opinions" based on the synthesized context.
  4. Interaction: The clinician interacts with agents via a conversational interface to refine the care plan.

6. How I Built the Project

Building MediSync AI was an exercise in Strategic AI Co-Development.

  1. Architecture First: I defined the types.ts and constants.ts first to establish a "source of truth" for the data models.
  2. Service Layer: I built the gemini.ts service using the @google/genai SDK, focusing on JSON Schema enforcement to ensure the UI could reliably parse AI responses.
  3. UI/UX Craftsmanship: Using Tailwind CSS and Lucide-React, I implemented a "Hardware/Specialist Tool" aesthetic. I focused on high-contrast elements and monospace fonts for data values to evoke precision.
  4. Iterative Refinement: I used the MeDo platform's multi-turn chat to refactor components for better responsiveness and to add the "Agent Orchestrator" sidebar.

7. Challenges & Overcoming Obstacles

Challenge 1: Context Window Management

Problem: Passing entire medical histories to the AI can be noisy. Solution: I implemented a "Synthesis Step" where Gemini first summarizes the data into a "Clinical Snapshot" before the specialized agents interact with it. This reduces token usage and improves focus.

Challenge 2: UI Information Density

Problem: Too much data can lead to "Alert Fatigue." Solution: I adopted the Bento Grid layout and used AnimatePresence from motion to allow users to toggle between "Overview" and "Deep Dive" views without losing their place.

Challenge 3: Real-time Interoperability Simulation

Problem: Simulating MCP (Model Context Protocol) in a frontend-only demo. Solution: I created a "Protocol Switcher" in the UI that changes the system instructions for the Gemini model on-the-fly, effectively "swapping" the agent's personality and toolset.

8. Lessons Learned

  1. The Power of Constraints: Building for a hackathon forces you to prioritize the "Wow Factor" while maintaining technical rigor.
  2. AI as a Structural Tool: I learned that AI is most effective when it's given a strict schema to follow. Unstructured AI responses are difficult to build apps around; JSON-mode is a game-changer.
  3. Design is a Language: A professional medical app must look the part. Using Inter and JetBrains Mono immediately elevated the project from a "demo" to a "platform."

9. Tech Stack

  • Frontend: React 19 (Vite), TypeScript
  • Styling: Tailwind CSS 4.0
  • Animations: Motion (formerly Framer Motion)
  • Icons: Lucide-React
  • AI Engine: Google Gemini 3.1 Pro via @google/genai
  • Utilities: clsx, tailwind-merge for dynamic class management
  • Platform: MeDo / AI Studio (The "Fantastic" foundation of this project)

10. Future Scalability: The Roadmap to 1.0

MediSync AI is designed to scale horizontally:

  1. Real FHIR Integration: Replacing mock data with live HL7 FHIR API connections for real-world hospital deployment.
  2. Multi-User Collaboration: Implementing WebSockets for real-time "Tumor Board" style collaboration between multiple doctors.
  3. Edge Deployment: Optimizing the agent logic to run on local "Edge AI" hardware for low-latency clinical environments.
  4. Predictive Analytics: Moving from "Risk Assessment" to "Predictive Prognosis" using historical patient cohorts.

11. Conclusion

MediSync AI is a testament to what is possible when human strategic vision meets the generative power of the MeDo platform. By focusing on Interoperability, Clinical Relevance, and High-Performance Design, we have created a solution that doesn't just "work"—it excels.

This project is my contribution to the future of healthcare, where technology serves as a bridge, not a barrier, to human care.

Ariadne-Anne DEWATSON-LE'DETsambali Strategic Lead & Architect

Built With

Share this project:

Updates