MeHelper

Inspiration

Healthcare inequality is one of humanity's most pressing challenges. While urban populations enjoy access to world-class medical facilities, 3.5 billion people worldwide lack essential health services. In remote villages across Africa, isolated communities in rural America, and disaster-stricken regions globally, the nearest medical facility can be hours or days away.

Our inspiration came from a sobering statistic: 1 in 4 deaths in rural areas could be prevented with timely medical guidance. During the COVID-19 pandemic, we witnessed how telemedicine became a lifeline for many, yet remained frustratingly inaccessible to those who needed it most. We envisioned a world where geography never determines access to life-saving medical knowledge.

The spark for MeHelper came from stories of mothers walking for days to reach a clinic, only to be told their child's condition could have been managed at home with proper guidance. We realized that AI, specifically open-weight models like GPT-OSS, could democratize medical expertise and bring intelligent healthcare guidance to every corner of the world.

What it does

MeHelper is an AI-powered medical triage system that transforms any smartphone or tablet into an intelligent medical assistant. Our system provides:

🩺 8-Level Comprehensive Medical Triage

Reassurance Level - Immediate comfort and context
Initial Assessment - Risk stratification (mild/moderate/severe/emergency)
Pathological Possibilities - Evidence-based condition analysis
First Aid Measures - Actionable home care instructions
Danger Signs Alert - Critical warning indicators
Vital Signs Analysis - Temperature and heart rate interpretation
Summary Report - Complete assessment with clear next steps
AI Vision Analysis - Medical image interpretation for visible symptoms

🤖 Multi-Modal AI Intelligence

Text Analysis: Natural language symptom processing using GPT-OSS-20B
Image Recognition: Visual symptom analysis for wounds, rashes, and swelling
Risk Stratification: Demographic and temporal risk factor integration
Emergency Detection: Automatic escalation for life-threatening conditions

🌍 Offline-First Design

Core functionality works without internet connectivity
Progressive enhancement for optimal performance in any environment
Local first-aid resource library
GPS emergency location sharing

How we built it

Our architecture leverages GPT-OSS-20B as the primary reasoning engine, perfectly aligned with the OpenAI Open Model Hackathon's vision of demonstrating open-weight model capabilities.

AI Orchestration Architecture

# Multi-modal AI pipeline
class TriageAI:
    def __init__(self):
        self.primary_model = GPTOSS20BService()  # Main reasoning
        self.vision_model = GeminiVisionService()  # Image analysis
        self.fallback = MockAIService()  # Offline capability

Our triage algorithm processes multiple data streams simultaneously:

$$\text{Risk Score} = \alpha \cdot f_{text}(\text{symptoms}) + \beta \cdot f_{image}(\text{visual data}) + \gamma \cdot f_{vitals}(\text{physiological})$$

Where:

$f_{text}$ represents GPT-OSS-20B's natural language analysis
$f_{image}$ captures visual symptom recognition
$f_{vitals}$ processes temperature and heart rate data
$\alpha, \beta, \gamma$ are weighted confidence factors

Technical Stack

Backend (Python/Flask)

GPT-OSS-20B Integration: Via Hugging Face Inference API with local fallback
Multi-Service Architecture: Parallel processing of text and image analysis
Medical Protocol Engine: Rule-based emergency keyword detection
API Design: RESTful endpoints for triage analysis and image processing

Frontend (Progressive Web App)

Vanilla JavaScript: Maximum compatibility and offline capability
Service Worker: Intelligent caching for offline-first functionality
Responsive Design: Mobile-first approach for global accessibility
Audio Feedback: Text-to-speech for accessibility in low-literacy areas

AI Integration Strategy

// Parallel AI processing for optimal performance
const analyzeSymptoms = async (data) => {
    const [textAnalysis, imageAnalysis] = await Promise.all([
        processWithGPTOSS(data.symptoms),
        analyzeImage(data.image)
    ]);
    return consolidateResults(textAnalysis, imageAnalysis);
};

Challenges we ran into

Open-Weight Model Integration Complexity

Initially, we explored specialized medical models but discovered many lack inference API support. This led us to architect a hybrid approach where GPT-OSS-20B provides the core medical reasoning while complementary models handle specific tasks like image analysis.

Medical Accuracy vs. Accessibility Balance

Creating prompts that elicit professional-grade medical analysis from GPT-OSS-20B while remaining comprehensible to non-medical users required extensive prompt engineering. We developed a multi-stage prompt architecture:

prompt = f"""
You are an expert medical triage AI. Analyze: {symptoms}
Provide assessment in exactly this JSON format:
{{"level_1_reassurance": "...", "level_2_assessment": {{...}}}}
Be precise, medically accurate, and prioritize patient safety.
"""

Offline-First Architecture Design

Building a medical application that works reliably without internet required rethinking traditional web development patterns:

Service Worker Complexity: Ensuring critical medical resources remain cached
Progressive Enhancement: Graceful degradation when AI services unavailable
Data Synchronization: Managing offline symptom tracking and emergency protocols

Cross-Cultural Medical Adaptation

Designing triage protocols that work across different healthcare systems and cultural contexts while maintaining medical accuracy posed significant challenges in prompt design and user interface considerations.

Accomplishments that we're proud of

🏥 Real-World Impact Potential

Built a system that could genuinely save lives in underserved communities where healthcare access is limited.

🤖 Advanced Open-Weight Model Implementation

Successfully demonstrated GPT-OSS-20B's capabilities in a high-stakes medical application, proving open models can handle critical decision-making tasks.

🌍 Global Accessibility Achievement

Created an interface that works equally well on a $50 smartphone in rural Kenya and a tablet in an Alaskan fishing village.

⚡ Performance Excellence

Achieved sub-3-second comprehensive medical analysis, even with complex multi-modal AI processing.

📱 Offline-First Innovation

Implemented sophisticated service worker architecture ensuring core medical guidance remains available even in areas with no cellular coverage.

🔬 Medical Protocol Adherence

Successfully implemented proper medical disclaimers, emergency escalation protocols, and safety-first design principles throughout the application.

What we learned

Open-Weight Model Potential

Working intensively with GPT-OSS-20B revealed the extraordinary potential of open-weight models for specialized applications. The model's reasoning capabilities in medical contexts exceeded our expectations, particularly in:

Complex symptom correlation and analysis
Risk stratification across age groups and demographics
Natural language understanding of medical terminology
Contextual awareness for emergency vs. routine conditions

Medical AI Ethics and Responsibility

Developing healthcare AI taught us critical lessons about:

Transparent Limitations: Clear communication about AI capabilities and boundaries
Safety-First Design: Every feature must prioritize patient safety over convenience
Cultural Sensitivity: Medical guidance must adapt to local healthcare contexts
Emergency Protocols: Robust escalation procedures for life-threatening situations

Offline-First Development Philosophy

Building for connectivity-challenged environments fundamentally changed our approach:

Progressive Enhancement: Starting with core functionality, then adding AI layers
Graceful Degradation: Ensuring the app remains useful when AI services fail
Local-First Data: Prioritizing client-side storage and synchronization patterns

Multi-Modal AI Orchestration

Coordinating text and image analysis taught us:

Parallel Processing: Time-sensitive medical applications require concurrent AI model execution
Error Handling: Robust fallback strategies across multiple AI services
Prompt Engineering: Medical context requires highly specific and structured prompts

What's Next for MeHelper

Vision

Make MeHelper the foundation of a Global Healthcare Equity Initiative, ensuring that geography and resources no longer determine access to primary care.

Roadmap

Phase 1 (6 months): Multi-language support, partnerships with NGOs, field testing in remote villages, stronger offline mode.
Phase 2 (12 months): Training platform for health workers, telemedicine integration, predictive health dashboards.
Phase 3 (18 months): Run models fully on smartphones, wearable device integration, federated learning for medical AI.
Phase 4 (24 months): Government partnerships, disaster response network, open health research platform, sustainable global model.