HealthByte
First Place
AI/ML48 hours4 peopleMay 2025
PythonOpenAI o4-miniGemini 2.5 flashReinforcement LearningMulti-Agent SystemsPrompt Engineering
Won 1st place among 20 university teams
Two-agent reinforcement learning loop (persona + editor)
Persona modeling across diverse demographics
Iterative content optimization with convergence tracking
HealthByte - Atlantic AI Summit 2025
First Place Winner | Healthcare Misinformation Simulation Platform
Team HealthByte:
- Eduard Kakosyan (Lead Developer), [email protected]
- Huy Huynh, [email protected], GitHub
- Hao Tang, [email protected], GitHub
- Tobi Onibudo, [email protected], GitHub
Project Overview
HealthByte simulates how different demographics react to healthcare content before it gets published. Instead of waiting to see if messaging lands well, communicators can test articles against synthetic personas and iteratively refine the content.
The Problem: Healthcare misinformation spreads faster than corrections. Traditional responses are reactive — by the time bad messaging is identified, the damage is done.
What We Built: A simulation platform that models diverse public reactions to healthcare content, letting writers optimize messaging proactively.
Results
- First Place at the Atlantic AI Summit 2025 among 20 university teams from across Atlantic Canada
- Built for Nova Scotia healthcare communicators, writers, and reporters
- 48-hour hackathon build
How It Works
Two-Agent Reinforcement Learning Loop
Persona Agent (OpenAI o4-mini)
- Takes demographic and belief profiles from structured JSON data
- Reads healthcare articles and generates three outputs:
- Acceptance Rate: Trust score (0-100)
- Sentiment: Emotional reaction classification
- Reasoning: Why they reacted that way
Editor Agent (Content Optimizer)
- Takes the persona feedback and rewrites the article
- Maintains factual accuracy while improving clarity and trust
- Targets specific audience segments that scored poorly
The Loop
- Runs up to 15 iterations of persona reaction → editor rewrite
- Converges when acceptance rates stabilize across target demographics
- Tracks improvement metrics throughout
Technology Stack
- Backend: Python with structured prompt pipelines
- Models: OpenAI o4-mini for persona simulation, Gemini 2.5 Flash for editing
- Data: Structured JSON for persona profiles and content versioning
- Dashboard: Custom web visualization for tracking content evolution
Dashboard
Try it: HealthByte Dashboard
- Watch personas react to healthcare content in real-time
- Track how articles change through each iteration
- Compare original vs. optimized content side-by-side
- See acceptance rates and sentiment shifts across demographics
Use Cases
- Public Health Campaigns: Test vaccine information across diverse communities
- Medical Communications: Improve patient education materials
- Healthcare Journalism: Help reporters craft clearer articles
- Policy Communications: Test government health messaging before release
Links: