MedGuard AI — Our Story
💡 What Inspired Us
Healthcare is one of the most critical domains in the world — and yet, when we looked at existing AI-based medical tools, we saw a frightening gap. AI models were confidently giving wrong diagnoses. No explanations. No evidence. Just a black box spitting out answers that could cost someone their life.
We asked ourselves: what if AI in healthcare had a guardian? That's where MedGuard AI was born.
🔨 How We Built It
We designed a multi-layered validation pipeline:
- AI Core — processes user symptoms and medical queries
- Structured JSON Output — every response is formatted with required fields:
diagnosis,evidence,confidence,recommendation - Python Watchdog — deterministically validates the output against predefined medical rules
- Confidence Thresholding — any response below a confidence score $C < \theta$ is automatically rejected:
$$ \text{Output} = \begin{cases} \text{Valid} & \text{if } C \geq \theta \ \text{Rejected} & \text{if } C < \theta \end{cases} $$
- RAG Support — responses are grounded in real medical datasets and knowledge bases
📚 What We Learned
- Explainable AI (XAI) isn't optional in healthcare — it's essential
- A deterministic watchdog is far more reliable and cost-efficient than a second AI validator
- Structured outputs dramatically reduce hallucination risk
- Building for safety means designing for failure first
⚡ Challenges We Faced
- Defining the right confidence threshold $\theta$ without making the system too restrictive or too lenient
- Balancing latency vs. reliability — every validation layer adds processing time
- Ensuring the rule-based system covered enough medical edge cases
- Coordinating across the team under hackathon time pressure — including aligning our frontend demo name (Diagnostic AI) with our core branding (MedGuard AI)
Built with passion at the GNEC Hackathon by Team Catalyst.
