10th Opinion

10th Opinion is a quick medical AI diagnostic system that uses 10 agents running with a specialised protocol on AWS Bedrock to analyse patient symptoms and provide diagnostic assessments.

Comment

Inspiration

I've seen friends misdiagnosed with anxiety when they had thyroid conditions, others waited 6 months to discover their fatigue was lupus. Single doctors miss things. But LLMs now cost pennies per query and can simulate different medical perspectives. Why trust one AI when 10 different viewpoints cost less than a coffee?

What it does

Runs 10 AI agents analyzing symptoms from different medical angles: pattern recognition, differential diagnosis, rare diseases, holistic view, consensus building, devil's advocate, evidence validation, hallucination detection, bias checking, and final synthesis. Complete analysis in 30-60 seconds for $0.02. Provides transparent reasoning and confidence scores.

How I built it

Next.js 15.3 + TypeScript frontend. AWS Bedrock (Claude 3.5 Haiku) for cost-effective AI. DynamoDB with 24-hour auto-deletion. S3 for image uploads. Each agent gets specialized prompts. Added retry logic for API failures and regex parsing for malformed JSON. Mobile route at /mobile for field use.

Challenges I ran into

UI/UX. Malformed JSON responses happen ~5% of the time.

Accomplishments I'm proud of

Multi-agent consensus catches diagnoses single models miss. Dedicated bias detection agents. $0.02 per complete analysis. Transparent reasoning users can interrogate.

What I learned

Multiple specialized prompts beat one complex prompt. Users tolerate 30-second waits if they see progress. Medical tools need disclaimers everywhere. Parsing AI responses needs defensive coding. Real mobile optimization means bigger buttons, not just responsive CSS.

What's next for 10th Opinion

Build native iOS/Android apps with edge computing for remote areas. Deploy models to edge devices for true offline diagnosis. Add Spanish, Hindi, Mandarin, and Arabic interfaces. Create provider dashboard tracking diagnostic accuracy and agent agreement patterns. Implement automatic physician escalation when agents disagree or confidence drops below 60%.

Share this project:

Updates