Sona | Devpost

Patients
New Patient
Scans
Reports
New Scan
Scan Results
New Chat with Bernie
Generating Report
Example Report

Inspiration

Radiologists play a critical role in healthcare, they’re the ones who catch life threatening diseases early, guide treatment decisions, and often provide answers when patients need them most. But training the next generation of radiologists is tough. Students rarely get hands on experience with real medical dashboards, and learning how to interpret complex scans often feels disconnected from real-world practice. We wanted to bridge that gap with an AI-driven educational platform that makes radiology training more interactive, realistic, and effective.

What it does

Sona is an all-in-one platform for radiology students. It helps them learn how to interpret thoracic scans using our custom AI model, practice generating reports, and retain knowledge through interactive quizzes. The interface is designed to resemble real medical radiology dashboards, so students train in an environment that mirrors what professionals actually use in hospitals and clinics.

Features

Thoracic Scanning:A custom CNN model trained to detect 16+ thoracic diseases.
Report Generation: Automatically generate test questions from patient reports using Anthropic LLMs to reinforce student learning.
Ask Bernie: An AI companion that answers any radiology question, whether about patterns in scans or deeper medical knowledge.

Technologies

Thoracic Model — PyTorch + DenseNet121 (with last layer replaced by sigmoid classification for 16 disease outputs) + GRAD-CAM visualization.

Frontend — Next.js + TypeScript.

UI Libraries — Lucide-React, Headless-UI, Framer-Motion.

LLM — Anthropic.

Databases/Backend — Firebase + Supabase (Postgres) integration.

Challenges we ran into

Integrating so many different technologies into one cohesive platform was one of the toughest parts. Managing two separate backends (Firebase and Supabase) due to subscription issues forced us to carefully design data flows so both databases could stay in sync. Balancing AI model training, frontend performance, and real-time interactions made development both challenging and rewarding.