PrediAI - Medical System
A revolutionary mobile health application designed for non-invasive early diabetes screening. Developed for the National GEMASTIK XVIII competition, PrediAI uses Deep Learning to analyze visual biomarkers on the tongue and nails, eliminating the need for painful needles.
The Challenge
Traditional diabetes testing requires invasive blood sampling (needles), causing pain and reluctance for many patients. Early detection is often skipped due to fear, cost, or lack of access to medical facilities in remote areas.
The Solution
We engineered a multi-modal AI system combining YOLO (for object detection) and MobileNetV2 (for classification). Users simply take a photo of their tongue or nail, and the app processes it via a FastAPI backend to predict diabetes risk. This is integrated with a Gemini-powered health assistant for personalized advice.
Key Features
Technology Stack
Kotlin
Native Android development with Jetpack Compose.
FastAPI
Backend service for hosting YOLO & MobileNetV2 models.
TensorFlow
Deep Learning framework for image classification.
Gemini API
Generative AI for chat assistant and lab analysis.
Development Process
Research & Dataset
Collected medical datasets for tongue and nail biomarkers linked to hyperglycemia.
Parallel Development
While the AI team trained MobileNetV2/YOLO models, I (Frontend) simultaneously sliced the UI using Jetpack Compose.
System Integration
Connecting the Android frontend to the FastAPI backend and integrating Firebase for auth.
Testing & Validation
Validating model accuracy (F1 Score > 0.90) and ensuring smooth UX flow.
Final Presentation
Pitching the product at the National GEMASTIK XVIII finals.
Engineering Challenges
Image Precision
Raw photos often contained noise. We implemented a YOLO Detection Service to precisely crop/localize the tongue and nail area before classification.
Complex Integration
Orchestrating multiple APIs (Maps, YouTube, Gemini, Inference) required a robust Repository pattern in the Android architecture to handle asynchronous states.