Client
Self-initiated
Year
2026
Role
AI Engineer · Full Stack
Stack
Python · ViT · Flask · Docker · Hugging Face
Status
Live on HF Spaces
Context
Radiology needs a second pair of eyes.
Pneumonia remains one of the leading causes of death worldwide, and chest X-ray interpretation can be subjective and time-consuming. PneumoDetect leverages Vision Transformers to provide instant, AI-powered pneumonia detection from chest X-rays, designed with a clean clinical interface that puts diagnostics at your fingertips.
The Vision
A first diagnosis that feels certain.
Two principles guided the build. One: the interface must feel clinical, not gimmicky. Clean, calm, glassmorphism with high-tech scanning effects that inspire confidence. Two: the model must be state-of-the-art. ViT (Vision Transformer) was chosen over traditional CNNs for its ability to capture global context in radiographic images.
The Product
Upload, scan, diagnose.
The interface offers drag-and-drop X-ray uploads with instant scan previews. Results come back with confidence metrics and probability distribution visualizations, making the model’s reasoning transparent to the clinician.
Architecture
Under the hood.
Frontend
Clean interface with glassmorphism effects, fluid animations, and real-time scan feedback.
Backend
Streamlined Flask backend handling image upload and ViT inference.
Model
Fine-tuned ViT model for pneumonia detection. Captures global attention in radiographic data.
Runtime
Containerized with Docker for portable, reproducible deployment.
Outcome
Performance metrics.
Challenge
Balancing speed and accuracy.
Medical imaging demands both high accuracy and low latency. Ensemble approaches improve accuracy but add inference time. Switched to a single ViT backbone with global attention, removing traditional preprocessing (CLAHE) entirely — cutting latency by 40% while maintaining strong classification performance.
Performance
Performance
ViT
Model Architecture
Flask
Backend
Docker
Container
None needed
Preprocessing
Dataset
Base Model
Pre-trained ViT
Fine-Tuning
Chest X-Ray dataset
Image Type
Radiographic (X-Ray)
My Contribution
End-to-end pipeline: ViT model fine-tuning, Flask backend, Docker containerization, and Hugging Face Spaces deployment.
Educational/research project — not intended for clinical diagnosis or medical decision-making.
Next case