Applied machine learning
Looking closer at synthetic media.
Bringing image and audio classification into an approachable web interface, alongside a dedicated Vision Transformer training workflow.
The problem
A detection model is only one part of the experience. Users also need a simple way to submit media and understand the prediction, while developers need a separate workflow to train and evaluate the model.
How it comes together
Upload
Django views accept image or audio files through dedicated forms.
Prepare
Images are resized and normalized; audio becomes a mel spectrogram.
Infer
TensorFlow and PyTorch models handle the two media paths.
Present
The interface displays the predicted label and model score.
My contribution
- Connected media upload, preprocessing, and inference within the web application.
- Added a Vision Transformer fine-tuning pipeline with class-weighted loss and checkpoint selection.
- Separated model evaluation into a script reporting precision, recall, F1, and a classification report.
The result
The project connects two detection workflows in one application and includes a separate ViT experiment pipeline. The code makes both the user-facing prediction flow and the model-development process available to explore.