Keras Content Moderation System

Real deep learning portfolio project for text moderation. It trains a Keras neural network on the UCI SMS Spam Collection dataset, exports the learned vocabulary and weights, and runs actual model inference in the web app.

Model

• Dataset: UCI SMS Spam Collection
• Framework: TensorFlow / Keras
• Architecture: bag-of-words input -> Dense(96, ReLU) -> Dropout -> Dense(32, ReLU) -> Dense(1, sigmoid)
• Test accuracy: 99.1%
• Test F1: 96.5%
• Exported artifacts:
  • public/model/keras_sms_moderation.keras
  • public/model/moderation_model.json
  • reports/keras_moderation_metrics.json

Train

python scripts/train_keras_moderation.py

Run

npm install
npm run dev

Deploy

npm run build
vercel deploy --prod