About This Project
Built a deep learning NLP system to classify 200,000+ PubMed RCT sentences using a Tribrid architecture combining token, character, and positional embeddings. Improved model accuracy from 72.5% to 85.6% through feature engineering, architecture tuning, and hyperparameter optimization. Developed an end-to-end ML pipeline covering data preprocessing, model training, inference API, and deployment using Flask and Docker. Implemented CI/CD workflows with GitHub Actions, reducing manual build time by 45%, and deployed a Chrome extension for real-time biomedical sentence classification.
Key Features
- Tribrid embedding architecture (token + character + positional)
- 85.6% accuracy on 200k+ PubMed RCT sentences
- End-to-end ML pipeline: preprocessing → training → inference API
- Flask + Docker deployment with CI/CD via GitHub Actions
- Chrome extension for real-time in-browser classification
- Color-coded sentence labeling across 5 medical categories
