A GenAI-powered email classification system that uses fine-tuned LLMs and a Streamlit chatbot interface to classify Microsoft Outlook emails by category (Finance, HR, Legal, Admin) and priority (High, Medium, Low).
Features
Fine-tuned sentence-transformers/all-MiniLM-L12-v2 using LoRA adapters.
Parameter Efficient Fine Tuning with LoRA
Dual-headed classification (category & priority)
Lightweight Streamlit UI with:
Support for .eml and .msg uploads and parsing
Manual text input
Chat interaction using Groq or OpenAI for email summaries and alt classifiers
OOP architecture will be added at a later time
Model Training Summary
Dataset: Sampled 12,000 emails from Enron corpus
Heuristic Labeling Spam Detection: Applied keyword-based heuristics to label emails as spam (e.g., 'offer', 'win', 'urgent') and filtered to ~5% to avoid skewing training.
Category & Priority Labels: Where explicit labels were not available, category and priority were inferred using logical keyword matching and subject/body context.
Note: As this is a weakly supervised setup, final evaluation metrics may reflect some noise in the labels. Manual labeling or active learning could improve this in future iterations.
Balanced by class using upsampling/downsampling
Fine-tuned with LoRA to reduce memory footprint
Evaluated with accuracy, F1, and confusion matrix
Quick Start
- Clone repo
git clone https://github.com/your-username/enron-email-classifier.git cd enron-email-classifier
- Install dependencies
pip install -r requirements.txt
- Run the app
streamlit run chatbot_ui.py