- Mechanistic Interpretability: Building tools to understand how transformers process information and form internal representations
- LLM Training Efficiency: Investigating novel architectures and optimization techniques for parameter-efficient language models
- Foundation Model Safety: Developing evaluation frameworks and alignment techniques for safer, more reliable LLMs
Interactive interpretability tool for transformer models
- Visualizes attention patterns across all layers for models up to 1.3B parameters
- Interactive heatmaps showing token-to-token attention weights and head specialization
- Impact: Helps researchers understand how LLMs process different text types
- Features: Layer-wise analysis, attention head clustering, prompt comparison mode
- Tech: PyTorch, Transformers, Streamlit, Plotly, NumPy
Explainable AI for financial applications
- Processes 10M+ transactions with <50ms latency using ensemble models
- SHAP-based explanations for regulatory compliance and model transparency
- Performance: 35% reduction in false positives vs baseline systems
- Tech: XGBoost, SHAP, Kafka, FastAPI, Redis
Deep learning for healthcare diagnostics
- CNN ensemble for pneumonia detection achieving 94.2% accuracy on chest X-rays
- Includes uncertainty quantification and confidence scoring for clinical decisions
- Dataset: Trained on 5,856 chest X-ray images with data augmentation pipeline
- Deployment: Streamlit web app with DICOM file support and batch processing
- Tech: TensorFlow, Keras, OpenCV, Streamlit, Matplotlib
End-to-end ML pipeline for real estate valuation
- Complete ML workflow from data collection to model deployment and monitoring
- Feature engineering with 80+ variables including location, demographics, and market trends
- Performance: Achieved MAE of $12,500 on California housing dataset (15% improvement)
- Production: Flask API with automated retraining pipeline and A/B testing framework
- Tech: scikit-learn, Flask, PostgreSQL, Docker, Pandas
- Advanced transformer architectures (Mamba, Mixture of Experts)
- Reinforcement Learning from Human Feedback (RLHF)
- Constitutional AI and alignment techniques
- MLOps best practices for large-scale model deployment
When I'm not decoding neural networks, you'll find me capturing the beauty of nature through photography 📸 or building sustainable tech solutions for environmental challenges! 🌱
"Understanding intelligence—whether artificial or natural—requires building it piece by piece." - Personal Research Philosophy