Predictive Maintenance Machine Learning
- Developed deep learning models (LSTM, CNN, Transformer, Ensemble) using PyTorch and PyTorch Lightning to predict equipment failure, achieving 11.7 RMSE on NASA benchmark data and enabling data-driven maintenance decisions
- Engineered end-to-end ML pipelines using Pandas and Scikit-learn to process 21 sensor channels across 700+ equipment trajectories, implementing automated feature engineering and data quality checks
- Designed MLflow experiment tracking to version models, log metrics, and compare results across 100+ training runs, enabling reproducible, scalable analytics workflows
Python
PyTorch
MLflow
Pandas
NumPy
Scikit-learn
Sentiment Forecasting
- Developed sentiment forecasting pipeline achieving 2.34 Sharpe ratio and -2.5% max drawdown across 180 days of backtesting, by fusing FinBERT transformer outputs with 9 topological features and 6 momentum indicators
- Processed 500+ financial headlines per run through FinBERT (110M parameters) with 32-batch inference, reducing sentiment scoring latency to 3 seconds on CPU
- Implemented Topological Data Analysis using 4-dimensional delay embeddings on 60-day rolling windows, extracting H0/H1 persistence features that captured market regime shifts missed by traditional indicators
Python
Hugging Face (FinBERT)
PyTorch
pandas
scikit-learn
yfinance
Streamlit
MoodVerse
- Developed a desktop diary management system with NLP-driven sentiment analysis and media recommendation, achieving 40% faster entry retrieval, 92–95% sentiment consistency across a validation set of 150+ entries, and producing emotion-aligned media recommendations that boosted perceived relevance by 30%
- Designed a sentiment-to-media recommendation engine that maps emotional tone to curated Spotify/IMDb suggestions, increasing user engagement with recommended content by 35% in prototype testing
Java
Stanford CoreNLP
Spotify Web API
TMDb API