|

Comprehensive comparison of game-theoretic and deep RL approaches for Texas Hold'em poker, implementing MCCFR, DQN, and NFSP algorithms.

• • Implemented MCCFR (95.3% win rate, +119.26 BB/100), DQN with round-specific models, and NFSP v2 with pot-aware reward shaping (93.9% win rate, +87.51 BB/100)
• • Built evaluation framework testing agents against Random and OddsAgentV21 with statistical analysis (confidence intervals, p-values, BB/100 metrics)
• • Developed GPU-accelerated training pipelines using TensorFlow/PyTorch with RLCard, including visualization tools and model persistence

ML system that recommends Steam games to users by combining user preferences, game statistics, and semantic text embeddings from reviews.

• • Achieved 92.79% AUC-ROC and 95.17% precision by combining statistical features, user encoding, and SBERT semantic embeddings, with 20.9% improvement over baseline
• • Tested multiple approaches incrementally on ~10,000 games and ~25,000 reviews, showing how each feature type (statistics, user patterns, text semantics) contributes to performance
• • Handled cold-start scenarios achieving 85.32% AUC-ROC for new users with limited history, making it production-ready for real-world deployment

Streamlit app that suggests recipes from text or pantry photos using image recognition, fuzzy matching, and nutrition-aware ranking.

• • Combines text input + EfficientNetB0 ingredient detection with OpenCV preprocessing
• • Fuzzy ingredient and recipe matching with duplicate handling and quick delete actions
• • Scores and ranks recipes by match overlap and nutrition, served via Streamlit UI

ML-driven dashboard predicting which YouTube trending videos sustain virality beyond a day, with survival analysis visuals.

• • Automated daily YouTube Data API pulls to build time-stamped trending snapshots
• • Random Forest classifier optimized for recall plus Kaplan-Meier survival curves
• • Streamlit UI highlighting top viral candidates and feature importance insights

Interactive data visualization exploring earthquake impacts based on infrastructure resilience and preparedness.

• • Built 3D interactive globe with D3.js for real-time earthquake visualization with dynamic tooltips
• • Developed educational storyline with scroll-based storytelling and energy calculator tools
• • 🏆 Won Best Project Award & People's Choice Award among 42 teams

Power outage prediction using machine learning techniques.

• • Built end-to-end ML pipeline with data preprocessing and feature engineering
• • Optimized Random Forest using GridSearchCV with cross-validation
• • Achieved 74.3% accuracy with precision, recall, and F1-score metrics

Hybrid LSTM/GRU model with Monte Carlo simulations for stock price forecasting.

• • Built hybrid LSTM/GRU model with Monte Carlo simulations for stock forecasting
• • Integrated technical indicators, sentiment analysis, and risk metrics
• • Deployed modular Streamlit app with yFinance and REST API

CNN-based SAR image classification with speckle filtering for land cover analysis.

• • Developed CNN model with ResNet50 for SAR image classification achieving 80% accuracy
• • Applied Lee and Gamma MAP speckle filters to enhance SAR image quality
• • Conducted extensive evaluation using precision, recall, F1-scores, and confusion matrices