Projects

BirdCLEF+ 2025 – Acoustic Biodiversity Challenge

Conference: CLEF 2025 | Technologies: Python, XGBoost, Spectrogram Analysis, Data Augmentation

• Developed a scalable one-vs-rest detection pipeline with 206 per-species XGBoost models trained on log-mel spectrogram statistics and geo-contextual metadata.
• Engineered 3,200-dimensional sparse binary features via adaptive STFT-based frequency binning with class balancing.
• Applied targeted augmentation and rebalancing, improving F1 scores from 0.00 → 0.72 for low-resource species.
• Achieved >0.90 mean AUC across species in final evaluations.

Data Ambiguity Quantification for Robust ML using Optimal Transport

Technologies: Python, Keras, Scikit-learn, Optimal Transport

• Designed a framework leveraging Optimal Transport to maximize Wasserstein distance between optimized distributions, providing a rigorous upper bound on generalization loss (10× improvement over baselines).
• Implemented Leave-one-out CV and entropy-based ranking to prioritize ambiguous samples for efficient annotation and improved classification accuracy.

NeuroFraudGAN – Credit Card Fraud Detection with Synthetic Data

Technologies: Python, TensorFlow, GAN, Auto-NAS, Optimization

• Synthesized 100k+ credit card transactions from JPMorgan’s AIRSyntheticData, reducing class imbalance by 40%.
• Applied Auto-NAS for neural network optimization, improving training efficiency and scalability.
• Achieved 96% accuracy and 0.97 AUC-PR through GAN-augmented training and hyperparameter tuning.

Examination of Toxic Tweet Classification on Twitter

Technologies: Python, NLP, Transformers, Scikit-learn

• Conducted large-scale analysis on 500k+ tweets for hate speech and toxicity detection.
• Built a two-phase pipeline: sentiment filtering with logistic regression → toxicity categorization with deep models.
• Benchmarked traditional ML (TF-IDF + LR) against transformer models (BERT), with class accuracies between 93–99%.

Evaluating ML & DL Architectures for NLP Tasks

Technologies: Python, TensorFlow, Keras, Scikit-learn

• Achieved 88% accuracy in sentiment analysis with optimized logistic regression.
• Attained 99% accuracy in NER using LSTM, outperforming classical models by 6%.
• Enhanced text generation quality with LSTM and GRU through embedding fine-tuning.

Social Media Marketing Optimization for Lead Generation

Technologies: A/B Testing, Tableau, Google Analytics

• Reduced CPC and CPM by 20% via A/B testing and targeting optimization.
• Built interactive dashboards for CTR, ROAS, and relevance scores, enabling data-driven marketing.
• Achieved 30% cost efficiency improvement through optimized budget allocation.

Data Visualization in Football

Technologies: Python, Tableau, Pandas, TensorFlow

• Analyzed 10,000+ player attributes, identifying KPIs with 0.87 correlation to overall ratings.
• Generated 500+ shotmaps and heatmaps for strategic team analysis.
• Built position forecasting models with deep learning on player attributes.

Image Compression Analysis (Coursework Project)

Technologies: MATLAB, Python, Linear Algebra, SVD

• Compared SVD-based compression with JPEG2000, evaluating CR, PSNR, and SSIM.
• Demonstrated JPEG2000’s superiority for medical imaging and archival storage.
• Highlighted SVD’s flexibility for research and teaching contexts.