Data Analyst

Built an Excel-based dashboard tracking $38.6M in spend across 55.7K transactions, using PivotTables, slicers, timeline filters, and KPI cards to surface fraud patterns, category performance, and weekly transaction trends.

Engineered a normalized PostgreSQL analytics platform ingesting 20M+ taxi records through Parquet ETL, a modular Flask REST API, and an interactive React dashboard, reducing fare and demand analysis from 7 hours to 15 minutes.

Investigated ChatGPT and interactive visualizations as AI learning tools across 36 participants, finding interactive visualization yielded 27% higher knowledge gains by applying paired and independent-samples t-tests on pre/post-test scores.

Designed an interactive D3.js platform processing 344K wage records across 6 demographic dimensions spanning 32 years, reducing manual analysis from 1.5 hours to 15 minutes.

Benchmarked 8 traditional and graph-based ML models for illicit transaction detection on Elliptic (203K nodes) and Ethereum datasets. GraphSAGE achieved 98.57% accuracy and 93% Macro F1 on Ethereum while Random Forest achieved 97% accuracy and 88% Macro F1 on Elliptic.

Benchmarked Logistic Regression, LSTM, BERT, and RoBERTa on 44,898 ISOT articles. RoBERTa achieved 99.91% accuracy and 99.92% AUC, with systematic ablation across model families informing architecture selection.

Evaluated privacy risks of Individualized DP-SGD models using black-box membership inference attacks on CNNs trained on CIFAR-10, SVHN, and MNIST, analyzing ROC-AUC scores across privacy budget groups.

Merged MoCA, Chameleon, and COD10K datasets into a unified benchmark and trained YOLOv5 for camouflaged animal detection, establishing baseline results across the combined dataset.

Undergraduate capstone — generated latent embeddings for CIFAR-10, CIFAR-100, and ImageNet using Supervised Contrastive Learning and Autoencoders, and trained the Perceiver architecture on limited compute.

Built a custom dataset of Pakistani textile prints via web scraping and trained generative models including DCGAN, StyleGAN, and VAE to synthesize fabric patterns across 6 pattern categories.

Analyzed universal approximation theory and exponential depth advantages in neural networks, validating theoretical results through experimental implementation.

Modeled human decision-making and cascading behavior using network models and game theory, applied to panic buying during COVID-19 in Pakistan and volatility in the Karachi Stock Exchange.