Projects

A collection of my cool projects and research work.

LLM Network Acceleration and Compression Research (1yr Capstone Project with Proofpoint)

LLM Network Acceleration and Compression Research (1yr Capstone Project with Proofpoint)

Finetuned 3B parameter LLMs with quantization, distillation, and pruning to achieve 50% faster inference with less than 2% accuracy loss for generalized cyber threat detection in emails, documents, and other communication channels.

Code Repo
Real-Time Facial Emotion Detection from Partially Obstructed Faces using CNNs

Real-Time Facial Emotion Detection from Partially Obstructed Faces using CNNs

Utilized VGG 16, outperforming ResNet50 by 7%, to detect emotions from partially obstructed faces, applied saliency maps for interpretability, and integrated it with OpenCV haar cascade for real time emotion detection.

Code Repo
Algorithmic Options Trading with Sentiment Analysis and Delta Hedging

Algorithmic Options Trading with Sentiment Analysis and Delta Hedging

Built an algorithmic options trading strategy using sentiment adjusted implied versus realized volatility spreads with delta hedging to remain market neutral, achieving consistent risk adjusted returns in backtests across volatile and normal markets.

Report
Reusable Cup Circulation and Tracking Application

Reusable Cup Circulation and Tracking Application

Designed and implemented a full stack application to manage reusable cup circulation, enabling user registration, borrowing and return tracking, and real time inventory monitoring to support sustainable consumption workflows.

Code Repo
Domain Adaptive Ophthalmic Image Segmentation using Transfer Learning

Domain Adaptive Ophthalmic Image Segmentation using Transfer Learning

Improved Diabetic Macular Edema detection on limited OCT images by finetuning MM Segmentation models, achieving mDice 79.38% and mIoU 70.64% with a UNet model using DeepLabV3 heads.

Report
Customer Profitability Analysis and Predictions using LSTM for Targeted Marketing

Customer Profitability Analysis and Predictions using LSTM for Targeted Marketing

Classified customer profitability into three tiers using K-means clustering on standardized monthly sales data, then built a customized LSTM model that achieved 96.65% accuracy and improved targeted marketing strategies to drive revenue growth.

Code Repo
Opioid Shipment Policy Difference-in-Difference Analysis

Opioid Shipment Policy Difference-in-Difference Analysis

Applied a difference in differences causal inference framework to estimate the effect of state level opioid shipment regulations on per capita shipments and overdose mortality, isolating policy driven changes from underlying time trends.

Code Repo