Graduate Teaching Assistant

• CSE : 574 - Computer Vision & Image Processing(Fall 2025) Professor Chen Wang
• Guided students & helped them in completion of Capstone Course Project on Computer Vision.
• Solved Doubts of students using online QA platform: Piazza and holding Office hours
• Designed Assignments & evaluated them for 150+ Graduate/Undergraduate students.
• Assisted with grading exams and maintaining course grading systems.
• CSE : 560 - Data Models & Query Language(Summer 2025) Professor Shamsad Parvin
• Guided students & helped them in completion of course Project of developing a Database.
• Solved Doubts of students using online QA platform: Piazza
• Designed Weekly Quiz, Assignments & evaluated them for 50+ Graduate/Undergraduate students.
• Assisted with grading exams and maintaining course grading systems.
• Coordinated with the professor to ensure smooth lectures, handled logistics,and schedule alignment.
• CSE : 396 - Theory Of Computation (Spring 2025) Professor Xiangyu Guo
• Evaluated programming assignments and theoretical problem sets for 80+ undergraduate students.
• Assisted with grading exams and maintaining course grading systems.
• Coordinated with the professor to ensure grading fairness and schedule alignment.

Software Engineer

Java | SpringBoot | Kafka | RabbitMQ | Camunda Modelling | Elasticsearch | AWS | Angular | Node.js | Nest.js | Drools | JUnit | Mockito

• Engineered and maintained 8+ microservices using Java, Spring Boot, Node.js, Kafka, and Couchbase DB, and led integration efforts with other upstream/downstream applications for Western Union.}
• Spearheaded development of BKYC Customer Onboarding, KYC wallet transfers modules in Western Union’s Pharos compliance platform, enabling smooth onboarding and compliance checks for transactions, ensuring adherence to global AML/CTF regulations with 99.98\% accuracy.
• Engineered a 3-tier Camunda BPMN/DMN workflow, automating 90% of compliance issue evaluations and cutting case resolution time from 48 hours to under 4 hours; this reduced manual effort by 65%.
• Authored API documentation, deployment guides, and knowledge articles for microservices, enabling smoother onboarding for new developers and reducing support queries by 30\%.
• Delivered 20+ production releases with <1% post-release defects, ensuring 99.9\% system uptime through active monitoring and rapid incident resolution.
• Assisted in Code Deployments, detailed understanding of CI/CD pipeline, Docker, and AWS containers.
• Project Integration SPOC, Backend Lead, Camunda Desinger during later years.

Softare Engineering Intern

Designed and Developed a Grievance Redressal System for Industry workers for a quick and effective way to resolve any disputes. Implemented two modules Grievance ETA updates and Grievance Ticket Tracker

Worked on MERN stack during the intership period

Generative AI for Emoji Creation using DCGANs

GAN | DCGAN | Conditional-DCGAN | Emoji Generation | Data Cleaning |

• Developed a Deep Convolutional Generative Adversarial Network (DCGAN) to generate novel and unique emojis by learning from the OpenMoji dataset. This project explores the power of generative models in creative design and content creation.

Key Contributions & Impact

• Model Architecture: Implemented a robust DCGAN with a 5-layer Generator that upsamples a 100-dimensional noise vector into $64\times64$ emoji images, and a 5-layer Discriminator to distinguish real emojis from generated ones.
• Conditional Generation: Extended the model to a Conditional DCGAN, enabling the generation of specific emojis based on text labels (e.g., "winking emoji," "crying cat face"). This adds a practical, user-driven component to the creative process.
• Latent Space Analysis: Demonstrated the model's deep understanding of emoji features by smoothly interpolating between different emojis (e.g., transitioning from a "Smiley" to a "Star-Struck" emoji) in the latent space.
• Training & Optimization: Carefully tuned hyperparameters, including using different learning rates for the generator (0.00015) and discriminator (0.000105) with an Adam optimizer, to achieve stable training over 500 epochs.
Project Performance & Visuals
Performance Highlights
The model successfully learned the core features of emojis—like their round, yellow shape and distinct facial expressions—and generated a diverse set of new designs. While some outputs had noise, the model's ability to create hybrid expressions showcases its creative potential. The final model's quality, measured by Structural Similarity Index (SSIM), reached a peak value of 0.24 during training.
Emoji Samples
Sample emojis generated by the DCGAN

Generated Samples
Sample emojis generated by the DCGAN

Conditionally generated emojis based on text labels

Interpolating between 'Smiley' and 'Star-Struck'

Self-Supervised Learning for Image Representation Learning and Classification of Alzheimer Disease

Research Paper accepted at CogMI 2025 Conference

Presented out work at CSE Demo Day 2025, UB

• We developed a self-supervised learning (SSL) framework to classify Alzheimer's disease from MRI scans and demonstrated its cross-domain adaptability to brain tumor classification. This work addresses the challenges of limited annotated data in medical imaging by leveraging large-scale unlabeled MRI datasets (~86,000 images).

Key Contributions

• Model Architectures: Implemented and evaluated three SSL approaches:
  • SimCLR: Contrastive learning with NT-Xent loss.
  • MoCo v3: Momentum contrast with Vision Transformers.
  • BYOL: Bootstrap learning without negative pairs using online-target networks.
• Downstream Classification: Fine-tuned SSL-pretrained encoders on a limited labeled dataset (450 images/class) for Alzheimer's classification using ResNet-50.
• Cross-Domain Transfer: Extended BYOL encoder to classify brain tumor types (e.g., glioma, pituitary, meningioma) using just 300 labeled samples/class.
• Explainable SSL representation : Using Grad-CAM, assissted the doctors to find the important regions in MRI scans and provideing other important statisitics.
Project Performance Details

Performance Highlights

• Accuracy on Alzheimer’s Test Set:
  • MoCo v3: Accuracy 87.0% Model Parameters : 87.8 M
  • BYOL Accuracy: 86.3% Model Parameters : 49.4 M
  • SimCLR Accuracy 82.6% Model Parameters : 26.6 M
  • ResNet-50 (supervised): 74.4%
• Brain Tumor Classification: 86.67% accuracy using BYOL features and minimal labeled data.
• Sample Grad-CAM Image Representations


Technical Pipeline

• Preprocessing: Data cleaning, augmentation (rotation, flipping, jitter), and class balancing.
• Training Strategy: 25–100 epochs of SSL pretraining on unlabeled data, followed by supervised fine-tuning with linear classifiers.
• Model Comparison: Visualizations of parameter counts, class activation maps, and diagnostic regions to interpret model decisions.

Impact

This project showcases the power of self-supervised learning in healthcare AI, demonstrating robust performance in low-label settings and transferable representations across medical domains. Future directions include integrating multi-modal MRI and clinical data and exploring real-time clinical deployment.

Project Poster

• Alzheimer Disease Classifier Website

Inteligent Traffic Management System

Object Detection | Object Classification | YOLOv3 | Logistic Regression | Custom Clock Algorithm | Python

Student Engagement Tracker

Web Developer Volunteer