Research Interests

I'm interested in computer vision, natural language processing, deep learning, generative AI, AI in healthcare, human-computer interaction, and especially multi-modal learning that integrates vision and language.

Publications and Research Projects

Developed and benchmarked machine learning and deep learning models for recognizing physical activity in 10–15-month-old toddlers using hip-mounted accelerometer signals during free play. The study compares classical ML, window-based deep learning, and dense prediction models under subject-wise evaluation, with activity-time estimation as the primary benchmark. Establishes ecologically meaningful baselines for toddler activity recognition across carried, sedentary, light-intensity, and moderate-intensity activity groups, showing that duration-based evaluation better reflects real-world physical activity behavior than frame-level accuracy alone.

Comprehensive survey on Small Vision-Language Models (SVLMs), covering architectural developments, training pipelines, and optimization strategies. Analyzes model efficiency and performance trade-offs, focusing on CNN and ViT-based vision encoders, multi-stage training taxonomies, and optimization techniques like quantization and MoE.

Novel deep learning framework for diabetic foot ulcer segmentation featuring learnable image preprocessor, hierarchical encoder, and boundary enhancer. Achieves state-of-the-art performance on DFUC (60.06/70.78%), FUSeg (79.06/85.76%), and AZH Wound (81.21/87.28%) datasets through innovative multiresolution positional attention and local-global feature aggregation.

Hybrid deep learning and GBDT ensemble for skin cancer detection using 3D-TBP images and synthetic lesion augmentation.

Comprehensive review of e-commerce recommendation systems addressing challenges like cold-start problems, sparse datasets, and innovative approaches including deep learning integration.

Optimized YOLOv5-based approach for real-time vehicle detection at road intersections using fisheye cameras, achieving 13.7% improvement in mAP @ 0.5 over baseline YOLOv5.

Robust deep learning framework with Process Convolution blocks and Dual Feedback feature accumulation for automated pneumonia detection from chest X-ray images, achieving 97.78% accuracy.

Industry Projects

Developed an automated e-KYC pipeline including Face-Match, Video Liveness Detection, and National ID (NID) info extraction services. The system reduced manual verification time by 80% and currently serves 500,000+ users with NIST-compliant facial recognition.

Developed two recommendation systems: (1) Food menu recommendation system using content-based, SVD-based, and hybrid models to suggest relevant menu items to customers, and (2) Coupon recommendation system integrating RFM analysis and demographic features to identify users likely to use coupons, achieving 18% increase in coupon redemption rates. Implemented KNN-driven customer segmentation analyzing 10M+ user data on Cloudera, improving overall coupon ROI by 15%. Built Flask APIs with automated CI/CD pipeline reducing deployment time by 50%.

Graduate, Undergraduate and Personal Projects

AI-driven mobile app (Flutter) automating transaction tracking, budget monitoring, plot generation from user prompts, and receipt-based expense categorization with >90% accuracy. Features FastAPI backend and PostgreSQL database.

Android app that takes pictures of book pages, converts image to text using EfficientNetB3 model, and converts to speech

Raspberry Pi-based entry system with three levels of risk detection. Integrated RFID sensor, camera for mask detection (ResNet-50, 99% accuracy), MLX90614 temperature sensor, and MAX30100 SPO2 sensor. Automated workflow with Firebase DB.

Detected 2D human pose from multiple videos using mmpose and mmdet APIs. Applied Python multi-threading to process 4 videos concurrently and combined frames to generate final output video.

Used Drebin and MalwareGenome datasets, conducted EDA, feature selection with RFE, and tested multiple tree-based ensemble algorithms. LightGBM achieved 98% accuracy. Created script to extract permissions from APKs for model testing.

Multi-class classification problem using LightGBM and XGBoost. Conducted EDA, feature engineering, handled imbalanced classes, and used model stacking for improved performance.

3D animation project using Blender software, creating animated content inspired by the popular "Among Us" game with custom character designs and scenes.

System Analysis, Design & Development course project featuring emergency help, restricted internet access, video calls, messaging, file sharing, and discussion sessions with mental health experts.

Conceptual design project for national disaster management system, focusing on system architecture and emergency response coordination frameworks.

Comprehensive Java tool implementing lossless compression (RLE, DPCM, predictive encoding, entropy encoding, Huffman encoding), lossy compression (DCT, wavelet transform), and steganography watermarking for image encoding.

Miscellaneous

• Graduate Teaching Assistant, CSCE 1030: Computer Science I, Fall 2024
• Graduate Teaching Assistant, CSCE 5214: Software Development for AI, Spring 2025
• Graduate Teaching Assistant, CSCE 5380: Data Mining, Summer 2025

• Runner-up in IEEE Video and Image Processing (VIP) Cup 2020 Link
• First Place on Live Leaderboard, Diabetic Foot Ulcer Challenge 2021 Link

• Deep Learning Specialization by DeepLearing.AI certificate
• The Data Science Course 2021: Complete Data Science Bootcamp certificate
• Taming Big Data with Apache Spark and Python - Hands On! certificate

Template stolen from Jon Barron's Site.