Problem: Standard CNNs struggle to capture global context in noisy CT scans, leading to false negatives in kidney abnormality detection.

Solution: Designed TransConvNet, a hybrid architecture merging CNNs for local feature extraction with Transformers for global attention.

Impact: Achieved SOTA performance 99.9% and implemented Grad-CAM for clinical interpretability, allowing doctors to visualize decision boundaries.

Problem: "Black box" deep learning models lack transparency in critical classification tasks.

Solution: Developed a CNN-based classifier for the FER-2013 dataset integrated with Grad-CAM (Gradient-weighted Class Activation Mapping).

Impact: The system visualizes exactly which facial features (eyes, mouth curvature) drive specific emotion predictions, addressing model bias and reliability.

A production-grade CRM designed for high-throughput event lead capture. Integrated Google Gemini API for unstructured-to-structured data extraction (JSON) from business cards, replacing brittle OCR templates. Engineered a custom fuzzy-logic deduplication algorithm to handle data redundancy in real-time.

An end-to-end MLOps demonstration focusing on the deployment lifecycle of computer vision models. While the core model uses CNNs, the project highlights infrastructure: Docker containerization for reproducibility, Nginx/Gunicorn for load balancing, and a CI/CD workflow on AWS EC2 for continuous deployment.

A data visualization platform analyzing tourism trends from scraped datasets. Focuses on data cleaning pipelines and interactive frontend visualization components to derive actionable insights from raw review data.