Projects
My research and engineering projects are listed here, including, when available, associated live systems, publications, and technical details. Further details are available by exploring each project.
MiddelMen (TrustThread)
Architected and deployed a scalable social-commerce trust platform with a decoupled React/Vite SPA on Vercel Edge and a Node.js/Express REST API hosted on Heroku (GitHub Actions CI/CD). Designed normalized relational schemas using Neon Serverless PostgreSQL and Drizzle ORM to enforce multi-role RBAC, verified seller reviews, and dispute logging; integrated Google OAuth 2.0 and bcrypt authentication. Engineered a dynamic SSR crawler-detection middleware injecting Schema.org JSON-LD metadata for SEO, backed by Cloudflare R2 zero-egress media storage.
Multilingual Mental Health Chatbot (RAG & LLM)
Designed and deployed an end-to-end bilingual (English / Roman Urdu) conversational system integrating Ollama Gemma-3 (1B) with advanced Retrieval-Augmented Generation (RAG) vector embeddings. Built a React frontend, Flask backend on Modal serverless GPUs, and auxiliary Express APIs. Trained PyTorch emotion-classification models achieving 86% test accuracy (BERT) and ~90% accuracy (XLM-RoBERTa), storing interaction logs in MongoDB Atlas.
LadderSort: Adaptive Stable Sorting for Partially Ordered Data
Engineered a novel deterministic, comparison-based sorting algorithm in C++ that dynamically adapts to input order, formally proving algorithmic optimality. Achieved up to 1.45× wall-clock speedups over std::sort and TimSort on Apple M1 architectures handling datasets up to 100M elements.
GPU/CPU Parallel Object Detection Engine
Developed a low-level object detection system processing 24-bit BMP images via custom memory representations, grayscale transformations, and Sobel edge detection. Achieved dramatic execution speedups through algorithmic parallelization, hardware-level memory optimization, and multi-threaded CPU/GPU workload distribution utilizing CUDA C and OpenMP APIs.
Written Digit Detection AI
Built a TensorFlow neural network for MNIST digit classification with high accuracy; preprocessed data and implemented model saving, reloading, and evaluation workflows with an interactive canvas drawer.
Discuss50 Web Platform
Developed a Flask-based web application for the CS50 bootcamp enabling students to submit weekly reviews and responses on lectures and psets; implemented user authentication, password reset, and secure data handling with SQL injection prevention.
Dijkstra's Algorithm Visualizer
Developed a Java-based visualizer for Dijkstra's shortest-path algorithm with a Bootstrap interface, implementing graph traversal and interactive visual feedback for educational purposes.