📅 IEEE Big Data 2024
DOI: 10.1109/BigData62323.2024.10825799

• 🚀 Novel graph clustering method for training differentially private GNNs on disjoint subgraphs
• 🔒 Enhances privacy and scalability without sacrificing accuracy

Developed a fully client-side ML system that predicts 3D object rotation from images and re-orients objects upright using AI in real time.

🔗 View Repository →

• ⚙️ ML CI/CD pipeline using a remote GPU server (SSH) for training.
• 🧪 3D pipeline for infinite synthetic dataset generation.
• 🔄 Consumed synthetic data to train orientation model.
• 🧠 Trained a Reinforcement policy to orientate models upright.
• 🌐 WebGPU deployment for models with client-side inference.
• 🚀 Demonstrated end-to-end ML engineering skills spanning synthetic data generation, deep learning, reinforcement learning, and distributed AI systems.

Designed and implemented a custom Blender + Python pipeline to generate high-quality, multi-angle synthetic datasets of 3D game character models with transparent backgrounds, tailored for computer vision and object detection tasks.

🔗 View Repository →

• 👁️ **>85% real-time accuracy** detecting in-game characters.
• 🧰 Automated Blender/Python scripting to facilitate infinite dataset creation.
• 🚀 Expertise in **synthetic data generation, computer vision workflows, and practical AI model training** in object detection.

A stylized Blender-based rendering system that scrapes, processes, and animates vintage gaming magazines into rotating, transparent 3D models.

🔒 Core assets (models, templates) are private to preserve originality.

• 🗂️ Scrapes 90s–2000s **PDF magazines** from web archives and converts pages into texture images
• 📘 Maps pages onto a **custom 3D magazine mesh**
• 🧠 Includes an interactive **page selection interface**
• 🛠️ Streamlines page swapping and batch rendering into a fully automated pipeline, simultaniously generating seamless, looping GIFs and individually sized videos for Instagram and TikTok.

Applied advanced Differential Privacy (DP) techniques to train neural networks that securely protect sensitive data, preventing any potential dataset reconstruction or leakage — crucial for privacy-critical industries such as healthcare and finance.

🔗 View Project →

• 🔒 Engineered DP-trained models that provide mathematically backed **privacy guarantees**.
• 🧠 Executed end-to-end development and evaluation of neural networks on the MNIST dataset, optimizing the trade-off between **model accuracy and privacy**.
• 🛠 Applied noise addition, gradient clipping, and privacy accounting methods, showcasing proficiency in **privacy-preserving machine learning frameworks**.
• 📊 Delivered comprehensive performance analyses demonstrating how to maintain predictive power while ensuring **robust data privacy**.