* Engine Development* 3D Scene reconstruction

I worked on several projects during my time at DFKI, with my primary focus on Sign Language Animation and Production. Simultaneously, I had the opportunity to contribute to the in-house Large Language Model (ChatGPT) application, incorporating built-in RAG (Retrieval-based augmentation). These projects served as a gateway to acquiring skills in technologies such as Blender, GPT-4, LLMs, Three.js, and Vue.js.In parallel, I delved into various topics on diffusion, experimenting with diffusion models. This experience provided hands-on exposure to all current generative AI-based models. Beyond Sign Language, I explored emotion modeling and behavioral agents.Summary of Key Achievements:• Achieved up to 50+% precise solution over current approaches pioneering a trajectory matching system for sign language animations using L-BFGS.• Realized a 100% efficient animation process by engineering and optimizing an animation composer leveraging Inverse Kinematics, pushing the boundaries of realistic animations in the sign language field.• Interpreted and analyzed nuanced motion capture data for sign language, creating a pioneering benchmark that emphasizes clarity and precision in animations.• Architected the first-ever solution using the concepts of inflection parameters for sign language.

One of the most impactful work experiences was my time during this internship. I am immensely grateful to the entire team at NVIDIA for providing me with this fortunate opportunity. Throughout my tenure, I witnessed the scale and magnitude of AI research, thanks to the seamless experience that facilitated laser-focused exploration.What began as a small project culminated in a publication at a major computer graphics conference. This achievement was made possible through the guidance and support of my mentor, Jonathan Grangskog, and my manager, Timo Roman. Before my stint at NVIDIA, I had limited exposure to running machine learning research on these large systems. My internship served as an invaluable learning ground, allowing me to delve into computing on such expansive systems.Summary of Key Achievements:• Accomplished 500% improvement -- 2 hours to 20 minutes -- in training speeds by optimizing the deep learning pipeline through effective distributed training methods.• Seamlessly blended advanced neural network architectures like ResNet50 with NNFM and color loss refining style transfer results.• Integrated CLIP-based ResNet50 - the foundation of modern image retrieval and diffusion methods - with NNFM and color loss for style transfer.• Authored a technical paper acknowledged at Eurographics 2023 and held an oral presentation of about 10 minutes at the conference.