Portrait of M.Sc. Thorben Voss Drop a Message

GPU-Powered 3D & AI Web Solutions Leading the Next Industrial Revolution

M.Sc. Thorben Voss | GPU & 3D Web Technologies | AI & Computer Vision

The next industrial revolution - powered by Artificial Intelligence - is already here. My mission is to strengthen businesses with cutting-edge 3D and AI solutions, preferably on the web and on-device. This ensures seamless accessibility while enhancing productivity, innovation, and competitiveness in complex domains.

My fascination with ML began in 2016 at the University of Passau with a seminar and bachelor's thesis on acoustic event detection using RNNs. Leveraging this foundation, I completed my master's at the Karlsruhe Institute of Technology (KIT), an excellence university in the TU9 alliance and a Helmholtz Association member, with a thesis on texture synthesis using CNN-based GANs in collaboration with Fraunhofer. All my theses, papers, and internships received the highest grade.

During my studies, I served as a research assistant at Fraunhofer IOSB and FZI, specializing in 3D visualization, prototyping, and automated avatar modeling for VR environments within the automotive and robotics sectors. Following that, I gained two years of professional experience as a software engineer in GIS and BI, where I developed TS web components to streamline governmental reporting processes.

My passion lies in GPU-powered, AI-driven immersive web experiences. Driven by this enthusiasm, I developed my own game engine utilizing WebGL, showcased at VOSS.ART. Through Progressive Web Apps (PWAs), I offer my projects to everyone, everywhere, on every platform. So if you're working on projects at the intersection of computer graphics and AI for the web, let's connect and innovate together! 🚀

Academics

Synthesis of Textures for Automotive Simulations Using Deep Learning

The pix2pbr demo is one of the outcomes of my master's thesis [1], which focused on automatically generating the pysically based rendering textures from any given RGB image. These textures encode surface properties in their channels - such as diffuse color, height variations, normals, and can even indicate whether a specific area is an electrical conductor. This refined information is then utilized in the rasterization process of interactive computer graphics to simulate physically based microfacet models in real time. The Fraunhofer IOSB, which supervised this work, can integrate the networks into its OCTANE simulation platform to enhance runtime-generated RGB textures.

To generate paired training data, I mapped freely available texture datasets onto virtual surfaces in Blender and rendered them using the Cycles path tracer, which simulates physically plausible light interactions. My neural networks are then tasked with inferring the textures required for real-time applications from these rendered images - or even directly from real-world photographs.

The project uncovered several challenging sub-problems, with conversion errors from basic CNN approaches shown in the heptagram. Some methods employed GAN architectures, where generated textures were assessed by a discriminator network. The project was completed within six months, awarded the highest grade, and achieved the following scientific and personal outcomes:

PyTorch deep learning framework logo TensorFlow machine learning framework logo ONNX open neural network exchange format logo Blender 3D modeling and animation software logo

Acoustic Event Detection Using Deep Neural Networks for Smart Homes

The aim of my bachelors's thesis [2] was to determine which neural network architecture is best suited for detecting acoustic events in a smart home. The thesis was conducted at the Chair of Complex and Intelligent Systems under the supervision of Prof. Björn Schuller and Dr. Maximilian Schmitt, where I already wrote a term paper on music recommendation systems [3]. To ensure practical relevance, I was the first student from the University of Passau to align the requirements with a company - the Austrian Loxone Electronics GmbH.

The networks were benchmarked against existing methods using cross-validation on the annual DCASE competition dataset, consisting of annotated audio files. The input spectrograms, computed using the software OpenSMILE from audEERING, represent time, frequency bands, and magnitudes from the discrete fast short-time Fourier transform. In the 3D background image, you see 1s of washing dishes.

A key challenge was handling inputs of varying temporal length, which ultimately made recurrent neural networks (RNNs) strong candidates. The approaches also had to robustly detect silence, which was unfortunately overrepresented in the data. Despite this, my methods kept pace with the state of the art, the thesis received top marks, and it yielded the following academic and personal results:

TensorFlow machine learning framework logo openSMILE logo - open-source Speech and Music Interpretation by Large-space Extraction audEERING logo - group behind openSMILE LOXONE logo

Experience

Software Engineer at Disy Infor­mations­systeme GmbH

Driven by my passion for bringing 3D content and neural networks to the web, I pursued professional experience at Disy Informationssysteme GmbH. In an award-winning environment, I worked on Cadenza, a leading software enabling government institutions to visualize and explore large datasets through interactive tables, charts, and maps for data-driven decision-making.

In international teams, we worked agilely with Kanban and Jira to implement customer requirements in TypeScript and Java, bringing dynamic, user-centric features to life on the web. My focus was on developing encapsulated frontend components within the Shadow DOM. These followed current data protection, security, accessibility, and web standards, integrating UX considerations while strictly meeting customer requirements on time. Experience them in the online map service of the State Institute for Environment BW.

With a dedicated monthly training day, the company fostered my growth through courses on modern TypeScript development and a specialized CSS course by Josh Comeau [4]. I thoroughly enjoyed my time there, both professionally and personally, even having the opportunity to take the stage as the male lead singer of the company band, ACdisys. Key learnings include:

Research Assistant at FZI

After being captivated by the Research Center for Information Technology (~FZI) through outstanding lectures on biologically inspired robotic systems - and earning top marks from Prof. Dr.-Ing. Rüdiger Dillmann - I was eager to contribute as a student research assistant.

My responsibilities included 3D modeling, visualization, and 3D printing of cutting-edge robotic systems, as well as creating virtual clones for collision detection and developing concepts for trade show presentations in the automotive sector.

I refined my FDM 3D printing skills on various Prusa printers, monitored prints with OctoPrint, and automated part removal using KUKA UR16 robotic arms. I also developed expertise in converting CAD models into polygon-based formats and designing integrated joints and snap-fit mechanisms, resulting in dozens of virtual and prototype industry assets.

Research Assistant at Fraunhofer

Since the Fraunhofer Institute for Optronics, System Technologies, and Image Exploitation (IOSB) offered exceptionally well-structured and engaging KIT lectures, I joined the institute to develop 3D models, VR applications, and, a few semesters later, my AI master's thesis.

During my work, I designed a digital strategy table in Blender and integrated it into our own Unity scene for exploring map sections in VR. For presentation purposes, I developed an automated pipeline using MakeHuman to generate a VR-ready avatar from a profile picture.

At Fraunhofer, I also learned Git and explored new technologies like the web-optimized 3D format .gltf and the Web Graphics Library WebGL to render 3D models in the browser using the GPU. My enthusiasm for these technologies ultimately led to the development of a .gltf importer and my own game engine - explore the resulting interactive projects on VOSS.ART!

Working Student at CashCard

At our family business, CashCard Automaten GmbH, we build cashless payment systems using contactless media, including POS systems, vending machines, and access solutions. Our software departement implements web apps, meal pre-ordering, and WinCash.NET. In collaboration with auvisus, we even support AI-based tray recognition via visionCheckout.

My responsibilities included creating the entire website, including special features such as the knowledge base section, consulting and implementing mobile-optimized online apps, and even 3D printing prototypes and custom parts for the new card top-up SKV10.

The role gave me valuable, sometimes involuntary insight into data protection and fiscal regulations like the DSFinV-K, along with changing hygiene standards during the pandemic. It also deepened my understanding of structural challenges in Germany's tech industry.

Founder of VOSS.ART

Web development has always been a creative outlet for me. I built dozens of websites and even sold some. At Fraunhofer, I learned that 3D models could be rendered in the browser using JS and WebGL, and I became captivated by the idea of bringing 3D to the web.

My first project was a racing game starring a plush frog - Froschidrive. It started as a Christmas gift for my girlfriend. In 2020, I turned it into a reusable game engine and got the sequel accredited as an internship at KIT. Play these and more games now on VOSS.ART.

Several semesters later, I decided to bring my master's thesis, also completed at Fraunhofer, and similar apps, to the web. Our texture synthesis networks can even be integrated into my game engine pipeline, enhancing lightweight RGB textures to full PBR quality during asset loading. Complex GPU-driven programs in the browser - what could be better? Stay tuned.

Beyond

Dive Beyond GPU-Accelerated Code

As you've likely noticed by now, I'm a technically skilled mind who tackles complex challenges passionately and independently - and solves them using cutting-edge technologies, even in uncharted waters. To wrap things up, a few personal traits:

My name has Scandinavian roots. It combines Thor, the god of thunder with a clear affinity for high voltage, and Ben, derived from Björn, meaning bear or chief. That's why the h in my name matters to me, so please don't forget it in your greeting.

I enjoy exploring nature on foot, from a mountain, a paraglider, or while windsurfing. This passion recently took me to Australia for a year, where I lived in a skyscraper, improved my English, worked on the JB Hi-Fi salesfloor in Brisbane, found opals in Lightning Ridge, and discovered that the best Australian food is actually Asian.

I am also a dedicated teammate in foosball, tennis, strategy games, competitive fencing at championship level, and I enjoy organizing group activities. As a trained baritone, I sing in choirs, bands, and at karaoke parties. I've been playing the violin for decades and have a deep appreciation for diverse musical genres.

Subtle details catch my eye, especially in everyday life. I enjoy building things from scratch, trying unconventional ideas, and staying open in controversial discussions. Sharing knowledge comes naturally, and I like making abstract concepts tangible through visual language or 3D models. Let's turn our visions into reality! 🌠