Projects

Chess Document AI pawnparse.com

Developing Optical Character Recognition for chess tournaments.

• Fine-tuning Vision and LLM Transformers for handwritten chess notation
• Developing text recognition for chess notation
• Data collection, Processing, Labeling and Design

Deep Stereo RGB-only Dense 6D Object Pose Estimation (Master Thesis)

Paper | Repository

Researched RGB-only 6D object pose estimation.

• Correctly identified which method in the literature will become state-of-the-art with further training and extended the method to a dual camera setup
• Created new real-world data with Kinect cameras, using camera calibration, and 50k synthetic samples with blender rendering and physics simulation on a Slurm cluster and generated 500k+ labels on a Slurm cluster with Python, Blender, Numpy and Numba
• Trained PyTorch model with 8 loss functions on a A100 GPU cluster with up to 8 GPUs per model with Docker containers inside Slurm jobs, optimized Hyperparameters
• Added new modalities to the model by combining it with depth regression models and tested various forms of image fusing neural network architectures for dual cameras

Deep Unsupervised Denoising For Interleaved X-Ray Channels

Paper | Repository

Researched deep unsupervised denoising for novel x-ray sensors. Successfully trained unsupervised denoiser in PyTorch, which required understanding of the specific statistics of x-ray sensors for correct loss function with Numpy and SciPy

• Trained deep neural nets for denoising in PyTorch
• Adaption of SOTA methods to the specific statistics of X-Ray images required understanding of the physics of the system
• Loss function had to fit the statistics of the noise generation process

Predictive Maintenance Internship

Worked on improving a deep predictive maintenance product.

• Promoted deep anomaly attribution for predictive maintenance on my initiative. Designed and developed a prototype using Autoencoders, Keras and Deeplift, which succeeded at identifying timing and sensor signals correlated with machine failure in real-world industrial IoT data, saving daily work for the 10-person customer service team
• Researched, tested and presented time series databases, prompting the team to switch to InfluxDB
• Developed data processing tooling with Pandas, SciPy and Bokeh

Deep Reinforcement Learning Agents For Pommerman

• Top three agent in student competition
• Designed and trained with PyTorch, Docker

Multi-Object Tracking

• Tracked 3D poses of objects in C++
• Filtering with a particle filter over the probability distribution

Genetic Algorithm for Build Order Optimization in StarCraft II (Bachelor Thesis)

Repository

• wrote a genetic algorithm in Java to optimize game-playing
• parallel evaluation of build orders with a C++ game bot

Solar plus Battery EV charging park energy and cash flow simulator

Repository

• Simulates a solar EV charging park in python with probability distributions
• Reports cash flows and balance sheet

Electric Vehicle Comparison Website eautoinfo.com

Repository

Created eautoinfo.com: compare most important aspects of EVs

• Frontend in React, Backend with REST and SQL
• Hosted with Microservices