A selection of robotics, AI, and data science projects — most course or research work isn't public on GitHub. Reach out and I'll happily share more.
Teaching a Unitree G1 humanoid to perform the Bavarian Schuhplattler, a dynamic full-body folk dance, through motion imitation and reinforcement learning. Deployed a full pipeline: motion capture, retargeting, RL policy training in simulation, sim-to-sim validation, and sim-to-real deployment. Now deployed on the real G1, this ongoing project is set to culminate in a scientific publication.
Developed 'SafeFlow' for Belimo's challenge at the START Hack flagship event: a software feature that uses existing Energy Valve sensors to detect catastrophic coolant leaks in data centers and automatically shut the valve within seconds. Required no new hardware, deployable via firmware update, protecting high-value AI infrastructure from water damage.
Built 'Chief AI', a multi-agent Discord orchestrator (OpenClaw + Discord + enterprise dashboard) for OpenAI's challenge track. The system scans channels and meeting transcripts, extracts action items, routes them to specialist agents, and triggers task creation, DM notifications, and email summaries. Out of 6,500 applications, the project earned the team one of 29 spots in Hack-Nation's venture lab, a 3-month startup incubator ending in an investor pitch.
Developed and trained a model to classify drones from acoustic signals, deploying a real-time inference pipeline on an NVIDIA RTX 5090 GPU for high-performance signal classification.
Developed a real-time application to classify and evaluate gym exercises using pose estimation and deep learning. Managed the codebase and built models to distinguish between multiple exercises and assess correctness. The project was deployed as a user-friendly web app and received the highest grade.
Fine-tuned transformer models like BERT, ALBERT, and DeBERTa for reading comprehension on the RACE dataset. Achieved strong accuracy using Google Colab and Hugging Face, despite limited resources.
Developed an end-to-end pipeline for integrating and managing biosignal and clinical data from multiple hospitals. Built a robust SQLite database, API, and user interface for clinicians and researchers. The project focused on data cleaning, interoperability, and usability, earning the best possible mark.
Authored a seminar paper comparing Diffusion Models and GANs for data generation, training, and output quality. Explored recent advances and practical applications in machine learning. The paper was awarded with the best possible grade.
Programmed a UR10e robot to autonomously draw SVG artwork on a whiteboard, validated in RoboDK and deployed on the real robot. Also implemented cleaning and trajectory replay routines for Franka robots using C++, ROS, and Docker. The project received the highest possible grade.
Developed a wearable smart camera device to help visually impaired people navigate buildings. Built the MVP in a two-week sprint, handling software, 3D printing, and device design. The project won the Start Award for best business potential.
Simulated and optimized a MISO PI controller to improve the path-following performance of an autonomous driving robot. Used MATLAB, Python, and ROS for simulation and software development. The project significantly reduced errors and crashes, earning the best possible grade.
Designed and built aerodynamic components for the xb023 Formula Student race car as part of the Aerodynamic Subteam. Contributed to both design and hands-on carbon fiber manufacturing, as well as sponsor outreach. Helped the team achieve top placements in international competitions.