Projects
Pruna AI
Applied Team
I was the first member to join the applied team, building Pruna's own models.
One of the main engineers behind Pruna's own model family: P-Image, P-Image-Edit, P-Image-Ideogram, P-Video, P-Video-Avatar, P-Video-2-Pro and P-Video-2-Pro-Reference, with over 100M runs on Replicate.
Turned MiniMax-H3 into an efficient image model, with over 600K downloads so far. Integrated into diffusers and ComfyUI.
Trained 5- and 8-step LoRAs for Qwen-Image-2.1, with over 500K downloads on Hugging Face.
Research Team
I joined Pruna as part of the research team, where I did research on proprietary caching algorithms for diffusion models, helped build Pruna's first image models, and worked on the open-source pruna package.
Researched and built Pruna's proprietary caching methods. They power the official endpoints of BRIA 3.1, BRIA 3.2, BRIA FIBO and FLUX.2 [flex].
Built Pruna's first image models, Wan 2.1 Image and Wan 2.2 Image, from the Wan video models.
Added caching and quantization methods to the open-source pruna repository, and made its test suite about 2.4× faster.
Talks
Munich NLP: how quantization, compilation, caching and distillation make models smaller and faster, with a hands-on demo.
Hugging Face Computer Vision Hangout, June 2025: speeding up diffusion models with Pruna and diffusers.
Research
NeurIPS 2025. Grew out of my master's thesis at TUM.
TMLR 2026. Brings diffusion caching to molecule generation.
Hackathons & Side Projects
Finalist at LauzHack 2023: clusters the complaints in thousands of phone reviews with LLMs and embeddings, shown in an interactive dashboard.
Makeathon 2025: an autonomous pipeline that turns hours-long Bundestag debates into 60-second shorts.
Predicts NFL game outcomes from Madden video-game player ratings.