Tech
Technical projects, some of them running services.
CrocoSign
We translate Polish Sign Language (PJM) videos into text in a spoken language. Our engineering thesis at PJATK: a dataset of 29,932 video-sentence pairs from 150 Deaf signers, and three gloss-free models compared on it. A research prototype, not a replacement for an interpreter.
MediMatch
Some drugs interact with each other, and some clash with a patient's allergies. As team 8 we built MediMatch, a checker on the Gemini API for people on long-term medication that flags both. It won first place at the Google Student AI Hackathon in Warsaw in June 2025.
Acoustic fields in 3D
Intense ultrasound in gas can act as an optical element for high-power lasers without being damaged by them, but only if the sound field can be reconstructed. During a summer studentship in the SonoPhotonics group at DESY in Hamburg, I co-developed the step of a PyTorch reconstruction framework, building on Schrödel et al. (Nature Photonics, 2024), from 2D to full 3D. Moving the simulation from NumPy on the CPU to PyTorch on the GPU, at 128³ × 512 spatial-temporal resolution, cut the large runs from hours to minutes.
Frank Tank
A toy tank, stripped of its electronics and rebuilt to follow a person. A Raspberry Pi 5 with a Hailo accelerator runs YOLO on the camera feed and steers towards whoever it sees; an Arduino drives the two motors; after the wake word "ok", Polish voice commands start and stop it and change its speed. YOLO on the Pi's processor alone lagged, so we moved detection to the accelerator. A course project at PJATK, built with Rostyslav Kovalenko in May and June 2025.
LLM-based GUI agents
We introduce GUI agents, AI agents that use an app's screen the way a person does: how they perceive the interface, plan, act and remember, and how they are trained and tested. A six-minute read on DEV, written with Rostyslav Kovalenko in January 2025.
Contact
About any of the work, or something we could make.
[email protected]