At DeepMark, I developed, tested, and evaluated deep-learning watermarking models for images and video using Python and PyTorch.
I profiled and optimized GPU inference and video-processing pipelines, reducing watermark embedding time by 38–56% across 360p–4K workloads. I also designed reproducible benchmarks to evaluate robustness against compression, resizing, cropping, rotation, codec changes, and upscaling.
At OneAI, I implemented ECG preprocessing and signal-analysis workflows, including wavelet filtering, R-peak detection, and signal-quality estimation. I also applied CNN-based classification to biomedical time-series data and evaluated preprocessing and model outputs.
In my projects, I trained and evaluated RawNet-style audio spoof-detection models with ASVspoof data, and implemented an audio retrieval pipeline for my B.Sc. thesis on music cover and remix detection. I'm pursuing M.Sc. studies in Electrical Engineering and Computing at the University of Belgrade — School of Electrical Engineering (ETF).

