Silver medal · top 1.5%

Biohub: Cell Tracking During Development

Detecting and tracking zebrafish cells in 3D over time during early embryonic development. My solution is an ensemble of three small 3D U-Nets with a transformer linker on top. Most of the gain came from handling the sparse labels: only about 1 in 28 cells is annotated, so out of the box the model learns that most real nuclei are background.

59th of 3,947 teams

• Solution
• Certificate

Morphottention

Attention for PyTorch with the dot product replaced by a morphological operation. The forward and backward passes are Flash-attention-style CUDA kernels, so the full attention matrix is never stored. It works with autograd, and a standard ViT can use it without other changes.

• Source
• PyPI

serron

Differentiable erosion, dilation and the operators built from them, plus learnable morphological layers, for PyTorch. With flat structuring elements the CUDA kernels run a separable van Herk–Gil–Werman scan, so forward and backward cost O(1) per pixel at any window size. Eroding an 8×3×512×512 batch with a 127×127 window takes 0.25 ms on an H100. Plain PyTorch needs 81 ms, CuPy 2.0 ms.

• Source
• PyPI
• Benchmarks

Lesion Segmentation in Dermoscopic Images Using FCN with Mathematical Morphology Enhancements

Plain FCNs segment skin lesions with fragmented masks and lose thin structures. We added learnable morphology to fix that: differentiable morphological layers inside the network, a boundary-aware loss, and morphological post-processing. On ISIC2018-T1 a stripped-down FCN with these changes reaches a Dice of 0.9070. That beats the original challenge submissions, later FCN methods and several larger U-Net and Transformer models, at a fraction of their compute. I led the team of five and wrote the morphological layers, the loss, preprocessing and evaluation.

Paper awaiting publication · code on acceptance

kx-plot

Matplotlib plots in one line for notebooks. A short spec string like “scatter-grid-leg” sets the plot kind, theme and options, and an unknown token raises an error. It has 15 plot kinds, 7 themes and colour-blind-safe palettes. The only dependencies are matplotlib, numpy and pandas, so it also works offline in Kaggle competition notebooks.

• Source
• PyPI
• Kaggle dataset

bt-dualboot

Syncs Bluetooth pairing keys between Linux and Windows so paired devices survive a dual-boot switch.

• Source

IssueHub

A JetBrains IDE plugin for working on GitHub and GitLab issues from inside the editor.

• Source
• Marketplace

BullBear

A paper-trading platform for stocks and ETFs, built by a team of six for FER’s Software Engineering course. Users trade on live market data, follow their portfolio, compete on return leaderboards and share “Mini-Funds”. I led the team and worked across the stack. CI/CD, deployment and the database were mine, as was a caching layer that batches market-data requests in cycles to stay under the data provider’s rate limits.

• Source

University of Zagreb, Faculty of Electrical Engineering and Computing (FER)

MSc, Computer Science (in progress).

University of Zagreb, Faculty of Electrical Engineering and Computing (FER)

BSc, Computer Science.

Thesis: Memory efficiency of self-attention mechanisms in vision transformer architectures.

XV Gymnasium (MIOC), Zagreb

High-school diploma, Mathematics & Computer Science.

Linux & Open Source Club · competitive-programming team (InfoKup, HONI, ACSL).