Computer vision · Wrocław, Poland

Computer vision, from dataset generation to the device it runs on.

Third-year Applied Computer Science student at Wrocław University of Science and Technology. Medical image segmentation, object detection and visual perception for robotics — trained, deployed and measured on real hardware, not only in simulation.

Get in touch GitHub LinkedIn

Selected work

DICOM to PACS inference pipeline

End-to-end clinical-style inference path: a study sent to a PACS server automatically triggers inference and returns the result as DICOM-SEG and DICOM-SR objects rendered in the OHIF viewer. Asynchronous queue with model versioning and a Grafana dashboard for latency and queue depth.

Orthanc · OHIF · Kubernetes · Grafana

Medical imaging

Breast tumour segmentation — 3D U-Net

Primary tumour segmentation on contrast-enhanced breast MRI against the MAMA-MIA benchmark — 1506 multicenter cases with expert segmentations, sourced from TCIA. Full MONAI pipeline: phase selection and resampling, intensity normalisation, patch sampling, augmentation, and a training loop evaluated with Dice and IoU on a held-out split.

Python · MONAI · PyTorch

Deep learning

Visual perception for an autonomous underwater vehicle

Object recognition for an AUV deployed on NVIDIA Jetson, with optimisation of camera image transmission. SLAM research covering mapping, pose estimation and resource usage on both a simulator and a physical ROV model.

ROS2 · NVIDIA Jetson · OpenCV

Robotics
2024 — present

NLP inference platform and LA Crime Data Warehouse

A multi-model text analysis system serving four NLP tasks as independent inference services, with shared tokenisation and queue-based batching. Separately, a Kimball star-schema warehouse over more than a million LAPD incidents, with idempotent SSIS ETL and an SSAS cube feeding a live Power BI report.

Java · Apache OpenNLP · OPUS-MT · SQL Server · SSIS · SSAS · Power BI

Further work

Experience

StettinerJul — Sep 2026

Computer Vision Intern

Python, OpenCV, NumPy, AWS SageMaker Ground Truth

Built a synthetic data generation and augmentation pipeline in OpenCV and NumPy with automatic bounding-box propagation for a 194-class automotive dashboard indicator detector — labels are produced with no manual annotation.
Expanded a 191-image base set roughly sixfold while keeping box geometry correct through every transformation.
Integrated the pipeline with AWS SageMaker Ground Truth — conversion between YOLO format and the output manifest under a unified class schema — and covered the geometric transformations with property-based tests ready for CI.

Technical stack

Programming: Python, C++, C, Java, Kotlin, Scala, SQL (PostgreSQL, Oracle)
Machine learning: PyTorch, TensorFlow, TensorFlow Lite, Scikit-Learn, MONAI, OpenCV, YOLO, 3D U-Net
Data: NumPy, Pandas, DICOM / DICOM-SEG, Power BI, MDX, synthetic data and augmentation, SageMaker Ground Truth
Infrastructure: Docker, Kubernetes, AWS (SageMaker), Grafana, Git, GitHub Actions CI/CD, Linux
Frameworks: ROS2, Django, React
Languages: Polish (native), English (C2, CPE), German (B2), Spanish (A2)