ML Engineer based in Olsztyn, Poland, working remotely (open to relocating to Wrocław or Łódź). I train models, check that their evaluation holds up, and build the backend that serves them. Most of my work is in medical data and speech.

• ML Engineer at EENGINE Software House. Speech recognition for call-center audio: fine-tuned Whisper for Polish (WER from 16% to 12%) and NeMo FastConformer for Romanian (WER from 30% to 17%), plus LLM-based extraction from call transcripts.
• Backend developer at Oncosort (finalist of Our Future Founders). Spring Cloud microservices for a medical data platform.
• MSc Computer Engineering (2026) and BEng Biomedical Engineering, medical informatics (2025), Wrocław University of Science and Technology.

Machine learning

• medical-ood-detection. My master's thesis: label-free out-of-distribution detection from K-Means centroid distances on frozen ViT and ResNet-50 embeddings. It covers 80 configurations, each with 5-fold cross-validation and Friedman–Nemenyi tests. AUROC is 0.999 on far-OOD and 0.81 on near-OOD, and the CNN beats the ViT on near-OOD.
• thyroid-ultrasound-vit. My engineering thesis, redone after I found data leakage in the original evaluation. A pretrained ViT-B/16 reaches 0.75 AUROC per patient on 157 held-out patients. It also includes shortcut tests, Grad-CAM and a FastAPI + React demo.
• pcb-yolo. PCB defect detection with a Canny edge map fused into the input channels of YOLO11n, which raises mAP@0.5:0.95 on DeepPCB from 0.742 to 0.765. It is written up as an IEEE-format paper.
• ecg-arrhythmia-classification. Heartbeat classification on MIT-BIH into the AAMI classes. The same CNN-LSTM scores 0.93 macro-F1 when beats are split at random, as in my first version, and 0.47 on unseen patients.

Software engineering

• cvrp-metaheuristics. A genetic algorithm, simulated annealing and baselines for vehicle routing, in plain Java. The genetic algorithm gets within 4.7% of the optimum on the Augerat A-set.
• library-management-system. A library web app with multilingual semantic search. A Spring Boot REST API (JWT, Liquibase, 26 Testcontainers tests) works with a small Python service that embeds book descriptions with sentence-transformers. Search puts the right book in the top 5 for 90% of test queries, against 38% for keyword search, and Polish queries find books described in English.

Machine learning

Backend

Frontend