Full Stack Engineer · One of Switzerland's largest health insurers
Reduced production incidents by ~75% in the broker support and management platform by redesigning its business-process workflows into reusable, composable subprocesses with tightened input-data contracts.
Automated the team's release process by migrating backend services' CI/CD pipelines from on-prem to cloud, replacing manual deployment steps with standardised builds and freeing engineers from release toil.
Eliminated uncontrolled production database changes by prototyping a version-controlled, declarative schema-migration workflow (Liquibase) — replacing ad-hoc, hand-run SQL with reviewable, repeatable, rollback-able migrations.
Improved team-wide code quality and engineering practices — design patterns, testability, and consistent style — by mentoring junior developers and reinforcing standards through code reviews and pair sessions.
Jul 2019 – Sep 2020
Zürich, CH
helmedica AG
Web Developer · Medical practice-management software
Automated incoming lab-result entry by building a matching component that auto-attached exact matches directly to patients and routed ambiguous ones to a staff review queue, replacing what was previously fully manual data entry.
Projects
2025 – 2026
Drug Review Condition AI Classifier (DistilBERT)
AI Engineering · 4-person team
Built a reproducible ML pipeline (cleaning → masking → baseline → fine-tuning → evaluation → auto-rendered PDF report) classifying free-text drug reviews into 10 conditions on a 98K class-stratified, drug-name-masked dataset derived from 215K raw reviews.
Established a TF-IDF and logistic regression baseline on masked data (0.879 accuracy, 0.836 macro-F1, 0.981 top-3 accuracy across 10 classes) as the floor the Transformer must clear; quantified drug-name label leakage via masking ablation (+2.0 / +2.7 points accuracy and macro-F1 from un-masking).
Fine-tuned DistilBERT with class-weighted cross-entropy and early stopping on macro-F1 (handling 8.6:1 class imbalance) to reach 0.917 accuracy and 0.880 macro-F1 on the masked test set — a +3.8 / +4.4 point improvement over the linear baseline.
2015 – 2017
Sunrise TV
Netstream
Won Netstream's internal hackathon: pitched and built a live-match sports-highlights replay feature for Sunrise TV, the consumer IPTV set-top box, built with one teammate.
Graph Adjacency-Matrix Generalisation
University Algorithms & Data Structures
Earned the course's top grade (5/5) and an exam exemption granted to only 3 of ~300 students by devising an original generalisation of the graph adjacency-matrix representation.
2019
From Nand to Tetris
Building a Modern Computer From First Principles
Built the Hack hardware platform in HDL: from logic gates and ALU to CPU, memory hierarchy, and full computer chip.
Implemented the toolchain in Python: a two-pass assembler with symbol resolution, and a VM translator supporting stack arithmetic, branching, functions, and returns.
Wrote Tetris in Jack on top of the platform: seven tetromino shapes encoded as bitmaps, rotation, collision detection, and line clearing.
2020
Contra Corona
CodeVsCOVID19 Hackathon
Built a multiplayer team game to raise awareness of the importance of social distancing during the pandemic, with mechanics that rewarded players for staying home.
Education
Oct 2025 – Expected 2027
Warsaw, PL
Vistula University
Computer Science · AI Engineering
Part-time, semester 4 of 7 · GPA 4.16 / 5.00.
Skills
Languages
JavaJava
JavaScriptJavaScript
TypeScriptTypeScript
PythonPython
Frameworks
Spring BootSpring Boot
AngularAngular
ML / Data Science
PyTorchPyTorch
Hugging Face TransformersHugging Face Transformers