# Mateusz Śliwa

**Data engineer · AI engineer · Python, machine learning, LLMs**

Last updated: 2026-10-13

- Email: [sliwa.mateo@gmail.com](mailto:sliwa.mateo@gmail.com)
- GitHub: [github.com/MSliwa01](https://github.com/MSliwa01)
- Web: [ecosort.pl](https://ecosort.pl)
- Location: Zielona Góra, Poland · remote or hybrid · open to relocation · available now
- Website: [mateuszsliwa.dev](https://mateuszsliwa.dev/)
- App Store: [EcoSort](https://apps.apple.com/pl/app/ecosort/id6785674008), [TrueSky](https://apps.apple.com/pl/app/truesky-weather/id6816346955)

I have an MSc in data engineering and a BA in economics. I built **EcoSort**, an iPhone app on the App Store, and the LLM pipeline behind it, which would rather hold a date back than publish a wrong one. I work with AI coding agents daily, build tools for them and test LLMs before I trust them.

## Projects

### EcoSort, creator and sole developer (since 2026)
_Waste collection dates, reminders and widgets for each address · 6 municipalities_
Demo: [mateuszsliwa.dev/ecosort](https://mateuszsliwa.dev/ecosort)
[ecosort.pl](https://ecosort.pl) [apps.apple.com/pl/app/ecosort](https://apps.apple.com/pl/app/ecosort/id6785674008)

- Built a pipeline that reads schedules one month or table section per LLM call: two reads with different prompts, a third on disagreement, code checks for holidays, Sundays and rhythm gaps, and my own YOLO detector as a fallback.
- On six hard scans (2,190 cells) the best of three prompt versions got **24.5% of cells wrong**. Over three full runs the pipeline published 82.2% of cells with **no wrong date** and held the other 17.8% back instead of guessing. The ground truth is 966 manual decisions, and the pipeline was tuned on these documents.
- Wrote the FastAPI and PostgreSQL backend with TERYT address matching, a React admin panel, the SwiftUI app, an unreleased Jetpack Compose Android app, a Gemini assistant with PII scrubbing and Docker Compose on a VPS.

### Win probability model for Counter-Strike 2, master's thesis (2025–2026)
_Python, pandas, Parquet, XGBoost, LightGBM, CatBoost, SHAP · FastAPI and React app in Docker_
Demo: [mateuszsliwa.dev/thesis](https://mateuszsliwa.dev/thesis)

- Parsed recordings of 9,196 professional maps into 16.8 million observations from 187,361 rounds and compared 9 models with match-grouped cross-validation, bootstrap tests and ablations. In quarterly walk-forward tests XGBoost reached **ROC-AUC 0.81–0.82**, with calibration error (ECE) 0.009.

### Skin lesion classifier, machine learning course project (2026)
_PyTorch, Optuna, MLflow, Grad-CAM, Streamlit, Docker_
Demo: [mateuszsliwa.dev/skin](https://mateuszsliwa.dev/skin)

- Trained a custom CNN and a fine-tuned EfficientNet-B0 on the HAM10000 images, with focal loss, class-balanced sampling, MixUp and test-time augmentation. EfficientNet reached **AUC 0.93** and 0.95 recall on malignant lesions.

### wwxd ("What would X do?"), open source (2026)
_Python, Typer, yt-dlp, faster-whisper, rapidfuzz · a Claude Code plugin_
Demo: [mateuszsliwa.dev/wwxd](https://mateuszsliwa.dev/wwxd)
[github.com/MSliwa01/wwxd](https://github.com/MSliwa01/wwxd)

- Turns one person's talks, podcasts and essays into a wiki of their positions; an agent answers "what would X do?" with timestamped quotes, and a script found 158 of 160 of them word for word. A blind judge from another lab (Google) preferred it to plain Claude on **34 of 40** questions. It doesn't predict opinions not yet voiced.

### storeshots and macsim, open source tools for AI coding agents (2026)
_App Store screenshots · an iOS Simulator for agents on Linux_
Demo: [mateuszsliwa.dev/storeshots](https://mateuszsliwa.dev/storeshots) [mateuszsliwa.dev/macsim](https://mateuszsliwa.dev/macsim)
[github.com/MSliwa01/storeshots](https://github.com/MSliwa01/storeshots) [github.com/MSliwa01/macsim](https://github.com/MSliwa01/macsim)


## Skills

- **Languages:** Python, SQL, TypeScript, JavaScript, Swift, Kotlin, Bash
- **Data and ML:** pandas, NumPy, Parquet, scikit-learn, XGBoost, LightGBM, CatBoost, PyTorch, SHAP, Optuna, MLflow, OpenCV, YOLO, OCR, Selenium
- **LLM and AI:** LLM evaluation, prompt design, Anthropic API, Gemini API, Whisper, Claude Code, Codex, Cursor
- **Backend and infra:** FastAPI, PostgreSQL, SQLAlchemy, Alembic, REST APIs, Docker Compose, Linux, Git
- **Mobile and web:** SwiftUI, Jetpack Compose, React

## Education

### MSc in data engineering, University of Zielona Góra (2025–2026)
_Specialisation in data mining systems · thesis on round win probability in Counter-Strike 2, defended 2026_

- Courses included machine learning with scikit-learn and TensorFlow, big data with Hadoop, Hive and Kafka, data warehouses with ETL, OLAP and SQL, Power BI, cloud load balancing, Docker and CRISP-DM.

### BA in economics, University of Zielona Góra (2021–2024)

- Courses included econometrics in Gretl, statistics in Python, financial analysis, accounting and Excel with VBA.

## Languages and interests

- **Languages:** Polish, native · English, C1 with the Cambridge C1 Advanced certificate from 2025 · German, A2
- **Interests:** Chess, 1850 on chess.com · football · cycling · esports data, which is where my thesis topic came from