Professional Experience

• Recent lead experience in a large (200k requests/day) GenAI project with RAG and MCP
• Solid experience mentoring junior data scientists across various roles
• Over 2 years of GenAI experience in engineering and data science roles
• Over 8 years of experience as a Data Scientist and Backend Engineer at companies including Samsung and CHECK24 (Germany's largest comparison portal with 15M customers)
• Master's in Theoretical Physics from the University of Warsaw

Personal Projects

• Spell
  

  Interactive demo of token‑level spellchecker I trained using Torch & 🤗 Transformers and a Streamlit app for dataset exploration

  ×

  

  Spell

  A token‑level spellchecker trained from scratch on Wikipedia edit data.

  Model & Training: PyTorch and Hugging Face Transformers. Trained on Wikipedia revision diffs to detect and correct spelling errors at the token level.

  Serving: FastAPI backend with on-the-fly conversion to ONNX format for low-latency “as you type” inference.

  Frontend: React + Vite interactive demo.

  Dataset Explorer: A Streamlit app for browsing and exploring the training dataset.

  Live: spell.janczechowski.com Source: github.com/przecze/spell
• stillwhaling.fyi
  

  Interactive map showing countries still hunting whales based on IWC.int data. Processed with Python pandas, visualized using D3.js

  ×

  

  stillwhaling.fyi

  An interactive map showing countries that still hunt whales.

  Data: Based on catch data from the International Whaling Commission. Processed with Python pandas.

  Visualization: D3.js interactive world map.

  Stack: React + Vite, no backend. Dockerized and deployed using an nginx-proxy stack.

  Inspiration

  • Kurzgesagt / Nightshift video on whaling

  Live: stillwhaling.fyi Source: github.com/przecze/stillwhaling
• ⚖️️ Unfairness ⚔️
  

  Prompting challenge where the player has to use jailbreaking, manipulation, or other techniques to maximize their score over 6 rounds of "ultimatum" game against AI

  ×

  

  Unfairness

  A prompting challenge where the player tries to manipulate an AI opponent across 6 rounds of the “ultimatum” game.

  Backend: FastAPI, making LLM calls via the OpenAI Python SDK through OpenRouter to Anthropic’s Claude Sonnet.

  Frontend: React.

  Database: MongoDB for storing game data.

  Inspiration

  • Gandalf by Lakera — the original AI prompt hacking challenge
  • Ultimatum game — the game theory concept behind the mechanics

  Live: unfairness.janczechowski.com Source: github.com/przecze/unfairness
• ShutUpGPT
  

  A game where the user has to trick ChatGPT into omitting mandatory safety instructions before providing the nuclear codes

  ×

  

  ShutUpGPT

  A game where the user has to trick ChatGPT into omitting the mandatory safety instructions before providing the nuclear codes.

  Backend: FastAPI, calling the OpenAI API directly with GPT-4o.

  Frontend: React.

  Leaderboard: A deliberately primitive local JSON file as a “database”.

  Inspiration

  • Gandalf by Lakera — the original AI prompt hacking challenge

  Live: shutupgpt.janczechowski.com Source: github.com/przecze/shutupgpt
• 500miles.fyi
  

  A joke website showing what walking 500 miles and 500 more would entail, personalized to the user's location

  ×

  

  500miles.fyi

  A joke website showing what walking 500 miles and 500 more would actually entail, personalized to the user’s location.

  Stack: Simple static page with minimal vanilla JS. No build step, no framework.

  CDN dependencies: Leaflet for maps, Bootstrap 5 for layout, Bootstrap Icons, and marked.js for Markdown rendering.

  Hosting: Dockerized and deployed using an nginx-proxy stack.

  Live: 500miles.fyi Source: github.com/przecze/500miles
• Siege!! WIP
  

  My attempt at a web reimplementation of 'Siege!!' - a freeware game from 2001 with pattern matching and tug-of-war castle defence

  ×

  

  Siege!!

  My attempt at a web reimplementation of Siege!! — a freeware game from 2001 by Beau Langston (Fallen Angel Industries) I fondly remember from my childhood.

  Two castles sit at opposite ends of a 1D battlefield. Players craft units by arranging blocks on a pattern-matching grid, then watch them march and fight in a tug-of-war.

  Engine: Phaser 3.

  Art: Tiny RPG Character Asset Pack by Zerie, processed with a custom Python asset pipeline.

  Still from the original 2001 game. As you can see, much prettier than my janky 2026 version.

  Live: siege.janczechowski.com Source: github.com/przecze/siege
• PL2023
  

  Various (mostly silly) analyses of Poland 2023 parliamentary election results

  ×

  

  PL2023

  Various (mostly silly) analyses of the Poland 2023 parliamentary election results.

  Stack: A Streamlit app with interactive data visualizations.

  Libraries: pandas for data processing, Plotly for charts.

  Data source: Official election results from PKW (Państwowa Komisja Wyborcza).

  Live: wybory.janczechowski.com/?lang=en Source: github.com/przecze/wybory
• The Inndex
  

  Experimental search engine for The Wandering Inn web serial — queries are hallucinated into text fragments via GPT‑3.5, then matched semantically against Volume 1

  ×

  

  The Inndex

  A prototype of an experimental search approach for The Wandering Inn web serial (Volume 1).

  The idea: instead of searching directly, the user’s query is first fed to GPT-3.5-turbo which hallucinates a plausible text fragment. That hallucinated fragment is then used for semantic similarity matching against the actual text — often finding relevant passages that a literal keyword search would miss.

  Frontend: Streamlit.

  Hallucination step: OpenAI GPT-3.5-turbo.

  Embeddings: Hugging Face all-MiniLM-L6-v2 via sentence-transformers.

  Matching: Pure np.dot over pre-computed embeddings — no vector database.

  Data processing: Text chunking and preprocessing with pandas.

  Live: inndex.janczechowski.com Source: github.com/przecze/inndex