Oskar Jarczyk oskar-j

Hey there, I'm Oskar 👋

🐍 Pythonista turning data into products for 12+ years in AI - 10 of them shipping commercial ML
📊 Data scientist at heart: recommender systems, NLP, fraud detection, forecasting, real-time big data
🤖 AI engineer nowadays: agentic systems, MCP, RAG and LLM evals - from prototype to production
🎓 PhD in social informatics: I studied how open-source teams on GitHub really work (200+ citations)
🌍 Based in Warsaw 🇵🇱, shipping for teams in 🇨🇭 🇬🇧 🇩🇪 🇺🇸 🇰🇷, among others

🧩 MCP servers & tools - plugging LLM agents into external data providers, voice agents and chat systems (down to low-level FastMCP customization)
🔎 RAG & natural language → SQL - multi-modal retrievers, smarter chunking, on-prem deployments
🧪 LLM evals & observability - quality-evaluation pipelines and metrics-driven tuning with Langfuse
☁️ Multi-cloud GenAI - AWS Bedrock & AgentCore, Azure AI Foundry, GCP Vertex AI, shipped with Terraform, Kubernetes/Helm and CI/CD

A taste of what I've been building lately - a good place to start. There's more in my repositories.

Project	Why it's cool
💸 visualize-my-expenses	Turn a month of budget rows into a Sankey diagram you can share as a PNG: income → budget → categories → what you actually bought. Reads CSV, JSON, Excel and bank exports (OFX, QIF, ISO 20022 camt), ~50 currencies, fully offline. `pip install vme-py`
🎯 thresher	Your classifier's 0.5 cut-off is almost never optimal. Thresher finds the threshold that maximizes accuracy - 7 search algorithms, pandas / Spark / Ray, library + CLI. `pip install thresher-py`
🐈 fastcat	Lightning-fast local lookups of Wikipedia category relations (DBpedia + Redis), 7+ languages, ~20x faster loading. `pip install fastcat`
🖼️ walsh-hadamard-transform	A complete, bit-exact image codec built on a transform from the 1920s - pure Python + NumPy, 7 file formats, 700+ tests. `pip install walsh`

💪 Can-do attitude - I've helped both fast-moving startups and large corporations turn ideas into working products
🎯 Detail-driven - I wrote a whole library because 0.5 is almost never the right threshold
➕ Curious by default - I built an image codec where additions and subtractions are all you need 😄
🕸️ I've spent years modelling how developers collaborate - so yes, I have opinions about your PR review process
💬 Ask me about: agents, evals, RAG done right, and what makes open-source teams thrive

📫 Let's talk: datageek.pl · LinkedIn