I’m Oskar Breitfeld. As a kid I wanted to become some kind of crazy inventor; today I connect energy engineering with automation, AI, and machine learning. I didn’t study computer science - I learn by building. And I genuinely hate stupid repetitive tasks - automating them became the obvious career.
sim: a day of load
spotify feed warming up, or offline.
uptime appears when the api answers.
recent activity appears when the api answers.
The foundation: processes, materials, energy - and where software can improve complex systems.
Energy and processes, with growing focus on what data can predict and optimize.
Agentic AI workflows with LLMs, Claude, MCP servers, RAG - automation people actually use.
Forecasting for the grid: exactly the energy × ML intersection I want to work in.
A shared Spotify link hijacks the speaker and kills the queue. Every time.
Built: forward the song via WhatsApp and the bot appends it cleanly. No interruptions, no hijacked speaker - it has survived every party since.
see the LinkedIn post about the flow ↗
Spot instances are cheap - until the wrong choice interrupts a workload or quietly costs more.
Built: a cloud optimizer ranking EC2 spot instances by price and interruption risk, forecasting prices with ML, and generating deployable infrastructure.
open the live product ↗
Management reporting from live planning data is repetitive, time-consuming, and too important to leave to free-form model output.
My contribution: turned an existing internal prototype into the deployed version on SAP BTP Cloud Foundry with the Agents SDK. Agents retrieve live Jira data; a deterministic pipeline validates and renders the final reports.
All the planning context sits in Jira, and there was no API to it - so every dependency question turned into an afternoon of manual digging.
Built: a semester-planning assistant for the Business Suite Portfolio Management team. It reads the Jira data from a SQL mirror and runs purpose-built report and dependency workflows on top of it.
Comparison portals have data everywhere and no coherent interface for AI agents.
Built: one MCP surface across CHECK24 verticals, with host-specific adapters, conformance checks, routing, and observability - deployed on Cloudflare Workers. Won the CHECK24 GenDev scholarship.
read the repository ↗Lab experiments are expensive; the relationship between process conditions and product yields is hard to see.
Built: Random Forest and XGBoost models on 795 literature datapoints, grouped by paper to prevent leakage and checked with SHAP against the physics.
read the 93-page thesis ↗
DSGVO-compliant RAG support, browser automation for a medical practice, and n8n/MCP workflows.
Malt profile ↗Co-founded the relaunch and represented Claude on campus - helping students build with AI instead of only chatting with it.
club site ↗A jump-and-run level made after scavenging PC parts as a kid. No plan, just curiosity, stubbornness, and far too many moved vertices.
Still in school, I ran a student company repairing and refurbishing wall clocks - diagnosing faults, soldering, and replacing electrical parts. KooKoo’s Singvögel clock ↗ is still part of the story.
The one interface I use without shortcuts. Playlists I built years ago and still put on.
01:00 School - Supertramp ↗
Traveling, skiing, swimming, hiking - movement after too many hours in front of a monitor.
languages: German, English, Spanish - and collecting more
you found the konami code (or typed `lab`) · things in here are held together with tape
One level. No engine knowledge. An unreasonable number of hours moving vertices. Still counts as my first shipped game.
WhatsApp in, my server in the middle, Spotify queue out. Runs quietly at parties. The guests never know why the music stays civilized.
Buying neglected fixed-gear bikes, rebuilding, reselling. Analog engineering - the only project here with zero uptime monitoring.
