About · Erfurt → Potsdam
Curiosity, competition, then building.
The short version is on the start page. This is the longer one.
The short story
Before I change a product I reproduce the problem on a real transcript and sweep every configuration in the fleet. Defaults stay conservative, per-customer opt-in rather than global switches. Every hand-over names what I did, what I did not do, and what could go wrong, written so that a non-engineer can follow it.
I grew up in Erfurt, went through a specialized STEM school and the federal computer-science competition, and moved to Potsdam for HPI in 2024. I work in focused sprints and use AI coding agents heavily, and say so: the architecture decisions, the test strategy and the review are mine.
-
01Curiosity
Math olympiads from sixth grade, the entrance exam for a specialized STEM school in Erfurt, the federal computer-science competition.
-
02Competition
Thirteen years of karate. A CNN written from scratch won the Thuringian round of Jugend forscht and a special prize at the national final (2024).
-
03Building
A winning team at the first Bundesliga Hackathon (2025). Then production code: Founding Engineer at Kobe AI, and a voice agent for 100+ salons at MitJulia (formerly Friseurtelefon).
-
04Impact
Medical volunteer in Red Cross civil protection since October 2024; a workshop day for 87 eleventh-graders with my scholarship foundation. Skills only count once they work outside your own CV.
How I work
- 3,200 / 0 comparisons on 400 generated catalogs against an independent re-implementation, 0 deviations
Evidence before code
I reproduce a problem on a real transcript and check every configuration before I change anything.
- 32 in 5 days commits on the takeøff website; 477 at Kobe in October 2026
Focused sprints
Short, intense build phases with a clear goal, then review and clean-up.
- 8+ projects with written agent contracts (AGENTS.md, CLAUDE.md)
Agents, openly
I use AI coding agents heavily and say so. The architecture decisions, the test strategy and the review are mine.
Skills, with evidence
7AI agents and evalsLLM agents with tools, MCP and voice, plus the evals that keep them honest. 267eval cases, grown from 67 (MitJulia, 2026)
A production voice agent for 100+ salons at MitJulia with an eval simulator (67 → 267 atomic cases, GPT-4o-mini as the caller); a six-stage agent pipeline on the Claude Agent SDK at Kobe, co-developed, with a rubric review agent that scores reports on five dimensions (0–3); extended the MCP server of the Kobe backend; the OpenAI Agents SDK and Realtime API for Voya; a talk on AI hallucination (Biases & Mishaps, 1.3).
- Claude Agent SDK
- OpenAI Agents SDK
- MCP
- ElevenLabs
- Twilio
- Voice agent in production at MitJulia, 100+ salons (Nov 2026); a simulator lets GPT-4o-mini play the caller across 267 atomic test cases, grown from 67.
- Kobe: co-developed a six-stage agent pipeline on the Claude Agent SDK (about 20k lines of Python), including a rubric review agent that scores reports on five dimensions (0–3), and extended the MCP server our CTO had set up (27 of 89 commits).
- Before one product change I checked a booking gate against an independent re-implementation: 400 generated catalogs, 3,200 comparisons, 0 deviations.
- Talk on AI hallucination in the course Biases & Mishaps (1.3).
Full-stack TypeScriptProduction code in two startups: web, backend and an Electron desktop client. 434 + 477commits at MitJulia and Kobe, 2026
434 commits in the MitJulia monorepo (Aug – Nov 2026) and 477 commits at Kobe in October 2026: Next.js and React, Node with Hono, Electron, Vite.
- TypeScript
- React
- Next.js
- Node.js
- Hono
- Electron
- MitJulia monorepo: 434 commits between August and November 2026.
- Kobe: 477 commits in October 2026 across the backend, the Electron desktop client and the pitch deck.
- takeøff website: bilingual, 32 of 33 commits in five days, live on Vercel.
Python and machine learningFrom a CNN written from scratch in NumPy to PyTorch benchmarks and YOLOv8. 7architectures benchmarked in PyTorch (PneuNet)
A CNN with its own backpropagation in NumPy (2023), seven architectures benchmarked in PyTorch, YOLOv8 segmentation on a CPU, and now the GorillaScout bachelor's project (detection, tracking and re-identification of wild gorillas).
- Python
- PyTorch
- NumPy
- OpenCV
- YOLOv8
- FastAPI
- PneuNet: a CNN with its own backpropagation in NumPy, then seven architectures in PyTorch; Jugend forscht 2024, Thuringian state winner.
- Real-time trash segmentation with YOLOv8 on a CPU, winning team at the Siemens × AWS hackathon 2025.
- Now: GorillaScout, a bachelor's project (team of eight) on detecting, tracking and re-identifying wild gorillas.
Cloud and dataServerless on AWS and Cloudflare, Postgres with Drizzle, Supabase, SQLite. 100+Lambda functions in one AWS backend (MitJulia)
An AWS SAM backend with over 100 Lambda functions and 9 DynamoDB tables; Cloudflare Workers in four projects, some with Durable Objects; Postgres and Supabase across several projects, with Drizzle and row-level security at Kobe; SQL, Database Systems I graded 1.7.
- AWS
- Cloudflare Workers
- PostgreSQL
- SQL
- Drizzle
- Supabase
- Vercel
- MitJulia: an AWS SAM backend with over 100 Lambda functions, 9 DynamoDB tables and S3.
- Cloudflare Workers in four projects, some with Durable Objects; Postgres and Supabase across several projects, with Drizzle and row-level security at Kobe.
- Database Systems I graded 1.7.
Engineering practiceArchitecture, tests, CI and code review, with AI coding agents under clear contracts. 1.0in both Software Architecture and Software Engineering
Software Architecture and Software Engineering both graded 1.0 (exam, system demo, code review); design patterns, ADRs and Scrum in the team projects of both courses; CI in a seven-year-old codebase with coverage raised from 84 to 95 percent (team); agent contracts (AGENTS.md, CLAUDE.md) in 8+ projects.
- Git
- GitHub Actions
- Vitest
- Playwright
- Claude Code
- Software Architecture and Software Engineering both 1.0: exam, system demo and code review.
- Set up CI in a seven-year-old codebase; the team raised test coverage from 84 to 95 %.
- Agentic workflow: AGENTS.md and CLAUDE.md contracts in 8+ projects, spec → ADR → code; 50+ parallel agent worktrees at the Voya hackathon.
- Design patterns, ADRs and Scrum in the SWA/SWT team projects.
Processes and requirementsBPMN 2.0, interviews and requirements, the enterprise view on software. 3BPMN projects, from interviews to process models
BPMN 2.0 in three projects; for the ESN Event Hub five stakeholder interviews that surfaced 15 pain points; Enterprise Applications 1.3; Global Design Thinking Workshop at the HPI d-school 1.0.
- BPMN 2.0
- Requirements engineering
- Design thinking
- BPMN 2.0 in three projects: Modeling (1.3), ESN Event Hub (1.3), one ongoing.
- ESN Event Hub: five stakeholder interviews, 15 pain points.
- Enterprise Applications 1.3; Global Design Thinking Workshop at the HPI d-school 1.0.
AlsoKotlin and Jetpack Compose, Flutter, C and C++, Smalltalk, Bash and Linux. 26klines of Kotlin in my largest app
Kotlin and Jetpack Compose (a 26k-line Android app), Flutter, React Native with Expo, C and C++, Smalltalk, Bash and Linux.
- Kotlin
- Jetpack Compose
- Flutter
- React Native
- C++
- Smalltalk
- Bash
Languages
- Germannative
- Englishfluent (C1)
- SpanishB1
- Frenchbasic (A1–A2, UNIcert I)
Beyond code
- 1st DanKarate13 years; German vice-champion U14 kumite
- since 2024Red CrossVolunteer emergency medical responder
- 23 countriesTravelincl. 17 days by van to the Arctic Circle
Interests
- Karate13 years, 1st Dan
- Hackathons
- 3D printing
- Event techtakeøff collective