Leon B.

I build and evaluate systems in which several AI models work together on long projects. Since July 2026 I have developed a local runtime in Python and SQLite that coordinates Claude Code, Codex, Gemini and OpenCode in separate roles: planning, execution and independent review. Git hooks restrict each agent to its assigned files, acceptance criteria run as executable tests, and no model may approve its own work. I used it for quantitative trading research: 366 logged strategy trials across 48 instruments, validated with walk forward splits, purged cross validation, deflated Sharpe ratios, White’s Reality Check and a locked holdout set. My daily work is reading model output critically. I catch fabricated sources, faked test results and overstated claims, and I turn them into concrete fixes. I study Sustainable Development (business focus) at Hochschule Bochum. I am a native German speaker, served two years in the German armed forces and am used to applying strict guidelines precisely.