GitLaw
Find and inspect relevant German federal law with traceable sources. The current Mietrecht pilot turns user-found retrieval failures into permanent regression cases; broader lawyer task-success validation is still pending.
I build AI systems that turn complicated workflows into clear, useful tools. My background combines hands-on engineering with founder, product and operations experience. I am strongest when I can go from a messy problem to a working product, test where it fails, and improve it with real users.
Find and inspect relevant German federal law with traceable sources. The current Mietrecht pilot turns user-found retrieval failures into permanent regression cases; broader lawyer task-success validation is still pending.
Turn online-harassment text or screenshots into a structured, reviewable evidence package. The end-to-end flow works; important classifier failures are documented and broad real-user validation is still next.
A reusable pattern for AI workers that must show evidence, expose failure and ask a human before consequential action. Public scenarios and the current reliability suite are synthetic; a real shadow workflow is the next proof.
Review documents against versioned rules and show what is missing before a human decision. The PDF intake/extraction core works; the public reviewer cases are synthetic and the system is not deployed in an authority.
Built and operated an Amazon business end to end across sourcing, logistics, fulfilment, listings and customer support; turned repeatable operational work into clearer processes.
Coordinated daily operations for 8+ staff, up to 150 covers per day and delivery workflows; resolved bottlenecks under peak-service pressure.
Scaled omnichannel e-commerce across marketplaces, website and B2B partners; managed acquisition channels and used customer feedback to improve positioning.
Full-stack software engineering, extended through hands-on work in LLM applications, RAG and agent systems.
Real users. Real corrections. Measurable outcomes.
I have built quickly across several domains. The next step I care about most is putting the strongest systems in front of target users, learning from their corrections and turning those failures into better products.