Applied AI engineer building useful systems for people navigating complexity.
I work across engineering, product and operations — from an unclear problem to a usable workflow, then into retrieval, APIs, evaluations, guardrails and deployment.
Berlin · open to AI engineering roles and selected collaborations
Portfolio · LinkedIn · Email
GitLaw — verified legal AI for citizens and law firms
A legal research and workflow platform across 5,936 German federal laws. GitLaw combines BM25 + semantic retrieval, a 94K-node citation graph, local citation verification, MCP tools and a closed-beta workflow tier being tested with a Berlin migration-law firm.
Why it matters: legal AI should not sound certain when its evidence is missing. GitLaw keeps citations, uncertainty and human review visible.
Python · FastAPI · React · TypeScript · FAISS · BM25 · MCP · AWS · Vercel
SafeVoice — bounded AI for victims of digital harassment
A court-preparation workflow that classifies harmful content, structures evidence and drafts a reviewable criminal complaint. The agent is limited by typed outputs, iteration and cost caps, idempotent tools, audit trails and a human checkpoint before anything leaves the system.
Why it matters: automation in sensitive contexts should increase agency without quietly replacing accountable human judgment.
FastAPI · SQLAlchemy · PostgreSQL · React · Pydantic Structured Outputs · Vercel
Human Systems Lab — small experiments for public systems that work better
A civic product studio for measurable interventions in public services: multilingual benefit navigation, law evaluation, consent-based data reuse, participation tools and an impact-evaluation suite with hard safety and access gates.
Why it matters: rights and services only create value when people can actually reach and use them. The lab tests the smallest real intervention before proposing a larger platform.
TypeScript · React · Cloudflare Workers · D1/SQLite · Drizzle
-
Agent Loop Lab — durable TypeScript workflows with checkpoints, selective retries,
onFailurepaths and explainable evidence-backed scoring.
TypeScript · Inngest · Express -
Atelier Engine — a creator-commerce prototype for made-to-order sneaker brands, with live product configuration, demand signals and creator-economics modelling.
React · Vite · Tailwind -
Echo's Space Station — an interactive pixel world with an autonomous AI companion, multiplayer systems, games, an economy and citizen-science concepts.
JavaScript · Canvas · Node.js · Socket.IO · SQLite
Civic AI MCP Toolkit extracts repeated infrastructure from several public-interest MCP servers into one reusable package: server factory, traced tools, structured logging, health checks, error envelopes, fixtures and CLI scaffolding.
Python · FastMCP · pytest
- Start with the user, decision and failure cost — not the model.
- Build a narrow vertical slice quickly enough to learn from.
- Make evidence, uncertainty and human responsibility inspectable.
- Turn repeated work into reusable infrastructure.
- Describe pilots as pilots and unfinished work as unfinished work.
I am especially interested in applied AI, document-heavy workflows, civic technology, agent reliability and product engineering.


