Applied AI Engineer building LLM agents, retrieval systems, and the evaluation harnesses that say whether they actually work.

• Built an AI agent governance platform — Grace — shipping an MCP tool-calling server, WebSocket transport, signed execution receipts and a Merkle audit chain on FastAPI + Next.js + Postgres + Redis. 939 test functions across 165 files, with CodeQL, Trivy, gitleaks and TruffleHog wired into CI.
• Deep on RAG and retrieval — self-corrective LangGraph agents with article-level citations (german-compliance-agent), and hybrid BM25+vector, GraphRAG/Neo4j and Atlas vector search compared on one corpus (Trinity).
• Evaluation-first: I build the eval before the claim, pre-register hypotheses, and publish negative results. Three of my repos report numbers that did not go my way.
• Backend-first: Python, FastAPI, SQLAlchemy — plus Docker, Terraform, GitHub Actions, GCP Cloud Run.
• M.Sc. Artificial Intelligence & Machine Learning, SRH Berlin — thesis on effort gaps in compressed LLMs (pre-registered, ActFloor).
• Previously Technical Analyst at Oracle India (2+ years).
• Based in Berlin.

• SlopCheck — grounds AI-written vulnerability reports against the source tree they describe. Measured on 557 labelled curl reports: it contradicts 29.4% of confirmed vulnerabilities, so it fails as a triage gate. Published as a negative result with a DOI and the corpus.
• Grace — cryptographic governance receipts for AI agents: 6-stage governed execution, Ed25519 + ML-DSA-65 signing, RFC-6962 Merkle chain, MCP server.
• german-compliance-agent — LangGraph agent answering GDPR questions with citations, and auditing contracts against them. (TODO: add citation-accuracy numbers over a labelled set.)
• ActFloor — pre-registered MSc thesis experiments on compression and effort gaps in LLMs, with the findings that failed to replicate written up alongside the ones that held.
• solar-forecast-de — day-ahead solar forecasting for the German grid. 2.57% nMAE on a 25-day walk-forward backtest, and −11.5% skill against the TSO baseline on synthetic data, where that baseline is unrealistically strong.
• Trinity — one ENISA corpus, three retrieval backends, one harness. (TODO: replace the saturated eval set before quoting a number.)

Reach me: LinkedIn