At ImmoScout24, I designed and shipped a production AI agent that handles 20,000+ daily support requests and moderates 30,000+ real estate listings. It combines semantic search, tool calling, multi-step planning, and escalation to human operators.
I built its hybrid RAG pipeline and evaluation platform, achieving 0.85 Recall@10 and 0.92 faithfulness. Confidence-based routing cut complex case handling from 12 to 4 minutes.
At ImmoScout24, I also built an incremental embedding update pipeline that reduced recomputation from 47 minutes to 3.2 minutes. I own the production MLOps stack, including Kubernetes deployment and monitoring.
Previously, at AutoScout24, I built search and ranking systems for vehicle listings across 24 European countries. The work improved search conversion by 8.3% and reduced listing processing time from 30 minutes to 8 seconds.
