I build reliable data and AI platforms that scale.

• Diagnosed PostgreSQL connection-pool saturation (393 of 500 connections in use, about 120 idle per API pod), corrected pool limits, and documented a repeatable validation workflow.
• Tuned Celery/Redis async processing, reducing p95 chat response from about 45s to about 18s during peak office traffic.
• Hardened OpenShift deployments (route/nginx timeouts, Celery Beat access, readiness probes, resource limits) and wrote runbooks for deploy, rollback, and on-call handoff.
• Wrote architecture diagrams and internal guides on incident response, RAG retrieval behavior, and LiteLLM/vLLM gateway routing.

Selected clients: Continental AG, Valeo, Fraunhofer, Maisel, HKN

• Built Kafka anomaly-detection pipelines on AWS (VPC, Fargate, ECR) for Continental AG at about 800,000 events/day; alerts in about three minutes and about 14% less unplanned downtime over six months on two pilot lines.
• Reduced Maisel SharePoint-to-data-lake ingestion from about 3.5h to about 28m with nine Airflow DAGs, 62 data-quality tests, and Prometheus/Grafana on a three-node Kubernetes/MinIO cluster (Hetzner).
• Built a hybrid LLM platform on Kubernetes for 45 Maisel users (Ollama, OpenWebUI, LLM proxy), cutting public-API inference spend by about 35% (about €1,200/month) while meeting EU data-residency.
• Redesigned Valeo BigQuery schemas and Dataflow ingestion at about 22GB/day, cutting a plant analytics report from about 5h to about 55m.
• Automated HKN email infrastructure for 280 mailboxes (Docker, SOAP/XML, GitHub Actions CI/CD) with nightly archiving of about 40GB/month.
• Delivered a Fraunhofer reinforcement-learning traffic pipeline (SUMO, Kafka, Protobuf, MLflow, Kubeflow): 68 experiments and about 11% lower mean vehicle wait time on the simulation benchmark.