AI/ML Engineer building multi-agent systems and the MLOps that keeps them honest
📍 Bonn, Germany · 🎓 CS graduate · 💼 Formerly AI/ML Engineer @ Madiba Consult GmbH
At Madiba Consult GmbH I built production AI systems for GIZ (Deutsche Gesellschaft für Internationale Zusammenarbeit):
- Multi-Agent Proposal Writer — 7 LangGraph agents turn a tender's terms-of-reference into a compliance-checked, section-by-section draft (Plan Compliance Grading A–F, section coherence checks)
- Expert Matching Engine — semantic retrieval over ~800 consultant CVs (hard-filter → embed → re-rank), 701/744 CVs ingested from Azure Blob into Qdrant
- Document Processing Pipeline — parsing and structuring hundreds of tender PDFs and consultant CVs into clean, section-aware text that fed both systems above
31 shipped projects across agentic systems, MLOps, deep learning, data engineering, and full-stack — each with a real GitHub repo, real README, and a working live demo. Browse the full ledger:
→ personal-site-kappa-woad.vercel.app
A few worth a direct look:
| Project | What it does |
|---|---|
| PrivaRAG | 100% local RAG system — Qdrant + Ollama, zero cloud, full GDPR compliance |
| ThreatMail | 6 LangGraph agents triage inbound email in real time, structured SOC incident reports |
| SentimentFlow | Fine-tuned RoBERTa sentiment pipeline with full observability and automated eval DAGs |
| DefectSense | Real-time industrial defect detection — PaDiM, exported to ONNX, served via FastAPI |
| FlowLake | Distributed Spark/Delta Lake ETL with automated data-quality checks |
| MedShield | HIPAA-aware medical records vault, patient-controlled encryption, emergency access |
Agentic / LLM
MLOps / Serving
Deep Learning
Data Engineering
Product / Backend
Multi-agent orchestration, RAG architecture, evaluation/observability for LLM pipelines, or how to get a Docker Compose stack running cleanly at 2am.
Arabic (native) · French (C1) · English (C1) · Russian (B1) · German (A2, improving)