Data Architect /Data Strategist/ Data Engineer / Data Scientist/ Data & AI Management

6/2025 – offen

Data Engineering for SAP–Databricks Pipelines in Environmental Sustainability Reporting

Design and implementation of enterprise-scale SAP-to-Databricks data pipelines to support environmental sustainability reporting for chemicals, packaging, and real substances. The objective was to enable accurate tracking of the global environmental footprint and sustainability metrics, supporting regulatory compliance and monitoring of chemicals under investigation across multiple business departments.

Comprehensive analysis of SAP business-process-driven data pipelines, including in-depth examination of SAP table schemas, joins, inheritance and self-inheritance patterns, and complex inter-table relationships. Each business process typically involved 5 to 12 interrelated SAP tables, requiring careful orchestration to preserve business semantics.

Definition and implementation of a Data Mesh–based data pipeline architecture, alongside Databricks schema modeling to support reliable daily ingestion of SAP data into the Databricks Lakehouse.

Implementation of PySpark-based polymorphic class structures in Databricks to capture and encapsulate SAP business logic for chemical and real substance processes. This abstraction enabled a highly automated, robust, and error-free flow of substance data from SAP to Databricks.

Post-processing of raw SAP data using specialized PySpark classes and Databricks utility functions for joins, data cleaning, filtering, and aggregation. These transformations produced customized data pipelines tailored to the specific needs of different business units.

Publication of processed datasets as governed data products, registered in the enterprise data catalog and made available for subscription by multiple departments. Creation of secure Databricks views governed by Unity Catalog authorization groups, enabling controlled access by downstream consumers. Integration with Power BI to allow business users to consume curated sustainability data products.

Implementation of data lifecycle management and end-to-end data lineage tracking within Databricks to ensure transparency, traceability, and compliance.

Ongoing, on-demand development of new sustainability data pipelines, frequently requested by business units to address evolving compliance requirements and newly defined sustainability data logic.

Building agentic workflows using Databricks Agent Bricks to automate dynamic data pipelines and workflows, addressing subtle toxicity nuances among chemical substances with very similar compositions.

Development of real-time and near–real-time data ingestion workflows using Azure Data Factory, Azure Fabric, Databricks Lakehouse, and Azure Event Hub to ingest data from external systems, perform preprocessing in Azure, and persist curated outputs as governed data products.

Azure Synapse Analytics, Big Data, Databricks, Microsoft Azure

11/2024 – 5/2025

1) Agentic AI with Reasoning and Reinforcement Learning:

Design and implementation of an enterprise-grade Agentic AI reasoning system based on System 2 reasoning, operating on large-scale enterprise and historical customer data. The system autonomously plans and executes complex collections of customer tasks, leveraging long-term customer interaction history.

Development of novel Reinforcement Learning (RL)-based AI agents to intelligently automate traditionally manual enterprise processes. These agents synthesize new, efficient AI workflows using Chain-of-Thought (CoT) and Tree-of-Thought (ToT) reasoning frameworks.

Implementation of Tree-of-Thought reasoning using Monte Carlo Tree Search (MCTS), guided by a custom-designed Upper Bound Cost function. Introduction of a new GRPO-based RL reward formulation for scoring and optimizing reasoning trajectories. Utilization of Proximal Policy Optimization (PPO) and Process Reward Models (PRM) to guide agent behavior and enable robust temporal reasoning.

Creation of a novel Risk Agent Model (RAM) to compute and evaluate the economic, ethical, and corporate risks associated with each AI agent decision. Integration of a Human-in-the-Loop (HITL) module that continuously monitors high-risk decisions, enabling human operators to approve, modify, or halt actions prior to execution.



2) Optimization of GPU Training Cost and Performance:
Optimization of foundational model training to significantly reduce GPU costs. Re-engineering of training pipelines for NVIDIA A100 GPUs by transitioning from pure FP32 computation to mixed precision (FP32 + FP16/BF16).

Detailed FLOP analysis and recomputation, including precise token-size selection based on target foundational model parameter counts, to improve computational efficiency without degrading model quality.

Low-level CUDA optimization using C++, enabling large-scale AI models to achieve high performance on lower-cost GPUs, reducing reliance on premium hardware such as A100 while maintaining training throughput and stability.

Informatik, Mathematische Optimierung, Neuronale Netze, Reinforcement Learning, Transformer

4/2024 – 10/2024

Development of Internal Corporate ChatBot based on Azure RAG Vectorization


Development of a RAG Indexer Pipeline that ingests EnBW corporates Documents (i.e., PDF, Powerpoints, Images etc.) from SharePoint and MSTeams into Azure Blob Storage and data lake.

Write python codes that use Azure Form Recognizer, Azure Document Intelligence, Azure AI Search to perform OCR text extraction from PDFs and Images.
Implementation of large document semantic summarization using LangChain and OpenAI. Development of document chunking and data vectorization techniques to build a Vectorization Index that it used by the corporate chatbot.

Usage of Azure ML Studio to create integrated vectorization indexing pipeline for the RAG, which is based on Azure AI Search Enrichment Skillsets. Creation of Azure Functions to handle Custom Skillsets during document processing.

Development and extension of a frontend chatbot python based application using Django

Development of CI/CD pipeline for the automatic End-to-End ingestion, cracking, chunking, enrichment and vectorization of incoming documents from share point and Azure Data Factory.

Django, Document Retrieval, Microsoft Azure