About Me

Designed and built an LLM web agent for multi-step workflows using agentic context engineering and selective multimodal reasoning, reaching a strict success rate of 69.7% to 95.9% across multiple benchmarks.

• Context engineering: structured DOM preprocessing as the agent's context foundation.
• Agent memory: knowledge-graph architecture for sequential states in long-running tasks.
• Cost-aware model routing: compute-heavy vision-language models are invoked only when pages are ambiguous or the system state is uncertain.
• Guardrails & human-in-the-loop: a validation layer combining model confidence with rule-based safety criteria; critical actions run only after human approval.
• Deployment: CI/CD production pipelines for MCP servers and software on Microsoft Azure and Microsoft Foundry.

Bandwidth extension of occluded in-ear speech signals (earbuds, headbands) via neural reconstruction.

• Curated and recorded a custom speech dataset; developed data augmentation strategies for different noise sources.
• Implemented, trained and compared real-time-capable models with PyTorch, Librosa and SciPy.

AI engineering and research software at TIB – Leibniz Information Centre for Science and Technology.

• TIB AIssistant: co-developed an LLM-based assistant that supports researchers across the entire research and writing process with reliable, source-grounded answers.
• OntoAligner: contributed to development and documentation of a Python library for ontology alignment (Best Resource Paper Award, ESWC 2025).
• NLP4RE: automation pipeline converting NLP4RE ID cards into the Open Research Knowledge Graph (ORKG) to answer competency questions over the knowledge-graph data.
• EmpiRE-Compass: led a team of 5 students building a neuro-symbolic dashboard for analyzing research knowledge (Best Posters & Tools Award, REFSQ 2026).

Built AI-driven automation for process management in ERP and CRM systems (React, Next.js, .NET).

• Lead developer of point-of-sale (POS) software as progressive web apps (PWA).
• Bachelor's thesis (grade 1.0): optimized visitor and freight distribution by combining routing algorithms (VROOM) with clustering methods for shorter routes and higher efficiency.
• Coordinated AI thesis projects with the university; mentored interns.

Mentored over 100 students in Machine Learning, Deep Learning, Data Analysis, and Software Engineering. Led interactive workshops, trained students in end-to-end project development, and organized comprehensive exams to assess their progress.

Research on agentic AI systems and their foundational methods, from training agents to representation learning with JEPA (Joint-Embedding Predictive Architectures). SciJEPA: JEPA embeddings for scientific knowledge graphs; supervised by Prof. Dr. Sören Auer.

Thesis: Supporting Multi-Step Workflows with LLM-Based Web Automation using Agentic Context Engineering and Selective Multimodal Reasoning, in cooperation with Volkswagen AG.

Teaching Assistant for six B.Sc. courses, including Artificial Intelligence, Databases and Software Engineering. Bachelor's thesis (grade 1.0) on route optimization with VROOM and clustering.