09.10.2026 aktualisiert
100 % verfügbar
Senior AI Engineer | LLM Systems, RAG & Multi-Agent Architectures | Python, LangGraph, AWS
Wien, Österreich
Weltweit
M.Sc. Applied Mathematics (with honors), University of ViennaÜber mich
AI engineer for production LLM and agent systems: RAG, multi-agent orchestration and the infrastructure to run them reliably. Built a multi-agent system that assigns ~10M government emails a year across 50 departments to the right case file with 90% accuracy. M.Sc. Applied Mathematics (honors).
Skills
RAGMLOpsLangChainAPIsKünstliche IntelligenzAlgorithmusAmazon Web ServicesAngewandte MathematikComputer VisionComputerprogrammierungDatenbankenContinuous DeliveryContinuous IntegrationDatenvisualisierungR (Programmiersprache)
LLM Systems & Multi-Agent Architectures
Design and production operation of LLM-based multi-agent systems: orchestrator/router patterns with dedicated specialist agents, tool calling, MCP (Model Context Protocol), agentic context engineering and prompt engineering. Frameworks and models: LangGraph, LangChain, AWS Bedrock, vLLM, Ollama, Hugging Face Transformers.
Retrieval-Augmented Generation (RAG) & Semantic Search
RAG pipelines and AI search assistants that give source-backed answers over large, German-language document collections. Embeddings (SBERT, transformer models), semantic search and vector databases: FAISS, Qdrant, pgvector.
Text-to-SQL & Data Agents
LLM agents that give non-technical users natural-language access to PostgreSQL databases, with modular tool interfaces for querying, reasoning over structured data and triggering email workflows.
MLOps & Cloud Infrastructure
End-to-end ownership from design to deployment and operation: FastAPI services, Docker, Kubernetes, CI/CD, AWS (SageMaker, Lambda, S3, EC2, Bedrock), experiment tracking and LLM observability with Weights & Biases and Langfuse.
Machine Learning, Deep Learning & Computer Vision
PyTorch and TensorFlow models for image classification, detection of manipulated, duplicated and AI-generated images, keypoint matching (OpenCV SIFT/SURF, LightGlue) and time-series forecasting.
Applied Mathematics & Optimization
M.Sc. in Applied Mathematics (with honors): numerical optimization, adaptive gradient-descent methods with convergence guarantees, approximation theory for neural networks (B-spline networks, universal approximation theorem, paper in preparation).
Programming, Data & Industries
Python, SQL, PostgreSQL, Git, R dashboards for non-technical decision-makers. Industry experience: public sector / government, financial market supervision, scientific publishing and research integrity.
Erfahrung
Berufserfahrung0 - 5 Jahre
Freelancer seit2025
Sprachen
DeutschMutterspracheEnglischverhandlungssicherSpanischgut
Projekthistorie
- Built and shipped a multi-agent system that automatically processes incoming government email across 50 departments: around 10 million emails a year, previously handled by hand, each attached to the correct case file among thousands with 90% accuracy.
- Architected it as a shared orchestrator routing to a dedicated agent per department, each learning new skills from experience (agentic context engineering) instead of needing constant retraining.
- Own the system end to end in production on the core AI team: design, deployment and operation.
- Built a Text-to-SQL agent (LangGraph) that lets non-technical staff pull data from PostgreSQL databases in plain English, removing their dependence on engineers for routine reports.
- Designed the modular tool interfaces the agent uses to query databases, reason over structured data and trigger email workflows, so it handles multi-step requests end to end instead of one-off queries.
- Tuned agent orchestration and prompt strategies so the system returns reliable, repeatable results in production, not just a working demo.
Tech: Python, LangGraph, LangChain, PostgreSQL, LLMs, prompt engineering
- Built computer-vision models that flag manipulated and duplicated images in scientific figures by matching them against a 100-million-image dataset, surfacing integrity problems in around 100 papers a year that are nearly impossible to spot by eye.
- Raised classification accuracy for the most common bioscience image types to 98%; implemented keypoint matching (OpenCV SIFT/SURF) and trained LightGlue models to match reused or altered regions.
- Co-built an end-to-end pipeline to detect AI-generated images, from labeling training data through automated evaluation.
- Part-time in 2025, alongside the freelance project.
Tech: Python, PyTorch, OpenCV, LightGlue, deep learning