Llama Experts in Munich

Co-founded an AI-native language learning startup, defining the product vision, AI architecture and technical roadmap. Designed and built the AI and backend stack, including LLM fine-tuning pipelines, custom agentic workflows, and scalable inference infrastructure. First product currently in private beta.

• Research and develop models for chatbots, NLP and LLMs (e.g. Llama, Qwen, OpenAI)
• Enhance chatbots with RAG, in-context learning
• Supervised fine-tuning (PEFT, LoRA), Huggingface or Unsloth
• Advanced training methods: Test-time training, (transductive) active learning, reinforcement learning
• High-throughput serving with vLLM
• Apply embedding models (e.g. SentenceTransformers), similarity/vector search or vector DB or ranking (e.g. LlamaIndex, Faiss, LangChain)
• Generate and filter synthetic data, clustering
• Detect hallucinations
• Evaluate chatbot models (Rouge, BLEU, F1-Score, Recall, Precision)
• Visualization of experiments (matplotlib)

AI-driven automation of purchasing on a printing company's website, including selecting delivery times, order options, ordering, payment, and uploading print data from the Lumiz Cloud.

• Head of overall technical integration NSC (National Secure Cloud, project with approx. 60 employees).

• Technical integration of all subprojects into one product, definition of interfaces and basic components of a cloud including hardware, technical architecture of the IABG platform.

• Development of a Cloud Management Platform (CMP) capable of creating private/mixed clouds of any complexity based on a textual description with one click or interactively.

• CMP also includes the complete hardware management lifecycle.

• Kubernetes, OpenStack and Hadoop are used as the foundation.

• The management layer includes Harbor, Gitea, Longhorn, Keycloak, Rancher and Jenkins, which are configured automatically.

• Private cloud can run any customer workloads, including a full Hadoop layer with HDFS, Spark, MapReduce, Mesos, HBase and around 20 additional ML/DL technologies.

• Hadoop worker clusters can also be installed automatically without Kubernetes on bare metal or commodity hardware.

• OpenStack with Nova, Neutron, Ironic, Swift, Cinder, Ceph.

• Development of a Java application Rudi: SOAP, REST, containers, DB.

• Technologies: Kubernetes (K3s, Rke2, Minikube, Harbor, Gitea, Jenkins, Longhorn, Keycloak, Rancher), OpenStack (Nova, Neutron, Keystone, Swift, Ceph, Cinder, Sahara, Magnum, Kayobe, Kolla, Bigrost, Ironic), Hadoop (HDFS, Ambari, Solr, Livy, Ranger, YARN, Tez, HBase, Kafka, Hive, Zookeeper, MapReduce, Spark, Oozie, Flink), virtualization (Kubernetes (K3S), VMware, Oracle), scripting (Ansible, Puppet, Juju, Shell, Groovy, Gradle, Maven).

• Architected an agentic, real-time offer orchestration engine where specialized agents (retrieval, pricing/optimization, and policy/guardrails) coordinate to personalise promotions across customer touchpoints using RAG with FAISS over Delta Lake and low-latency Databricks Model Serving. Collaborated with product managers and commercial stakeholders to shape the roadmap and evaluate emerging agent patterns for production.

• Designed an agent-based data quality service that orchestrates schema detection, entity normalization, and validator/exception-handling agents to clean multi-retailer SKU feeds at scale. Wrapped model calls in PySpark UDFs for distributed inference, automated via Databricks Workflows and CI/CD.

• Developed a multimodal, agentic extraction pipeline where vision, parsing, and compliance agents collaborate to derive brand, packaging, and volume from scanned images using Claude 3 Sonnet with Swin Transformer encoders. Orchestrated via Azure Event Hub with outputs persisted to Delta Lake.

• Implemented a GS1 taxonomy classification service built around cooperating agents for inference, drift monitoring, and auto-retraining governance using Falcon 180B (LoRA-tuned) with a batch pipeline on Databricks.

• Created a hybrid agent workflow where a retrieval agent surfaces candidate matches via embeddings and a reasoning/verification agent (Mixtral 8x7B) adjudicates receipt-to-SKU alignment, integrated into a streaming Databricks pipeline.

• Built a multimodal attribute inference pipeline structured as cooperating vision-language, rules/consistency, and compliance agents to fill NutriScore, nutrition fields, and packaging types from names and images using LLaMA 3-8B with CLIP embeddings.

• Developed a GenAI-powered orchestration system that ingests recipes from multiple websites, parses ingredients through structured extraction agents, and dynamically links them to real-time retailer offers via tagging, semantic reasoning, and business-rule agents.

• Building the company with a focus on connectivity for IIoT and Industry 4.0, iRPA/process automation, advanced robotics and smart systems, sensors and services
• Project management and software architecture for IoT gateway development (since 2020) with protocol translation, IT/OT convergence and GRC
• Developing RPA bots for automating and monitoring industrial processes with an agent-based AI approach (since 2020)
• Implementing unsupervised clustering and anomaly detection for time series data in big data streaming pipelines (since 2021)
• Introducing a Docker-based release train for OTA updates with DevSecOps and CI/CD (since 2018)

• Software development of a B2B SaaS for AI-based search in internal candidate pools of recruitment agencies
• Design of a multi-tenant, hybrid architecture with dedicated GPU servers and secure cloud integration
• AI-Engineering
• LLMOps
• Python
• FastAPI

• Deployed and optimized Vision-Language-Action (VLA) and diffusion policy models on NVIDIA Jetson Orin and Jetson Thor, meeting real-time inference latency targets for humanoid robot control loops.
• Built TensorRT engine pipelines (PyTorch → ONNX → TensorRT) with INT8/FP8 post-training quantization, calibration dataset design, and quantization-aware validation, reducing inference memory footprint by over 3× on Jetson without accuracy regression.
• Developed custom CUDA C++ plugins and CUDA Graphs for latency-deterministic, real-time policy execution – meeting hard runtime and memory constraints on embedded GPU targets.
• Developed an inference engine for VLA models on top of llama.cpp bringing different VLA policies under single runtime, packaging each as a single self-contained GGUF that needs no Python or PyTorch.
• Profiled and tuned GPU execution using NVIDIA Nsight Systems and Nsight Compute, identifying CUDA kernel bottlenecks, memory bandwidth saturation, and SM occupancy issues across Jetson Orin and Thor compute profiles for cross-layer performance optimization.