I'm an ML/AI Engineer who takes generative AI from a whiteboard idea to a system real users depend on. My focus is LLM applications, retrieval-augmented generation (RAG), and agentic, multi-agent systems, plus the evaluation, guardrails, and LLMOps that make them trustworthy in production.
Currently I'm building LLM platforms at ITUC GmbH. Before that I co-founded the edge-AI startup ROVEONE, and I was the first ML engineer at truemetrics (YC S23), shipping time-series ML on AWS. This year our agentic systems won 1st place at the InnoTrans 2026 Hackathon.
Outside my day job, I'm building a robotic arm from scratch, from the motors and electronics up, and teaching it to handle objects on its own with vision-language-action (VLA) models: it learns first by copying demonstrations (imitation learning), then improves by trial and error (reinforcement learning). More in Physical AI. In my free time I fly racing drones, build small self-driving cars, run, and train karate.
“
Move fast on evidence. Move carefully on power.
my rule of thumb
02Skills
What I bring to the table
The full GenAI lifecycle, owned end to end: from POC to production.
↻ Monitor in production → learn → iterate
01
Stage 01 of 05
Discover & Prototype
Turn a fuzzy business problem into a working GenAI proof-of-concept in days, not quarters — rapid prompting, model selection, and a demo stakeholders can click.
What I do
Use-case framing, model choice, prompt prototypes
You get
Clickable GenAI demo in days
Prompt Engineering
Claude / GPT / Gemini
Llama · Qwen · Nemotron
Streamlit
Jupyter
Design Thinking
02
Stage 02 of 05
Retrieve & Ground
Production-grade RAG: ingestion, chunking, hybrid search, re-ranking, and knowledge graphs — so every answer is grounded and cited, never hallucinated.
What I do
Ingestion, chunking, hybrid search, re-ranking
You get
Answers cited to your own data
RAG
GraphRAG
Vector DBs (Qdrant, pgvector)
Neo4j
Embeddings & Re-rankers
Hybrid Search
Docling / OCR
03
Stage 03 of 05
Orchestrate Agents
Multi-agent systems with supervisors, specialists, and verifiers — tool calling over MCP, guardrails, and human-in-the-loop by design.
If it isn't measured, it isn't shipped. LLM-as-a-judge, golden datasets, regression suites, latency/cost tuning, and fine-tuning when prompting hits its ceiling.
What I do
Golden sets, LLM-as-judge, regression suites, tuning
You get
Measured quality, latency & cost
LLM Evals
LLM-as-a-Judge
RAGAS
LoRA / PEFT Fine-tuning
Quantization
Observability (Langfuse)
05
Stage 05 of 05
Deploy & Scale
Containerized, CI/CD-driven, cloud-native deployments with streaming APIs, model serving, monitoring, and cost control — LLMOps end to end.
What I do
CI/CD, containers, serving, monitoring, cost control
You get
Production service that scales
FastAPI + SSE
Docker
Kubernetes
vLLM
AWS · GCP · Nebius
GitHub Actions
MLflow
Airflow
01 / 05
Tech stack
Tools I ship with. Hover a category or a tool.
LLM products, RAG and multi-agent systems, designed for real users rather than demos.
Classic ML and deep learning, from sensor time-series to reinforcement learning.
Pipelines, tracking and cloud infrastructure that keep models reliable in production.
Solid software engineering: typed code, clean APIs and streaming backends.
Robots that learn: VLA policies, simulation, edge inference and the electronics underneath.
I work comfortably across teams and clients in four languages.
Hover or tap any technology to see where I've used it.
03Experience
Where I've shipped
Scroll to follow the path. ★ marks milestones and achievements; hover a card for the full story.
Architect and ship LLM-powered products end to end — from rapid GenAI proof-of-concepts to production deployments.
Build retrieval-augmented generation (RAG) and agentic workflows with tool calling, grounding, and guardrails.
Own LLMOps: CI/CD, containerized deployment, experiment tracking, and monitoring for production readiness.
Run systematic LLM evaluation, benchmarking, and cost/latency optimization.
Translate business problems into AI solutions with cross-functional and client stakeholders.
LLMs
RAG
LangChain
LangGraph
FastAPI
Docker
GitHub Actions
MLflow
Airflow
Hover for details
2023
Oct 2023 – Aug 2024
Role
Technical Co-Founder & Lead ML Engineer
ROVEONE · Oct 2023 – Aug 2024
★ Co-founded the company and joined the board Oct 2023
★ Shipped on-device threat detection on the Apple Neural Engine
Co-founded an edge-AI civilian protection startup; served on the board, led the privacy-first iOS threat detection engine.
Designed a Bayesian sensor-fusion engine combining audio, motion, and contextual signals; deployed optimized YAMNet/CoreML models on the Apple Neural Engine.
Built a GDPR-compliant GCP data and active-learning pipeline for ingestion, labeling, model improvement, and analytics.
TensorFlow
CoreML
Python
Swift
GCP
Docker
Hover for details
2022
Apr 2022 – Oct 2023
Role
ML Engineer — Intern & Freelancer
truemetrics GmbH (YC S23) · Apr 2022 – Oct 2023
★ Promoted from intern to the startup's first ML Engineer
★ Production ML pipeline live on AWS
Progressed from intern to the startup's first ML Engineer; continued as freelancer with end-to-end ownership.
Analyzed and visualized large-scale mobility/sensor data for logistics insights.
Developed, fine-tuned, and evaluated time-series ML models for real-world predictions.
Built and deployed a production ML pipeline on AWS — processing, feature extraction, experiment tracking, inference.
Python
TensorFlow
tsfresh
AWS
MLflow
Scrum
Hover for details
2018
Sep 2018 – Oct 2020
Role
Working Student — Transmission & Hybrid Systems
IAV GmbH · Sep 2018 – Oct 2020
★ Built a reinforcement-learning framework for automotive applications
Developed a reinforcement learning framework for automotive applications.
Automated data preprocessing pipelines for efficiency.
Implemented and evaluated RL and supervised learning algorithms.
Agentic supplier-risk assistant for Siemens Mobility rail procurement — a Nemotron supervisor fans out to four specialist agents (data, documents, email, web) and returns cited, explainable buy/avoid recommendations. Built and deployed with ITUC colleagues in under 6 hours.
1st place, InnoTrans 2026 AI Track. A four-agent system — Dispatcher, Analyst, Inspector, Writer — that turns a plain-language question into a verified railway control-room brief. Every number comes from a tool call, never from the model.
AI clinical scribe for German hospitals: a wearable streams the consultation, German medical speech-to-text plus fine-tuned NER and a LoRA-tuned Gemma model turn it into a SOAP note, and a RAG companion helps fill the hospital's KIS. Built at the Berlin Big Hackathon 2026, then grown into an MVP.
Work in progress: an agentic AI co-planner for railway electrical engineering. It reads CAD drawings, calculation sheets and norms together, drafts the authority-ready documents with citations and flags every discrepancy, while the certified engineer reviews and signs.
Ongoing Physical AI build: a self-built robotic arm with custom electronics and embedded control, driven by vision-language-action policies trained via imitation learning and refined with reinforcement learning.
Deep-learning interface for the BBCPy brain-computer-interface toolbox (TU Berlin): train, tune and benchmark PyTorch models on motor-imagery EEG with a standardized, automated ML lifecycle. Built during my master's thesis.
A generic toolkit that exposes GNU Radio as an OpenAI Gym environment, so reinforcement-learning agents can train directly on software-defined radio scenarios.
Designing and building a robotic arm from the motors up, then giving it a brain: embedding VLA models trained with imitation learning from teleoperated demonstrations and refined with reinforcement learning — closing the loop between foundation models and the physical world.
Learning
Vision-Language-Action (VLA) policies
Imitation learning from teleop demos
RL fine-tuning
Sim-to-real transfer
Hardware
Arm mechanics & servo actuation
Sensors & cameras
Power electronics
Wiring & PCB prototyping
Embedded
Microcontroller firmware
Motor control loops
On-device inference
ROS 2 integration
06Learning
Always learning
Degrees, certificates and courses on one timeline. I'm always building new skills; scroll to follow the path, hover a card for details.
Validated H.AI (Humanized AI), an agentic co-planner for railway electrical engineering, through structured customer discovery with practising planners; MVP and business-model fundamentals.
