ML Engineer in Berlin, around 8 years of experience. I build production ML systems: LLM/RAG agents, MLOps platforms, recommenders and large-scale NLP.
I like the messy end-to-end of ML systems: retrieval that actually works in production, platforms that let a team ship without waking up at 3am, agent loops that adapt instead of running through a fixed plan. My recent focus is on agentic architectures and retrieval-augmented systems for domains where correctness matters.
Most recently at Pure App, where I:
- Built a RAG agent on LangChain, LangGraph, Weaviate, and Langfuse
- Built the MLOps platform from scratch on Kubernetes and AWS, with 4Airflow, MLflow, and Prometheus/Grafana
- Built a PyTorch GNN-based recommender
Before that: Senior Data Scientist at Upday (multilingual NLP at scale) and Data Scientist at Bayes Esports Solutions (real-time outcome prediction).
An adaptive plan-and-solve agent that answers legal research questions about the EU AI Act. Given a query, it plans a research strategy, executes it against a retrieval layer over the official regulation text, and revises that plan based on what it finds, by chasing references, reformulating queries that come back empty or finishing early when it has enough. Every claim in the final answer is cited to a specific provision and linked to the official EUR-Lex source. Built with LangGraph and traced end-to-end with Langfuse.
MSc in Electrical Engineering (Summa Cum Laude), Universidad de Chile. Research with publications in Nature Astronomy and IEEE from earlier astroinformatics research.
- P. Huentelemu et al., Correntropy based filtering for supernova detection, 2016 International Joint Conference on Neural Networks (IJCNN), Vancouver, BC, 2016, pp. 3322-3329
- F. Förster et al., The delay of shock breakout due to circumstellar material evident in most type II supernovae, Nature Astronomy, 2397-3366, 2018