I build the evaluation and data infrastructure that makes AI systems survive production.
I am joint first author of COMPL-AI, the first technical interpretation of the EU AI Act and an open benchmarking suite built on it. At LatticeFlow AI I built LLM evaluation infrastructure — 23 evaluators across 10+ models — that cut QA cycles from two days to four hours. At Omnilex I built multi-jurisdiction legal AI pipelines processing roughly one million documents across three countries. I hold an MSc in Machine Intelligence from ETH Zürich.
LLM evaluation
Data pipelines
Trustworthy AI
Python / TypeScript
AWS / Azure
Writing
I write about the layer around a model in production, and where it breaks.
What a misbehaving RGB controller taught me about designing tools for language models.
Experience
Data Engineer — Omnilex
Feb 2026 – May 2026 · Zürich, Switzerland
Built ingestion and transformation pipelines for legal content across 3 jurisdictions (Switzerland, Germany, Austria), processing ~1M documents from APIs, scraping outputs, and bulk sources.
Developed TypeScript-based data workflows for normalization, citation-aware chunking, embeddings, classification, and entity extraction.
Contributed to RAG-ready indexing and Azure-based search infrastructure for precise, traceable legal AI responses.
Introduced data contracts and validation checks that caught 50K+ duplicate entries before they reached production.
Machine Learning Engineer — LatticeFlow AI
Jan 2025 – Nov 2025 · Zürich, Switzerland
Built v0 evaluation infrastructure for a new AI product from scratch: 23 evaluators, integrations with multiple data and chat-model providers, assessing 10+ LLMs.
Automated evaluation workflows for QA and regression testing, reducing review cycles from ~2 days of manual inspection to ~4 hours.
Extended evaluators with targeted dataset generation to increase coverage across model behaviours and failure modes.
Research
COMPL-AI — Benchmarking LLM Compliance with the EU AI Act
Research Intern · Secure Reliable Intelligence Lab, ETH Zürich · Oct 2023 – Mar 2024
Joint first author (equal contribution) and lead author on COMPL-AI — the first technical interpretation of the EU AI Act, mapping its six ethical principles onto 27 concrete benchmarks and evaluating 12 prominent LLMs against them. Worked on benchmark design, evaluation pipelines, and model integration via Hugging Face Transformers. The framework is positioned as a reference point for the EU’s GPAI Code of Practice.
Give a system a corpus from a company it knows nothing about, and no schema. Agents propose changes to a shared knowledge state; a minimum-description-length energy and a Metropolis-Hastings rule decide what survives, so the ontology is annealed rather than fixed in advance. Thirteen reversible operators, each with its inverse asserted by round-trip tests.
Search over legal sources that differ in authority, jurisdiction and trust, built around one decision: canonical truth is separated from the serving projection, so the index is a rebuildable view rather than the system of record. Thirteen services, moved from Google Cloud onto self-hosted Kubernetes, with 59 architecture decision records and two dozen automated checkers holding the boundaries in place.
MSc Computer Science — Machine Intelligence, ETH Zürich
Sept 2022 – Dec 2024
Thesis (top grade): Speech Recognition for Children with Congenital Disorders Using Adaptive Methods — adapting Whisper to non-normative child speech from a single speaker.
