InkSight
Offline-to-online handwriting conversion (TMLR): trained PaLI-style vision-language models up to 5B parameters on 64 TPUs (JAX, T5X) using unpaired data, achieving 87% human-validated stroke conversions.
ML engineer working between data and modeling — LLM post-training, AI agents, and evaluations at scale.
I'm a Software Engineer at Google in Zurich (YouTube Trust & Safety), where I post-train frontier LLMs for policy-violation detection and build AI agents improved with automatic prompt optimization.
Previously, I worked as an Applied Scientist at 2K Games on real-time ML matchmaking for NBA 2K, as an SDE at Amazon Web Services on distributed ML anomaly detection in CloudWatch, and conducted research and engineering internships at NII Tokyo, Google Zurich, and LinkedIn. I hold an M.Sc. in Data Science from EPFL and a B.Sc. in Computer Science from RWTH Aachen.
Research code, benchmarks & open-source tools
Offline-to-online handwriting conversion (TMLR): trained PaLI-style vision-language models up to 5B parameters on 64 TPUs (JAX, T5X) using unpaired data, achieving 87% human-validated stroke conversions.
Evaluation dataset of 1,118 human-verified multi-hop questions extending existing QA benchmarks with commonsense, arithmetic, and symbolic reasoning—benchmarked across models up to 70B parameters.
Modular conceptual framework and open-source library developed at EPFL for composing, debugging, and scaling structured interactions between autonomous AI agents and humans.
A fast CLI utility built with Lune that transforms monolithic Roblox place files (.rbxl / .rbxlx) into clean, diffable directory trees tailored for Git version control and collaborative workflows.
Selected preprints and peer-reviewed work
Blagoj Mitrevski*, Arina Rak*, Julian Schnitzler*, Chengkun Li*, Andrii Maksai, Jesse Berent, Claudiu Musat
* Equal contribution (random author order)
Julian Schnitzler*, Xanh Ho*, Jiahao Huang*, Florian Boudin, Saku Sugawara, Akiko Aizawa
* Equal contribution
Martin Josifoski, Lars Klein, Maxime Peyrard, Nicolas Baldwin, Yifei Li, Saibo Geng, Julian Paul Schnitzler, Yuxing Yao, Jiheng Wei, Debjit Paul, Robert West
Industry, research labs & academic background
Zurich, Switzerland
Post-training Google's frontier LLMs for YouTube policy-violation detection and building an AI video-review agent (prompt, tools, logic) improved with automatic prompt optimization (APO).
Dublin, Ireland
Built ML-based matchmaking and skill rating for NBA 2K matching millions of players in real time; analyzed large-scale player data and prototyped new modeling approaches.
Dublin, Ireland
Built and maintained a distributed ML anomaly detection system for customer time-series metrics in CloudWatch Alarms (Java, Python); added 2 time-series models driving 22% YoY adoption growth at 99.999% uptime.
Tokyo, Japan
Created MoreHopQA (co-first author, 1,118 human-verified multi-hop questions); built dataset pipelines and evaluated/fine-tuned up to 70B-parameter models on multi-GPU clusters.
Zurich, Switzerland
Co-authored InkSight (TMLR); trained PaLI-style vision-language models up to 5B parameters on 64 TPUs (JAX, T5X) using unpaired training data and an OCR-based evaluation pipeline.
Dublin, Ireland
Built an automated evaluation system over datasets up to 100B rows (Spark, Scala), cutting iteration time from days to under an hour; owned prompt design for LinkedIn Premium's GPT-4 Writing Assistant.
Lausanne, Switzerland
Focus on machine learning, NLP, and large-scale data science; contributed to the Flows collaborative AI framework.
Aachen, Germany
Foundations in computer science, algorithms, and process mining (including open-source contributions to PM4Py).
Languages
Data & Pipelines
Deep Learning & LLMs
Research collaborations, engineering & inquiries
Whether you want to discuss machine learning research, open-source tooling, or engineering projects, feel free to reach out via email or connect on LinkedIn, X, or GitHub.