Daniel Lutziger

Machine Learning Engineer

I train deep learning models on GPU clusters and have built production systems at UBS for 13 years.

Profile

Research-grade deep learning: pretrained a multimodal Transformer foundation model in PyTorch on ~121,000 hours of time series from ~13,000 people, on an NVIDIA GPU cluster (A100, H100, RTX 4090). Against training from scratch, it more than doubled balanced accuracy on a fully held-out dataset.
13 years of production engineering at UBS: today tracing data lineage for a finance platform and building self-service tools for the migration to a new one.
LLMs in practice: RAG prototype on a locally hosted LLM at UBS, a fine-tuned LLM classifier, and a knowledge-graph chatbot with a local Llama model.
MSc Data Science (magna cum laude), University of Zurich; co-author at ACM RecSys 2025.

Master thesis project

June 2025 – December 2025

Built a multimodal set-then-sequence Transformer encoder in PyTorch for polysomnography (EEG, EOG, EMG, ECG, respiration). Channels are encoded as an unordered set, so the model handles the incomplete montages found across clinical sites.
Built preprocessing pipelines harmonizing ten heterogeneous public cohorts (~13,000 subjects, ~121,000 recording hours): resampling, channel normalization, modality grouping, subject-wise splits to prevent leakage.
Pretrained with contrastive objectives (LOOC + SimCLR, InfoNCE) and compared frozen-encoder transfer against training from scratch on a fully held-out cohort (SHHS, 5,793 subjects): 76.2% vs. 37.2% balanced accuracy for sleep staging, 75.4% vs. 60.4% for apnea event classification.
Ran ablations on pretraining data scale, modality subsets and imbalance losses (focal loss, class-weighted CE), and reported where pretraining did not help (night-level AHI severity).

Tech stack: Python, PyTorch, Hugging Face Transformers, NumPy, pandas, Matplotlib, Linux HPC cluster with NVIDIA GPUs (A100, H100, RTX 4090)

August 2013 – Present (apprenticeship full-time 2013–2017, part-time 2018–2025 alongside studies, full-time from January 2026)

Product Owner & Site Reliability Engineer (since December 2023; full-time since January 2026)

Data lineage for the finance platform: Tracing data from outbound feeds back through each transformation to the master tables, and visualizing the lineage as diagrams in a dashboard so users can follow the workflow.
New finance platform (self-service file transfer): In a fast-paced team that builds heavily with AI coding tools, built a self-service application that lets users find files in the current finance system with filters and fast search, and copy them directly into Azure storage for the new platform.
LLM proof of concept:Built a Retrieval-Augmented Generation prototype for internal knowledge management on a locally hosted LLM with a Chroma vector store. This laid the technical groundwork for the department's AI plans.

Tech DevOps Engineer (2018 – December 2023, part-time alongside studies)

Deployment strategy: Developed the Ansible-based deployment strategy for a wealth management finance platform, rolling out code to multiple instances across locations.
Workflow automation (Flowable PoC): Built a Flowable proof of concept as a possible replacement for the Redwood RunMyJobs job control system; my production cost estimate was confirmed three months later by an independent external estimate.
Internal tooling (DevOps Dashboard): Led the Spring/React rewrite of an outdated tool for managing day-end-processing scenarios and environments (multi-task handling, attachments, automated stakeholder emails), then merged it into a DevOps dashboard with consolidated SSO and a unified database, reducing infrastructure costs.

Apprentice Software Engineer (2013–2017, full-time)

iOS app (Swift) for UHNW clients, Angular migration of a client-advisor web app, live chat for ubs.com.

Tech stack: Java, Spring, React, Vite, Angular, Swift, jQuery, Python, Ansible, Linux, Flowable

October 2022 – December 2024

Built the “Groundhog Alpha Telemetry Application” in React (web) and React Native (mobile) for real-time sensor data visualization and live monitoring of a tunnel boring prototype. Our ETH Zurich student team placed 2nd at the Not-a-Boring Competition 2023 and won it in 2024.

Tech stack: React, React Native, JavaScript/TypeScript

Blooming Beats (UZH, 2025, lead developer & designer): Explainable, content-based music recommender built on 219,673 plays from ten years of listening history; published at ACM RecSys 2025, Best Pictorial at CHItaly 2025. Next.js/React, D3, FastAPI.
RDF Movie Chatbot (UZH, 2024, team of 2): Conversational agent that answers movie questions from an RDF knowledge graph with no external AI APIs. Flair NER for entity extraction, fuzzy matching against the graph, templated SPARQL queries for facts, embedding nearest-neighbour search for recommendations, and a local Llama 3.2 1B used only to phrase verified results. Crowd-sourced fallback answers weighted by Fleiss' κ. Dockerized.
Ski IMU Logger (personal, ongoing): Boot-mounted RP2350 + BNO055 sensor with MicroPython firmware logging 9-axis motion data at 50 Hz. Working prototype; data collection this ski season, modeling once labelled data exists.

Major: Data Science · Minor: Information Systems

August 2022 – December 2025 (completed while working part-time)

Graduated magna cum laude (5.4/6.0).
Thesis (5.75): “Towards Scalable Foundation Models in Sleep EEG and Polysomnography” (Idiap Research Institute; supervised by Prof. Dr. Manuel Günther, Dr. André Anjos, Dr. Olivier Pallanca). See Idiap entry above.

August 2018 – July 2022 (completed while working part-time) · Elective focus: Data Science and Analysis

Grade 5.07/6.0; exchange at the University of Exeter, UK (2020–2021).
Thesis (5.5): NLP pipeline over ~25,000 news articles with a fine-tuned LLM classifier and a Neo4j knowledge graph.

I. Al-Hazwani, D. Lutziger, C. Kirchdorfer, L. Huber, O. R. Aschwanden, J. Bernard, L. Boratto. “Blooming Beats: An Interactive Music Recommender System Grounded in TRACE Principles and Data Humanism.” ACM RecSys 2025 (Demo Track), Prague. DOI: 10.1145/3705328.3759337
Best Pictorial Award – CHItaly 2025 (extended version of the above work).
Best Project Award – Impact on Research – Interactive-Visual Data Analysis course, University of Zurich (2022).
Abschluss im Rang, Zürcher Lehrbetriebsverband ICT (top 10% of ICT apprentices, canton of Zurich, 2017).

ML / Deep Learning:PyTorch, Hugging Face Transformers, scikit-learn, self-supervised & contrastive learning, transfer learning, time-series biosignals
Generative AI: RAG with Chroma, LLM fine-tuning, local LLMs (Llama, Qwen), Claude API, OpenAI API, embeddings and similarity search
MLOps & Engineering:Weights & Biases, NVIDIA GPU cluster (A100, H100, RTX 4090), Docker, Git, CI/CD, Ansible, Linux, Azure storage, Python, SQL, Java/Spring, React

German (C2, native) · English (C1) · French (B1) · Italian (A2)