I build and validate machine-learning pipelines end to end; from data acquisition and signal preprocessing to modeling and rigorous evaluation. I work mainly with speech and physiological signals, with a recent focus on automated evaluation for generative, LLM-driven speech systems.
I'm a Machine Learning Engineer with a PhD from EPFL and 7+ years building deep learning and time-series ML pipelines end to end; from data acquisition and signal preprocessing to modeling and rigorous, often automated, evaluation. Most of my work is with speech and physiological signals, and I've recently focused on evaluation and tooling for generative, LLM-driven speech systems. I work in Python with PyTorch, TensorFlow, and scikit-learn, and build reproducible pipelines with Docker, REST APIs, and automated tests. Along the way I've collected specialized speech, physiological, and face/voice datasets, mentored Master's students, and reviewed for IEEE/ACM Transactions on Audio, Speech, and Language Processing.
Benchmarked 9 text-to-speech voices across 15 domains into a single quality score from intelligibility, audio quality, speaker similarity, and synthesis success, cutting reliance on manual listening tests. Validated against a human ranking study, with automatic and human rankings agreeing on the top-3 and bottom-2 voices.
Developed deep-learning models that estimate breathing signals directly from raw speech waveforms in TensorFlow, achieving a 25% reduction in input lag compared with spectral-based models.
Developed a distance-bin detection method that reduced MSE by ~50% versus a fixed-bin baseline, then fed the detected bins into an existing ML model for vital-sign estimation, cutting downstream errors by ~31% (breathing rate) and ~87% (heart rate).
Developed models for COVID-19 detection, emotion recognition, heart-rate estimation, and hypoglycemia detection in diabetic patients, using speech signals with both data-driven and hand-crafted features in PyTorch and scikit-learn.
Investigated presentation attacks on CNN-based face-recognition systems, including makeup-induced spoofing, and collected and curated two large face-biometric datasets from 100+ participants, including acquisition protocol design and attack generation.
At the Lemanic Life Science Hackathon at EPFL, worked with a multidisciplinary team on a prototype to help clinicians assess children's reading. Selected WavLM base+ for phoneme recognition on children's speech and built the initial pipeline; the team reached a working prototype in three days.
Full list – 15+ publications, ~700 citations, h-index 9 – on Google Scholar →
I'm open to Machine Learning Engineer, Applied Scientist, and Research Engineer roles in the Zürich area. The fastest way to reach me is email.