--:--

Mauro
Luzzatto

I am a Machine Learning Engineer at Epidemic Sound, where I train, evaluate, and use latent diffusion models for audio. I focus on audio generation models that help creators customize music for their videos.

→ github.com/MauroLuzzatto → linkedin.com/in/mauro-luzzatto → mauro.luzzatto@proton.me
01Now

Currently shipping Adapt, Epidemic Sound’s AI music editing tool

Role Machine Learning Engineer, Generative AI
Tenure At Epidemic Sound since Feb '22
Base Stockholm since 2021
From Switzerland

I work on Adapt, Epidemic Sound’s AI music editing tool. Day-to-day, that includes training latent diffusion transformers, building data infrastructure to train them at scale, and prompt engineering to get the most out of the models.

Before this, I joined Epidemic Sound as a Data Engineer, building the data platform and migrating event-tracking infrastructure to Snowplow.

Previously, I worked as a Data Scientist using NLP and machine learning at IBM in Zurich, and conducted research in legal NLP at ETH Zurich.

I studied Computer Science and Energy Science at ETH Zurich, with an exchange semester at KTH in Stockholm.

02Focus

Things I think about most days

A.

Generative AI for audio — and how to train, evaluate and scale the generative models.

Diffusion · Data · Evals
B.

Music editing — fine-tuning generative models to follow edit instructions.

Editing · Fine-tuning · Instruction-tuning
C.

ML infrastructure — data pipelines, training optimization, multi-GPU workflows.

Scale · GPU · Pipelines
D.

Software craft — testing, docs, clean code, observability.

Engineering · Quality
03Selected Work

Open-source and side projects

01 / 02

explainy

Python scikit-learn SHAP GH Actions

A library for generating ML model explanations in Python — feature-importance for individual samples through a small, scikit-learn-flavored API. Four algorithms, global or local, contrastive or not.

→ Repo Docs Paper
02 / 02

Entropy in Legal Language

word2vec Gensim NLP

A method for measuring word ambiguity in legal corpora using word2vec embeddings — applied to opinions of the U.S. Supreme Court and the German Bundesgerichtshof, comparing common-law and civil-law systems.

→ Repo
04Research

Papers, supervision, and slower work

2026 Epidemic Sound AI Research Blog: When music generation models get mixed signals — investigating conditioning conflicts in Adapt Epidemic Sound
2026 Supervising a KTH Master’s thesis — instruction-tuning a latent music diffusion model, extending SAO-Instruct to music. Ongoing. KTH
2025 explainy — a companion paper on generating explainable Machine Learning predictions in Python ACM
2021 Entropy in Legal Language — measuring entropy in U.S. Supreme Court and German Bundesgerichtshof opinions with word embeddings ETH
05Stack

Tools I reach for, most days

Machine Learning
Python PyTorch Hugging Face Diffusers scikit-learn spaCy
Generative AI
Latent Audio Diffusion LoRAs DiT VAEs Google ADK Gemini API
MLOps & Infra
GCP AWS Dataflow / Beam Modal Docker Terraform GitHub Actions Weights & Biases
Data
BigQuery dbt Airflow Snowplow SQL
Web & Visualisations
FastAPI Streamlit Gradio htmx TailwindCSS
Certifications
Google Cloud Professional Data Engineer AWS Solutions Architect — Associate AWS Machine Learning — Specialty Azure Architect Technologies (AZ-300)