Julien Herzen


I'm an engineering lead at Isomorphic Labs, a sister company of DeepMind, where we are building on innovations such as AlphaFold to reinvent drug discovery using computational and AI-first approaches.

Before Isomorphic Labs, I was the lead data scientist at Unit8, working on democratising data engineering, data science and machine learning in the industry. I was a part of Unit8 leadership team, contributing to growing the company from 6 to 110 people in about 4 years (we were even recognized as the fastest growing Swiss company in 2022 by the Financial Times!).

Before that, I was a lead data scientist at Swisscom, a PhD student at EPFL, and a researcher in the Silicon Valley.

I have experience designing machine learning models and systems, and writing production-grade software in different contexts (see some projects below). My PhD thesis was nominated by the jury for the EPFL Patrick Denantes Memorial Prize. I wrote a few academic papers (some of which got best paper awards), and have a couple of (granted) patents.

Some Projects

Darts: Time Series Made Easy in Python

Darts: Time Series Made Easy in Python

For a long time, the time series forecasting experience in Python was not really great. I created Darts github stars as an attempt to provide a unified and user-friendly API for time series forecasting (and more). The library puts a strong emphasis on modern machine learning techniques. Among other things, it offers the possibility to easily train models (included some of the latest and coolest competition-winning algorithms) on large time series datasets, and obtain probabilistic forecasts. I'm leading the development and contributing regularly, along with a great team of contributors.

pip install darts

Darts on Github JMLR Paper Introductory Medium article

VeGANs: Various existing GANs (and Other Generative Models) in PyTorch

When working and experimenting with GANs, I grew a little frustrated by the fact that most open-source implementations of research papers could only be used to reproduce that paper's result in a limited context. So I created VeGANs github stars, a small library making it easier to use many different sorts of GANs (and other generative models) with PyTorch. The idea is to only provide a Generator and Discriminator PyTorch modules, and let VeGANs train them for you with a GAN algorithm of your choice. The library was then revamped and vastly improved by Thomas Neuer, a Unit8 colleague.

pip install vegans

VeGANs on Github
Digits generated conditionally using VAEGAN
Number of trips between Swiss Cantons

Country-scale Mobility Mining Using Telco Data

At Swisscom, I was product owner and tech lead for the Mobility Insights Platform. We were transforming raw network data coming from cell towers (~ 2 Mio data points / second) into intelligible statistics about mobility in the country. I led the development and architecture of the platform, interfaced with business, and developed some key algorithms (such as for positioning and inferring paths over road and railway networks). For some time I was also in charge of the content of the Swisscom open data portal, and started a project to spot anomalies in (massive) network time series.

Swisscom Mobility Insights

Mining Democracy

During my time at EPFL, I started a little side project whose goal was to spot patterns in voting advice applications data (such as Smartvote) and open government data. Among other things, we discovered that the Swiss Röstigraben can be quantified, that some votes' results can be predicted with very little information, and that politicians do not always optimally cover the "ideological space" of their constituents. Since then, the project has been significantly strengthened and improved by others, and more research has been done on the topic of vote prediction.

Paper (ACM COSN 2014 - Best paper award) Predikon website
Embeddings of Smartvote answers (politicians)

Dynamic Spectrum Allocation

Distributed Resource Allocation for Wireless Networks

Over the course of my PhD, I invented, implemented and analysed new distributed algorithms for resource allocation in computer and wireless networks. Among other things, I worked on new algorithms that can reach globally optimal configurations (in some probabilistic sense) of spectrum usage, transmit durations and transmit power using only local information about their surroundings. I also used machine learning to predict the performance of wireless networks in new (unseen) configurations.

PhD thesis A nice paper (IEEE ICNP 2015) Another pretty good paper (IEEE SECON 2015) Another nice paper (IEEE Infocom 2013)

Graph Embedding for Scalable Routing

I proposed a new algorithm for obtaining graph embeddings in a distributed fashion. The embedding algorithm relies on spanning trees, it uses a custom metric space (with $l_{\infty}$-norm) and it can be used for geometric routing in computer networks. It requires low memory (polylogarithmic in the network size), making it scalable for Internet-scale graphs.

Paper (IEEE ICNP 2011) Slides Technical Report
Greedy Graph Embedding