Lewis Tunstall took an unusual path to Hugging Face: he cold-emailed the company’s co-founder with a book proposal and got a response. That book, NLP with Transformers, co-authored with Leandro von Werra, became the seed for both of them joining the company. Now nine months into his role as a machine learning engineer, Tunstall is focused on a problem that bedevils practitioners far more than training ever does: deployment.
The core issue he’s tackling is one anyone who’s tried to ship an NLP model has hit. Modern transformer models are enormous, and that parameter count translates directly into latency. “If you’re trying to build a chatbot you need this model to be very fast and responsive,” Tunstall explains. “Most of the time these models are a bit too slow if you just take an off-the-shelf model, train it, and then try to integrate it into your application.” His recent work on the transformers library centers on export functionality that lets developers move models into more efficient formats—leveraging both Hugging Face’s own tooling and the broader open-source ecosystem.
It’s a pragmatic pivot from the theoretical physics background Tunstall came from, and it reflects where the field is actually hurting. Training gets all the headlines, but inference latency and throughput are what determine whether a model ever reaches real users. The gap between a model that works in a notebook and one that responds in milliseconds inside an application is where ML engineering lives. Tunstall’s focus on that gap—rather than on pushing architecture benchmarks—says something about where Hugging Face sees its mature-phase value proposition.
His origin story also highlights something underappreciated about how open-source communities actually operate. A cold email, a genuinely useful contribution idea, and a receptive founder were enough to launch both a book and two careers. The barrier to entry in this ecosystem remains surprisingly low for people willing to build in public.
