Prezi leans on Hugging Face experts to sharpen its AI presentation pipeline

Prezi, the online presentation platform, has been quietly reworking the machine learning that powers its flagship AI product. The company joined Hugging Face’s Expert Support Program to get hands-on guidance, and the collaboration has already changed how Prezi thinks about model selection. Máté Börcsök, a backend engineer at Prezi, explained that the team already had a working version of Prezi AI when they signed up. The system takes a user’s prompt and builds a draft presentation by coordinating closed models and various asset providers. But their Hugging Face expert reviewed the entire flow and suggested concrete improvements—most notably, adding an open-source re-ranker model to their search pipeline. That single change helps Prezi find better images and text for each presentation ‘cheaper, faster, and better than an LLM,’ as Börcsök put it.

What makes Prezi’s use case tricky is that presentations are inherently multimodal. They combine images and text, which means the team juggles vision models, text models, and vision-language models simultaneously. Börcsök was candid about the challenge: ‘We are not Machine Learning Engineers, and we are learning this together on the way.’ That’s precisely where the expert support pays off. With new models dropping every week, having someone who can cut through the hype and flag which releases are actually useful saves the team significant time. The guidance extends to fine-tuning VLMs and curating training data—work the team could technically do themselves, but would take far longer without direction.

Börcsök also highlighted Inference Endpoints as a standout feature. The Endpoint Model Catalog offers curated models that require zero configuration, which removes a huge amount of friction. He specifically praised the sleep functionality that shuts down endpoints after a few minutes of inactivity, meaning Prezi isn’t burning money on idle GPUs. Support for single and quad A100 instances covers models that need serious compute, and deployment is refreshingly simple—one click for the latest version, or roll back to any older version using a Git hash. His comparison was blunt: ‘None of these features are easily available on AWS.’

For teams without dedicated ML engineers, this kind of program is effectively a shortcut to production-grade machine learning. Prezi got a review of their existing architecture, practical advice on model selection, and ongoing feedback on everything from embeddings to object detection. The Inblix angle here is that Hugging Face is increasingly positioning itself not just as a model hub, but as a consulting layer for companies that want to ship AI without building an in-house research team. Whether that model scales remains an open question, but for Prezi, the speedup is already tangible.