Prezi Case Study: Accelerating ML Roadmap with Hugging Face Expert Support

Prezi leverages Hugging Face to accelerate multimodal AI development

Prezi has integrated the Hugging Face Expert Support Program and Inference Endpoints to accelerate its machine learning roadmap, specifically for its Prezi AI product. By transitioning toward smaller, more efficient open-source models and utilizing expert guidance, Prezi is optimizing the creation of dynamic presentations that combine text and images.

Optimizing Prezi AI with Open-Source Models

Prezi AI allows users to generate draft presentations from a prompt and description. While the system initially relied on closed models and various asset provider services, the integration of Hugging Face's expert guidance has led to several technical optimizations:

  • Implementation of Re-rankers: To improve the search system used to find suitable images and text for presentations, Prezi added an open-source re-ranker model. This approach is faster, cheaper, and more effective than using a Large Language Model (LLM) for the same task.
  • Multimodal Integration: Prezi utilizes a combination of vision models, text models, and vision-language models (VLMs) to handle the inherently multimodal nature of visual communications.
  • Model Selection: The Expert Support Program helps Prezi's engineering team—who are not specialized ML engineers—filter through the high volume of new model releases to identify those that provide actual utility over hype.

Deployment and Infrastructure via Inference Endpoints

Prezi utilizes Hugging Face Inference Endpoints to manage the deployment of its ML models. Key technical advantages cited include:

  • Zero-Configuration Deployment: The Endpoint Model Catalog provides a curated list of models that work with Inference Endpoints without requiring manual configuration.
  • Cost Management: The ability to set endpoints to "sleep" after a few minutes of inactivity prevents unnecessary spend.
  • Hardware Support: The platform supports single and quad A100 instances, which are necessary for larger models.
  • Version Control: Prezi can deploy the latest model versions with a single click or roll back to previous versions using Git hashes, a workflow described as more convenient than similar processes on AWS.

Impact of the Expert Support Program

For teams without dedicated Machine Learning Engineers, the Expert Support Program provides a bridge to production-ready AI. Prezi uses this partnership to access best practices and state-of-the-art models for specific tasks, including:

  • Embedding and Re-ranking: Optimizing how assets are retrieved.
  • Object Detection: Improving visual asset management.
  • Fine-tuning: Learning how to fine-tune new vision-language models and curate the necessary data.

As Máté Börcsök, a backend engineer at Prezi, notes:

"Our expert guides us to work on the right things, sharing best practices and state-of-the-art models for embedding, re-ranking, and object detection and showing us how to fine-tune new vision language models and collect and curate data."

Sources