We live in an era where access to cloud resources for training and deploying AI models is just a few clicks away. However, there are certain scenarios in which access to these models needs to be restricted, or the devices where these models run have limited resources. In this talk, we will explore different strategies for integrating AI capabilities into resource-constrained scenarios, such as edge devices, offline systems, and mobile applications.

We will start by exploring some of the scenarios where offline or restricted access is essential, like remote areas with limited connectivity or applications where privacy is crucial like those used in healthcare and finance. We will then explore the process of fine-tuning base models like Phi-3 to suit specific use cases, leveraging different optimization techniques and we will finish by showing how to use tools and frameworks for efficient deployment on mobile platforms (like the ONNX Runtime).

At the end of the talk, participants will be equipped with the knowledge needed to overcome the challenges of deploying AI models offline and bring cutting-edge AI to mobile and embedded devices.