David vonThenen

AI can now listen to your keyboard and guess what you're typing. This session shows how deep learning models can reconstruct text from keystroke sounds, then breaks down how these attacks work and how to defend against them. It's a live, hands-on look at the thin line between innovation and exploitation in modern AI security. Bring your curiosity and maybe a little paranoia.

This session walks through how to build and train your own Small Language Model using open source tools. You'll learn how to select datasets, reuse BERT's tokenizer, design a smaller student model, and apply knowledge distillation to preserve accuracy. We'll share code, show common pitfalls, and help you deploy an efficient, real-world SLM that runs fast on modest hardware.

This session breaks down the training process of multi-modal machine learning models for real-time detection of debilitating diseases using video and audio data. Attendees will learn how to build and integrate video and audio classifiers, curate multi-modal datasets, and deploy these models in real-world settings. The session includes practical demonstrations and code samples to help participants implement their own multi-modal classification models, equipping them with tools and methodologies to apply in the healthcare industry or other fields requiring complex real-time detection.

This session describes and simplifies the process of building NLP models by guiding attendees through creating a basic model, followed by a more advanced redaction model from scratch. The session offers practical insights into model development, including dataset preparation and model training, while providing code to build and implement the models discussed. By the end of the session, participants will walk away with a proven framework for approaching NLP problems and the tools to apply these concepts immediately in their work.

Discover the power of voice-assisted IoT and Edge interfaces in this practical session. Learn to create effective voice-activated devices, using LLMs like Falcon and OpenLLaMA. Gain insights into design decisions, essential components, and collaborative frameworks.

Creating and deploying Microservices is easy. The real problem is how to manage and support these services out in the wild and in production. What happens when these services stop working or worse yet when they are running but running slowly? Which service instance is the culprit? This session talks about how you can leverage Jaeger, OpenTracing, and Prometheus in order to give better visibility into the distributed nature of a Microservice architecture in a Kubernetes environment.

Persistent applications can be complex to manage and operate at scale, but tend to be perfect for modern schedulers such as Apache Mesos, Kubernetes, and etc. Internal direct attached storage and external storage are both options to run your applications. This talk outlines how 2 Layer Scheduling and Software Defined Storage enables deployment of managed frameworks and tasks, while maintaining high availability, scale-out growth, and automation.