Open Source Summit North America 2025

As Generative AI unfolds its transformative potential, and enterprise needs have shifted toward large-scale, distributed deployments, the requirements for inference serving have fundamentally changed. To meet these new demands, NVIDIA has introduced Dynamo— a high-throughput low-latency inference framework designed for serving generative AI and reasoning models in multi-node distributed environments.

This session provides a technical overview of Dynamo’s architecture, focusing on how its design addresses the core challenges of large-scale, distributed generative AI inference. We will walk through concrete deployment scenarios—including disaggregated serving and dynamic GPU scheduling—and examine how Dynamo manages resource allocation, request routing, and memory efficiency for high-throughput, low-latency inference.

We will also share practical implementation examples and discuss engineering best practices for optimizing workload performance, scalability, and cost using Dynamo. We’ll outline the steps and considerations for deploying Dynamo, highlighting key architectural differences and compatibility factors. By the end of the session, attendees will have a clear understanding of how to deploy and operate Dynamo in production environments to support advanced AI workloads.

Generative AI applications, in general, excel in zero-shot and one-shot types of specific tasks. However, we live in a complicated world and we are beginning to see that today’s generative AI systems are simply not well equipped to handle the increased complexity that is found especially in business workflows and transactions. Traditional architectures often fall short in handling the dynamic nature and real-time requirements of these systems. We will also need a way to coordinate multiple components to generate coherent and contextually relevant outputs. Event-driven architectures and multi-agent systems offer a promising solution by enabling real-time processing, decentralized decision-making, and enhanced adaptability.

This presentation proposes an in-depth exploration of how event-driven architectures and multi-agent systems can be leveraged to design and implement complex workflows in generative AI. By combining the real-time responsiveness of event-driven systems with the collaborative intelligence of multi-agent architectures, we can create highly adaptive, efficient, and scalable AI systems. This presentation will delve into the theoretical and practical sides.

Join us to discover how Intuit's GenAI framework is reshaping AI development, enabling swift integration of AI functions across varied business units. We'll focus on a robust framework of reusable agents and tools derived from open-source technologies like LangChain/LangGraph, facilitating diverse functionalities from simple data retrieval to complex processes such as query generation, optimization, pipeline creation and debugging. This framework dramatically reduces the time required for data workers to operationalize data pipelines and supports diverse customer interactions through notebooks, no-code approaches, REST integrations, and Python libraries, catering to a wide range of needs including agent developers and teams in pre-production settings. Our meticulous evaluation process ensures that each tool and agent is rigorously tested against high-performance benchmarks to guarantee reliability and consistency before deployment. By centralizing these AI components, Intuit has not only accelerated development timelines but also upheld a high standard of quality, establishing a benchmark for crafting scalable, effective AI solutions in the dynamically evolving tech landscape.