WHERE DATA MEETS INTELLIGENCE

As enterprises increasingly leverage vast public and private datasets, generative AI and agentic systems are transforming the landscape of AI-driven solutions. These systems demand unparalleled scalability, speed, and efficiency to process massive data volumes while autonomously orchestrating complex workflows. SambaNova Systems offers its revolutionary memory-centric design, engineered to power trillion-parameter models and multi-agent systems with record-breaking interactive inference performance.

This talk will delve into SambaNova’s innovative three-tier memory system and reconfigurable dataflow architecture, which overcome the "memory wall" challenge by enabling seamless switching between hundreds of agents in microseconds. Attendees will explore how these technologies optimize data access, minimize latency, and scale across diverse real-world applications—from real-time decision-making to autonomous multi-agent collaboration—delivering transformative solutions for enterprises worldwide.

Modern data lakes promise affordability and scalability, but using them can be a headache. Cloud data warehouses make querying easy, but they come with a hefty price tag and extra complexity. What if you could get the same ease of use without the cost and lock-in?

In this session, we’ll show you how to leverage open-source software to build a fully functional, queryable analytics powerhouse using DuckDB, Fivetran, and Polaris Catalog. We’ll walk through how to:

1. Load data that is automatically converted to Iceberg open table format

2. Run SQL queries using DuckDB’s new Iceberg extension

3. Run transformations directly on data stored in your data lake with a new dbt adapter

4. Get started easily with a practical, hands-on demo

No vendor lock-in, no unnecessary complexity—just an open-source-powered approach to enabling advanced analytics and AI. If your data warehouse is holding you back or eating away at your budget, this session is for you!

Modern data workloads demand fast, interactive, and scalable visualization—without the cost and complexity of server-side rendering. The local-first approach leverages modern browser capabilities, WebAssembly, and in-browser computation to achieve high-performance analytics while reducing cloud costs.

In this workshop, we’ll explore:

1. Why Local-First? The benefits of running everything client-side for cost-efficient, scalable visualization across thousands of users.
WebAssembly (WASM) for Data Apps: How Perspective harnesses WASM to power ultra-fast, browser-native analytics and even replace traditional Docker-based containers for data workloads.

2. Perspective + DuckDB: A full in-browser analytics stack that enables high-speed querying and visualization without a backend.
Streaming Data with InfluxDB: How to visualize high-frequency, real-time IoT and log data with sub-second latency.
Databricks + Perspective: Enhancing large-scale analytics with interactive dashboards inside Jupyter notebooks.
Through live coding and guided exercises, attendees will build their own browser-native analytics dashboards, connect to real-time data streams, and learn Perspective’s API in Python, Node.js, and Rust.

Difficulty level - Intermediate – Some experience with Python, JavaScript, and data analytics will be helpful, but beginners can follow along with guided exercises.

To participate, bring a laptop with: 
Git
VS Code
Docker
Python (3.8+)
Node.js (16+)

Not all AI agent use cases are created equal. While code generation agents can be tested against clear benchmarks, operational agents tackling real-world problems face a fundamentally different challenge: how do you evaluate an agent that must navigate complex, dynamic systems without a predefined playbook? Take root cause analysis in distributed systems: an agent must understand intricate service dependencies, parse through inconsistent logs, and reason about potential failure modes. Unlike coding tasks with definitive right answers, these scenarios have no ground truth. Traditional testing approaches break down completely. This talk breaks down our approach to building a deterministic simulation environment that generates and tests realistic failure scenarios at scale. We'll expose why existing evaluation methods fail—from infrastructure mimicry to LLM-generated tests—and demonstrate a lightweight simulation technique that enables precise, reproducible agent testing.

The modern data landscape is dominated by complexity: tables, schemas, pipelines, warehouses, and more. Yet the most successful data platforms share a common principle—they make data itself invisible to the end user. When data infrastructure functions optimally, it's like good plumbing: you only notice it when something breaks. Organizations often fixate on the mechanics of data while losing sight of what truly matters: metrics, dimensions, and semantics. When users engage with meaningful abstractions rather than technical details, they make better decisions faster. In this talk, you'll learn strategies for making data invisible through real-world abstraction success stories, designing effortless interactions, and implementing governance through abstraction. Walk away with practical ways to assess your data stack, advocate for user-centric approaches, and measure progress—making your data platform not just powerful, but invisible in all the right ways.

At the foundation of AI project failures lies a critical gap between data teams and business reality. On top of this gap, data quality issues, unexpected privacy concerns, and tools that don't align with actual business problems arise to hinder or block implementation. As we've built our own AI product—AI Decisioning—and implemented it with customers, we've learned that successful AI implementations depend on embedding data teams within business units. Embedding doesn't mean breaking apart your data team and dispersing it throughout every other department. It means establishing focused partnerships where data team members are deeply integrated into business teams' daily workflows and decision-making processes while remaining connected to the central data organization. This embedding creates a virtuous cycle: data teams gain deep domain knowledge, business professionals see improved data quality and gain data facility, and together, data and business teams implement AI solutions that solve real problems. In this talk, we'll share concrete examples of how data teams (especially data scientists) and marketing have worked together in successful AI Decisioning implementations. We’ll derive strategies to implement this organizational pattern and enable a company to move from analytics to actions and from data teams as service providers to active collaborators. While our case studies focus primarily on marketing partnerships, the embedded partnership model we present applies equally to other business functions including product development, operations, and customer service teams.

Building Cost-Effective LLM Routers: Boost Accuracy 25% While Cutting Costs 90% | This session reveals how to build intelligent model routers that dynamically direct inputs to the optimal large language model (LLM) for each specific task. Attendees will learn practical implementation strategies for multi-model LLM systems that significantly improve performance metrics—achieving up to 25% higher accuracy while reducing operational costs by as much as 90%. The presentation covers essential routing methodologies, evaluation frameworks, and scalable architectures for production deployments. Developers and ML engineers will gain actionable insights for overcoming technical challenges in multi-model LLM systems, optimizing both performance and cost-efficiency in generative AI applications. Perfect for teams looking to maximize ROI from their AI infrastructure while maintaining high-quality outputs.

The first two years of the GenAI revolution are bending the OSS way: Open Source models have reached state of the art, and most of the ecosystem around AI is open-source. The key to AI adoption is properly organizing and using business knowledge. In industry, LLMs give way to Small Specialized Models (SSMs), utilized by Domain Expert Agents (DXAs). Their work should be structured according to the domain requirements, requiring structured output. Organizing and using domain knowledge for AI has long been a domain of Knowledge Graphs (KG). At Neo4j, we are in a moment where our KG leadership powers the rise of GraphRAG, a better context traversal that we lead alongside Microsoft, Amazon, Google, and other GenAI partners. We also integrate with many OSS AI startups to build a better AI stack around GraphRAG. Neo4j has joined LFAI to bridge the enterprise AI adoption with startup innovation, centered around structured knowledge. In this talk we describe OAKS, a set of projects, communities, and technologies that comprise the Open Agentic AI Knowledge Stack. We show where the most value will be created and how the OSS AI ecosystems come together to build and deliver it.

OAKS consists of structured input, knowledge transformation, and structured output. We show the Agentic AI architectures emerging around AI memory, graph-based agentic workflows, and frameworks including scalable message passing, knowledge encapsulation, and colocated knowledge and computation for web-scale routing. We invite the community to join us!

Building High-Throughput Data Orchestration: Instacart's Journey to 20M Daily Workflows | Explore how Instacart built an enterprise-grade orchestration system handling 20 million daily workflows across diverse technical domains. Learn implementation details of their cloud-native platform combining Apache Airflow and Temporal for robust scheduling and execution. Deep dive into YAML-based workflow definitions, GitOps deployment patterns, and observability solutions that enable reliable scaling. Practical insights from years of production experience, applicable to both startups and enterprises building scalable data infrastructure.