GenML 2025

– The LTXV model architecture – The basic specs, what's special about it, and why we chose this.

– How do you train LTXV? – What are the components you need to train a video model, what teams need to do, and what specific technical elements?

– Examples from our experiments – some behind-the-scenes look at how different elements affect the quality of the model. Data, losses, scale.

Foundation model training requires a large amount of compute power and, traditionally, substantial manual interaction, little signal for improvement, and slow iteration times. In this talk, we will review the challenges and the engineering behind foundation model training at Poolside. We will detail the machinery we have built to automate and streamline every stage of a large-scale training process. We start by reviewing how we explore the hyper-parameter space for the optimal combination for the model. Later, we will review the compute-intensive pre-training stage and how we distribute the training workload across thousands of GPUs. We then continue to the fine-tuning stage, where we apply reinforcement learning, and conclude with the final stage of preparing the model for production.

Pretrained transformers have revolutionized natural language and vision tasks, but can they do the same for tabular data? In this talk we’ll discuss TabPFN, a recently introduced pretrained transformer designed for tabular data predictions. We’ll present its core idea, main limitations and variants. We will then discuss its evaluation on public benchmarks, as well as our own comparison between TabPFN and its variants using real-world proprietary datasets. We’ll conclude with some practical insights and recommendations for data scientists working with tabular data.

Building production-ready AI agents typically requires juggling multiple tools, complex integrations, and fragmented workflows that slow development and increase operational overhead. This session demonstrates how Amazon Bedrock AgentCore provides a unified platform that eliminates this complexity, offering everything needed to develop, deploy, and monitor AI agents at enterprise scale through a single comprehensive solution based on building blocks for your needs.

AI agents are no longer just prototypes, they are powering customer support, analytics, and decision-making in production environments. But while agents can act autonomously and chain multiple reasoning steps, tracking their behavior is far from trivial. Traditional monitoring is not enough when the system makes complex tool calls, generates unpredictable outputs, and continuously evolves.

In this talk, I’ll share hard-earned lessons from deploying AI agents in real-world production systems. We’ll cover the core challenges of observability: capturing agent traces, logging prompts and tool usage safely, monitoring costs and drift, and defining metrics that go beyond accuracy to measure reliability and trust. I’ll walk through best practices to make AI agents more transparent, debuggable, and trustworthy, before things break in production.

This session explores the advancement of AI reasoning, moving beyond conventional architectures toward structured thinking and decision-making. The discussion will examine the theoretical foundations that enable models to perform complex, multi-step tasks. From an NVIDIA perspective, we will share key insights from the development of the Nemotron family of reasoning models, offering practical perspectives on current architectures and the processes behind building advanced reasoning systems.

Your code is running fast, results look correct, but your API bill is skyrocketing. Hidden beneath the surface was a subtle async pitfall: all requests were being sent at once, even when the process stopped early. The outcome? Ten times more calls than intended, and costs are climbing out of control.

In this talk, we’ll dive into a real-world example of an LLM validation script that looked fine on the surface but was quietly sending unnecessary requests. With just five lines of Python, costs were cut by 90% while keeping performance intact. We’ll explore how async methods can silently create inefficiencies, how techniques like semaphores can control concurrency, and how being intentional about how your code runs can save massive time, money, and frustration in the long run.

The talk is based on https://towardsdatascience.com/how-we-reduced-llm-cost-by-90-with-5-lines-of-code/

The rapid integration of Generative AI technologies, and particularly their multimodality, presents an exciting frontier for tackling complex and lengthy tasks. However, it also opens the door to potential security risks in the cybersecurity landscape. In this talk, we will delve into the security of multimodal large language models (LLMs) by demonstrating various image-based attacks on multimodal systems, illustrating the vulnerabilities these models may possess. Throughout the talk, we will be focusing on the creation of an image based heat-maps for stating where are the most sensitive points for a successful attack, and exploring those attention points from the eyes of an attacker or in other words demonstrating why they are a magnet for bad actors. These research aids not only at explaining why a set of points were successful but also expanding the researchers and engineers mindset in how to approach developing new detectors or security mechanisms in a new blue ocean such as the multimodal presents.

Large language models are revolutionizing how we build intelligent systems, but deploying them on massive datasets in production brings serious challenges — from runtime bottlenecks and spiraling costs to maintaining output quality at scale. In this talk, I’ll walk through a real-world case study of running LLMs over big data pipelines, sharing lessons on optimizing inference performance, managing compute costs, and ensuring accuracy.

As AI systems are increasingly deployed in sensitive domains such as healthcare, finance, and security, ensuring that private data remains protected while being processed by AI systems has become a critical challenge. Homomorphic Encryption (HE) provides a powerful foundation for privacy-preserving computation, allowing models to process encrypted data without ever decrypting it.

In this talk, we introduce the principles of HE and its integration with deep learning. We discuss the challenges of adapting transformer-based architectures to operate efficiently under encryption and present recent innovations that make encrypted inference practical. Finally, we demonstrate how these methods can be applied in real-world settings where sensitive data must remain private.

This approach bridges rigorous cryptographic theory with real-world AI practice, paving the way toward trustworthy, privacy-preserving AI systems.

Aidoc has recently developed one of the largest and most advanced clinical foundation models in

healthcare, setting a new benchmark for accuracy and scale in the industry.

That was made after Aidoc refined its supervised learning approach over the past 9 years,

bringing it to a level of maximum efficiency, which has enabled the company to develop 3–4

In this talk we will embark on a journey to build an AI agent designed to assist teachers with grading. Although the task appears simple, it harbors significant underlying complexities, which will be demonstrated when building it with leading AI agent frameworks. Ultimately, we will demonstrate how AI21 Maestro's unique capabilities are leveraged to effectively address this problem.

AI hype is everywhere. After four years of building production-tested AI products, I still have to explain that LLMs aren't a drop-in fixes for classic data problems.

Amidst the buzz, it’s easy to overengineer – turning every retrieval task into a complex RAG-flow and every collection of documents into a knowledge-graph.

Teams struggle with over-complicated solutions where classical tools would suffice, mistaking similarity for relevance, or betting big on agentic flows when bag-of-tricks NLP would save time and budget. Meanwhile, costs drop but technical debt grows, as data curation, cleaning, and operational know-how are neglected for architecture debates and API wrappers.

Instead of just highlighting the common pitfalls, this talk offers concrete, field-tested alternatives; Simple Python for complex retrieval, lightweight NLP, and cost-effective labeling. Attendees will discover how embracing practicality over hype can lead to more robust, impactful, and maintainable AI products.

The BigQuery Data Science Agent is an AI-powered tool designed to assist data scientists and analysts with their workflows. It leverages LLMs like Gemini to understand natural language prompts and automate various data science tasks.

This session will show how this innovative tool, powered by Gemini, streamlines data science workflows through intuitive natural language interaction, empowering both data scientists and analysts to achieve more with less effort.

In verbit we use temporal workflows for ML pipelines, with engineering and research collaborating on them together, in a seamless manner. We've essentially created a "plugin" system for ML algorithms that runs in production, allowing research to continuously "upgrade" our algorithms without needing to touch engineering team production code, and vice versa. This allows both for engineering to run fast with initial ideas, and research to run more incrementally without creating a bottleneck in the product delivery pipeline.

For enterprise-scale applications serving the world's largest companies, the risk of AI hallucinations and inaccurate outputs is significantly amplified.

The challenge is to plan a methodology that enables the creation of complex, open, and highly customizable AI agents without sacrificing critical requirements like ease-of-use, secure data access, and prompt monitoring.

Oracle’s Fusion & NetSuite solutions, include such methodology as they heavily integrate AI agents.

In construction, critical project information is often buried in visually rich formats (architectural drawings, CAD exports, specification sheets etc), rather than clean, well-organized text. These files hold essential details for design, compliance, and execution, yet their visual complexity makes them difficult to search or integrate into AI workflows.

Using AWS Bedrock Knowledge Bases, we can transform these complex documents into structured, queryable data and power Retrieval-Augmented Generation (RAG) workflows. This session will cover parsing visual elements into text, enriching them with metadata, and tuning Bedrock KB for precise, reliable retrieval.

The result is a living knowledge base that turns raw construction data into opportunities for dozens of AI-driven capabilities. Attendees will take away practical methods for ingestion, retrieval optimization, and integration of visually rich data, unlocking a whole new world of opportunities for them.

For years, data security and data understanding have focused on extracting patterns, the “trees” in the forest, entities, tokens, regular expressions. But many of the world’s most sensitive or valuable files contain none of these clues. A board meeting summary, a product roadmap or a clinical protocol each convey meaning beyond any single data element.

In this talk, we present a generative AI approach to file-level classification: a model that infers what a document is rather than what it contains. Our research combines foundation-model fine-tuning with large-scale generative auto-labeling, and novel data processing techniques that preserve semantic variation while removing redundancy. The result is a classifier that can describe previously unseen document types, cluster emergent concepts across domains, and adapt data sensitivity to organizational context, all without predefined label sets or per-tenant training.

This work demonstrates how leveraging LLMs, domain fine-tuning, and dataset engineering enables systems that not only see the trees, but finally understand the forest

Everyone is excited about conversational AI. Everyone is implementing their own chatbots, until they have to make a conversation behave in production.

Replacing a rigid, quiz-style signup with a dynamic onboarding chat sounds great, but executing it is far from simple. It requires designing a conversation that adapts dynamically, collects data, run actions, and completes all of this within a reasonable timeframe. The real headache isn’t “adding an LLM”; it’s the engineering of an agent that can make decisions, and acts on them using automatic tool triggering, which includes presenting actual assets during the conversation.

In this talk, I’ll show how we treated onboarding as a conversation-engineering problem and shipped a production agent using our internal Python-based Agents SDK. We'll walk through the core building blocks and key considerations of creating and maintaining a real-world onboarding agent in production. By the end of the talk you will learn how to structure conversation flows that adapt dynamically, and how to engineer your agent to reliably achieve specific conversational goals.

Training production-grade classifiers usually demands large labeled datasets – costly in both money and time. We present a practical framework that builds high-quality, task-specific classifiers by optimizing the prompt of a state-of-the-art LLM without retraining or fine-tuning. This is done by treating the prompt as a learnable parameter vector and refining it iteratively through designed loss signals.

This mindset shift enables rapid creation of classifiers that consistently outperform hand-crafted baselines, leading to significantly shortened model development cycles and improved performance – without requiring specialized hardware.

Attendees will learn how to formulate effective loss signals for prompt optimization, design a robust automated framework, and operationalize guardrails and evaluation so that a "few labeled examples + simple criteria" reliably become scalable, accurate classifiers.

In this talk we will showcase SLMs and their different architecture.

We will start by analyzing the current market state of SLMs vs LLMs, what triggers companies to make the transition and what needs to be considered when doing so. We will show how to train SLMs, discuss current challenges, and explain how knowledge distillation can be used to overcome them. Then, we will examine the different SLMs architectures, encoder-only (bi-directional) vs decoder-only (autoregressive) models, and demonstrate on what use cases encoders prevail.

Training powerful and aligned large language models (LLMs) remains a significant challenge for real-world applications. While Proximal Policy Optimization (PPO) is foundational in RLHF, it faces hurdles in reward model reliability and sample efficiency. This session will trace the journey from SFT to Group Relative Policy Optimization (GRPO) and beyond through classical PPO, offering improvements in stability and learning efficiency. Crucially, we will delve into the frontier of Reinforcement Learning with verifiable rewards for LLMs, demonstrating how to integrate objective, programmatic, and tool-based feedback. This approach moves beyond subjective human preferences to achieve provably correct and interpretable LLM reasoning behaviors. Attendees will gain actionable insights into building robust, interpretable, and truly aligned LLM-anchored systems.

Agentic AI unlocks powerful new possibilities – autonomous decision-making, adaptive workflows, and dynamic interaction with the digital world. But with this autonomy comes a profound reshaping of the attack surface. In this talk, we’ll explore the emerging threat landscape through the lens of the OWASP Top 10 for AI agents, highlighting vulnerabilities unique to systems that can plan, act, and learn on their own. We’ll dive into how frameworks like MCP and tools for autonomous web scraping, such as Amazon Operator, are accelerating both innovation and risk. Finally, we’ll close with a brief discussion of defenses – what builders and practitioners must do today to secure their agentic systems.

The 2025 buzz around the Model Context Protocol (MCP) highlights its role in enabling AI models to connect securely with external tools and data via standardized protocols. But how does the underlying communication really work?

In this session, we'll explore MCP's transport layer, focusing on STDIO and Streams. Attendees will uncover the principles driving seamless integrations, common failure points (e.g., stream disruptions in real-time AI agents), and troubleshooting strategies, such as error handling and retries.

Through code examples and live demos, gain actionable insights to build robust MCP implementations, ensuring reliable interactions between AI tools and your LLM.

In this talk, I will discuss our ongoing efforts to make the training and adaptation of large foundation models more efficient and accessible. While foundation models (such as GPT5) achieve impressive performance, they require substantial memory and computational resources, limiting their broader application. I will present several recent methods that reduce training memory, accelerate convergence, and lower inference costs while preserving predictive accuracy. These include techniques that exploit the low-rank structure of gradient updates to enable memory-efficient fine-tuning, as well as methods that apply structured sparsity to identify compact and effective subnetworks within large models. Together, these advances enable the training, fine-tuning, and deployment of large models more quickly and efficiently, supporting a broader range of research and industry applications.