LLMs have democratized AI, making it more accessible for everyone. But today’s chat bots still feel very disconnected. Wouldn’t it be great if you could use AI to tap into your personal knowledge and data, and use it to drive the tools you already know and love? Imagine using a chat bot to create your next pitch deck, or generating bespoke 3D scenes for your next home decoration project using inexpensive tools like Blender. In this talk, I'll show you how this is possible with tools like Genkit and MPC, the Model Context Protocol.

You will learn:

• What Genkit is and how its key abstractions make building AI apps easier and faster
• How you can leverage MCP to allow LLMs to interact with existing apps and systems
• How to architect intelligent applications that orchestrate complex tasks by connecting to and reusing your existing APIs, data sources, and specialised tools.

I will even demo an agentic app built using Genkit, MCP, Swift, and Keynote that you can use to craft presentation decks using just natural language - including doing the research!

Automated UI testing is essential but notoriously fragile, often failing to keep pace with rapidly evolving interfaces and documentation. In this talk, we introduce a novel multi-agent architecture that leverages large language models (LLMs) to transform static documentation into dynamic, executable UI tests—without writing code. Our system extracts structured user journeys from product manuals, refines them into atomic tasks using LLM agents, and executes them autonomously in live browser environments.
This talk is for developers, QA engineers, and AI practitioners building or maintaining test automation pipelines in enterprise environment. Attendees will gain insights into architecting LLM-powered agents for structured extraction, and improving test resilience. We'll explore how this system achieves high execution accuracy and adaptive self-correction, even under ambiguous or incomplete specs.

What if your AI assistant could travel through time, rewriting the rules of documentation as it goes?
In this talk, we’ll take you on a journey to the future of knowledge management, where your assistant doesn't rely on outdated documents. Instead, it can ingest real-time information via chat, making sure that this new data takes priority over the old. Powered by cutting-edge Generative AI, LLMs, Bots, and a Retrieval-Augmented Generation (RAG) architecture, our system is capable of time-traveling knowledge, collecting, structuring, and storing it as the ultimate source of truth.

We’ll dive into the research and technical architecture that made it possible, exploring how we integrate dynamic knowledge injection, prioritize real-time inputs, and maintain consistency across the system. You’ll learn how we built an AI assistant that adapts to evolving needs without compromising on reliability or user experience.

So, buckle up! We’re about to take your AI into the present. No flux capacitor required!

Target audience: developers, ML engineers, product managers.
Expected duration of the talk: 40'

As individuals and organizations strive to make smarter, faster decisions, AI is evolving beyond prediction into a new era of causal, intelligent, and collaborative decision-making. This session explores how developers can build cutting-edge Decision Intelligence Platforms by combining three key technologies. First, we'll look at how graph analytics uncover causal relationships in data, revealing patterns and dependencies that power smarter insights. Next, we'll dive into a simulation engine that enables scalable scenario planning, helping teams test “what-if” situations before acting. Finally, we’ll explore generative data storytelling and an alerting and recommendation system that transforms raw data into meaningful, real-time narratives and guidance. Together, these elements create a platform that doesn’t just react to data—but understands it, plans around it, and collaborates with users to make proactive decisions.

AI coding assistants like Copilot, ChatGPT, and Cursor have transformed how we write software — but often in unpredictable ways. Vague prompts, inconsistent results, and scattered workflows are now part of our daily reality. The promise of AI-powered development is huge, but without structure, it's easy to fall into chaos.

In this talk, I’ll introduce User Story Matrix (USM) — an open-source CLI tool that helps developers orchestrate AI-assisted development with clarity and control. USM lets you define structured user stories, generate implementation blueprints, and guide AI tools through testable, repeatable steps.
Whether you're working solo or in a team, USM gives you a lightweight, developer-first workflow to make AI coding deliberate and scalable — not just experimental.
This talk is aimed at developers, tech leads, and AI tool builders looking to reduce noise and get real results from AI coding. You’ll walk away with a fresh mindset, a practical CLI demo, and a repeatable process to bring predictability back to your AI-assisted workflow.

Are you building AI applications that need a reliable, scalable vector store? This talk provides a comprehensive guide to configuring PostgreSQL as an optimal vector store for AI applications using the pgvector extension.

We'll begin with a brief introduction to vector stores and PostgreSQL fundamentals before diving into the pgvector extension and its capabilities. The presentation will explore essential vector operations and compare indexing approaches (HNSW vs. IVFFlat), offering practical guidance on hyperparameter tuning for your specific use cases.

We'll also cover advanced optimization techniques including:
- Reduced precision
- Quantization methods
- Various scaling strategies such as partitioning, sharding, and replication to handle growing data volumes

We'll also see a real-world example of a semantic search application using pgvector, demonstrating how to implement and optimize it for production.

Whether you're implementing semantic search, recommendation systems, or other vector-based AI applications, you'll leave with actionable knowledge to configure PostgreSQL as a high-performance, production-ready vector database.

Modern AI development is rapidly shifting from isolated LLM prompts to orchestrated systems of collaborative agents. But with so many tools and frameworks emerging, interoperability remains a key challenge. In this talk, we introduce three powerful and complementary technologies: CrewAI, a lightweight framework for autonomous multi-agent collaboration; LangGraph, a graph-based orchestration engine for LLM workflows; and Google’s Agent-to-Agent (A2A) protocol, a new open standard enabling agents to discover and communicate across frameworks and infrastructures.

You’ll learn how each technology works, what problems they solve, and how they differ in design and application. The final part of the talk is a live demo: an AI agent that receives a user request and coordinates with two separate agent stacks—one built with CrewAI and another with LangGraph—via the Google A2A protocol, showcasing seamless cross-framework problem-solving in action.

Audience: AI developers, architects, and technical leads building agent-based systems or evaluating agent frameworks.
Length: 40 minutes.
Innovation: This is a brand-new talk featuring a live demo of multi-framework agent interoperability using Google A2A.
Takeaways: Attendees will leave with a clear understanding of CrewAI, LangGraph, and A2A; practical insights on when and how to use each; and a blueprint for building scalable, interoperable, multi-agent systems.

We are in the AI Agentic phase, where AI entities are able to autonomously reason and act (ReAct) and this will literally change the way we Live, Play and Work. A new army of Autonomous Digital Workes will join the humans at work as well as in private life changing the entire IT industry: software, hardware, infrastructure and services. This moltitude of digital resources will require a new "internet" working differently, al layer 7, able to Connect, Secure, Observe and foster Collaboration exactly as it was in the beginning of the internet revolution. We are at the dawn of the IOA, Internet of Agents.

What happens when Generative AI meets the world of dining? At TheFork, we’ve transformed restaurant discovery into a personalized experience for every user. This talk will take you through our exciting journey of adopting GenAI—balancing innovation with privacy, legal compliance, and the creation of a solid AI blueprint to support multiple internal initiatives. We’ll share how we’ve leveraged diner reviews to enhance the user experience, from generating smarter recommendations to crafting personalized restaurant descriptions. Join us as we explore the challenges, lessons learned, and the strategies that are shaping TheFork into an intelligent, user-centric dining advisor. Join us as we explore the challenges, lessons learned, and the strategies that are shaping TheFork into an intelligent, user-centric dining advisor.

This 30-minute talk will motivate and introduce Physics and Quantum methods that will improve current machine learning algorithms and perhaps help solve the growing energy demands of LLMs. Starting from Landauer's principle, we will discuss the limitations of current semiconductor technologies and how emerging computing paradigms could power future training and inference.

Why do even successful GenAI projects often plateau after reaching production?

Moving from PoC to a live system is only the beginning — the real challenge lies in scaling it, making it trustworthy, and continuously improving its impact. Whether you’re navigating your first demo or already running a GenAI solution in production, this talk is for you.

Here I’ll share hard-earned lessons from building production-grade multi-agent chatbots (e.g. “talk-with-your-documents”) in a corporate environment — lessons grounded in hands-on experience, packed with practical tips and references to GenAI tools and frameworks that actually work.

We’ll cover:
- why involving business stakeholders early and often is not just a nice-to-have, but a critical success factor — and how to do it.
- how to build robust evaluation pipelines using real user data and expert input to assess relevance, accuracy, and overall system performance over time.
- how to strike the right balance between elegant simplicity and complex agents architectures, with a focus on composable patterns and advanced strategies for RAG (chunking, pre-retrieval, post-retrieval, and more).
- the main challenges of unstructured data — and how smart preprocessing can be helpful in compensating for poor document quality.
- post-deployment strategies: onboarding users, increasing trust through transparency (think interactive source tracing and agent flow visualization), and iterating based on real-world queries and feedback.

This is a new, original talk for AI Heroes, aimed at developers, AI engineers, and product teams building GenAI solutions. Get ready for a session filled with hands-on advice and real-world lessons learned from taking GenAI projects to production across a variety of contexts and scenarios.

Expected length: 40 minutes

As AI floods the digital landscape with content, what happens when it starts repeating itself?
This talk explores model collapse, a progressive erosion where LLMs and image generators loop on their own results, hindering the creation of novel output.

We will show how self-training leads to bias and loss of diversity, examine the causes of this degradation, and quantify its impact on model creativity.
Finally, we will also present concrete strategies to safeguard the future of generative AI, emphasizing the critical need to preserve innovation and originality.

By the end of this talk, attendees will gain insights into the practical implications of model collapse, understanding its impact on content diversity and the long-term viability of AI.

The foundational RAG architecture we all know - built on embeddings, vector databases, retrieval algorithms, and generative models - served us well in 2023. Now it's time to explore the powerful enhancements that have emerged to take RAG to the next level.

We'll explore HyDE, an approach that leverages an LLM to generate hypothetical documents, creating a stronger bridge between user queries and relevant content. The result is a marked improvement in retrieval accuracy without requiring additional models.

Visual content handling gets a major upgrade through advanced embedding techniques, particularly with the introduction of ColPali. This retrieval model works directly with document images, using vision-language models to create rich, contextual embeddings from document pages. It's particularly powerful when dealing with PDFs, scanned documents, and other visually structured content that traditional RAG systems struggle with.

For those focused on efficiency, we'll examine SPLADE, a neural retrieval model that creates sparse representations of both documents and queries. By combining the speed advantages of traditional inverted indexes with the accuracy of modern neural approaches, SPLADE offers a compelling solution for production environments.

The discussion isn't just theoretical - we'll ground everything in practical implementation. Through code snippets, engineering best practices, and live demonstrations, you'll see these enhancements in action and understand how to integrate them into your own systems.

Whether you're building your first RAG system or looking to enhance an existing one, this session will equip you with actionable insights and practical solutions to make your retrieval systems more robust and capable of handling real-world challenges.

Replica Italia is a synthetic, privacy-safe digital twin of the Italian population. This presentation explores the technical challenges and data sources involved in its development.​
Creating a digital twin required integrating diverse datasets: we utilised open data, statistical modelling, and simulations to represent demographic, behavioural, and geolocation attributes. Key data sources included public demographic statistics, consumer behaviour studies, and geospatial information.​ The resulting platform provides instant access to AI-generated panels of over 60 million synthetic Italian users, facilitating concept testing, surveys, and consumer response simulations. It also enables the enrichment of customer profiles and the creation of high-quality synthetic audiences for targeted advertising, all without accessing personal data.
This talk provides an in-depth exploration of the methodologies employed to integrate diverse data sources in creating a national-scale digital twin.