Panos Vagenas is an Advisory Engineer at IBM Research, leading development efforts in the intersection of Artificial Intelligence, Information Retrieval, and Data Management. At IBM since 2020, Panos has made key contributions to various innovative AI technologies, such as Docling. With an engineering career spanning from 2014, he has significant experience in designing and developing enterprise software solutions for industry-leading organizations. Panos holds an MSc in Computer Science from ETH Zurich and has received various awards including the IBM Outstanding Technical Achievement Award, the IBM Technical Collaboration Achievement Program Leadership Award, the ACM SIGMOD Best Demonstration Award, and the ETH Zurich Excellence Scholarship.
Panos Vagenas is an Advisory Engineer at IBM Research, leading development efforts in the intersection of Artificial Intelligence, Information Retrieval, and Data Management. At IBM since 2020, Panos has made key contributions to various innovative AI technologies, such as Docling. With an engineering career spanning from 2014, he has significant experience in designing and developing enterprise software solutions for industry-leading organizations. Panos holds an MSc in Computer Science from ETH Zurich and has received various awards including the IBM Outstanding Technical Achievement Award, the IBM Technical Collaboration Achievement Program Leadership Award, the ACM SIGMOD Best Demonstration Award, and the ETH Zurich Excellence Scholarship.
Docling is rapidly becoming the de-facto standard in open-source document AI. The project has achieved remarkable adoption with over 30,000 GitHub stars, more than 500,000 monthly downloads, and trending to #1 repository globally on GitHub. Now incubated as a Linux Foundation AI & Data project, Docling provides enterprise-grade capabilities for parsing complex layouts, extracting tables, and converting unstructured documents into AI-ready structured formats.
In this workshop, we will use Docling to build workflows ranging from document extraction and conversion into different formats, to advanced RAG and agentic AI applications — covering a wide variety of document AI use cases.
This session is meant for AI tech leaders and builders who would like to learn how to make the most out of their unstructured data using Docling.
We first introduce Docling, the emerging de facto standard for open-source document processing for GenAI and beyond, with over 40K GitHub stars, 1.5M downloads/month, and repeated #1 rankings on GitHub and HuggingFace trending boards. We illustrate the challenges it addresses and outline its architecture, including our AI models that power its vision capabilities, as well as the unified data model.
After the introduction, we move to the workshop part, where we walk through various Docling capabilities, starting from basic usage, covering document conversion and exploration, moving on to rich native chunking capabilities, over to end-to-end scenarios such as advanced RAG pipelines including multimodal support and visual grounding, and last but not least agentic workflows leveraging Docling MCP.
By the end of the workshop, participants will have a solid understanding of how Docling significantly simplifies, accelerates, and standardizes working with unstructured data, and they will be able to apply to this knowledge to various scenarios be it typical GenAI use cases like RAG or ultimately any document-related workflow.
Tech leaders, engineering managers, software developers, and practitioners who would like to make the most out of their documents — for GenAI (e.g. RAG) & beyond
Git, Python, Pip installed. Experience with Python, basic understanding or GenAI workflows such as RAG.
Workshop materials will be available here: https://github.com/docling-project/docling-workshops
Open Source Software (OSS) may have taken over the world, but did someone take control of OSS itself in the process, locking it into a proprietary platform, controlled by a single vendor? Is this still a sustainable path today, in the time of trade wars, volatile tariffs, GenAI mandates and countries fighting for data sovereignty? The future of OSS as a global public good is at risk. The need for a resilient, community-owned alternative has never been more urgent.
This talk will introduce you to Radicle, a new, decentralized code forge that’s built on a peer-to-peer architecture.
We’ll explore how Radicle works and how it extends Git to create a secure and resilient network for code collaboration. We’ll cover its peer-to-peer replication protocol, its decentralized identity system, and its novel approach to code discovery.
This talk is for developers of all levels interested in the future of Open Source. You’ll leave with a better understanding of the challenges involved in designing and running distributed, peer-to-peer systems. You’ll also learn how you can start mirroring / migrating your code to the Radicle network, so you can #FreeYourCode !
Interactive storytelling is often treated as a writing problem. In practice, it quickly becomes a systems design challenge.
While building Choose Your Fate, a platform for creating and reading interactive stories, we came to understand what makes the medium genuinely challenging: branching paths that create contradictions across a story, pacing that breaks down as choices multiply, reader agency that feels real but requires careful construction to actually be meaningful. The structure holding the story together turns out to be as demanding as the story itself.
This talk is about what we learned treating narrative as an engineering and craft problem.
We’ll walk through how stories behave like graph systems once you introduce branching, why complexity grows faster than it looks, and how specific structural patterns (diverging, converging, looping, state-persistent) can keep a story coherent while preserving genuine reader agency. A recurring pattern in existing systems is flat choices that carry no real consequence, or meaningful decisions locked behind a monetization layer rather than a narrative one.
We’ll look at the product side too: what it takes to build tools for writers who don’t think in systems, and how you design interfaces that support system-level thinking without asking creators to become engineers.
We’ll also cover how AI fits in this creative space without alienating the creators: not as a story generator, but as something closer to a structural editor that can flag pacing problems and suggest where choices belong.
Attendees will leave with:
– A clearer mental model for why interactive narratives break down and how to prevent it
– Structural patterns that balance branching complexity with story coherence
– Design lessons from building a creator platform for interactive storytelling
– A realistic view of where AI helps in creative tools and where it doesn’t
What is logging, in one word? Cost. Undeniably.
Development cost. Performance cost. Operational cost. Maintenance cost.
But is logging truly costly, or merely expensive?
As with education, the answer depends on outcomes, and those outcomes only become visible over time.
By approaching logging the way we approach learning, we can transform a controversial expense into an operational necessity.
We employ a discreet but continuous source of insight across business, development, architecture and design, security, and incident response.
When designed intentionally, logs tell stories about both system and human behaviors.
They reveal how our services actually run and where assumptions quietly fail.
In that sense, observability through logging is not a sunk cost, but a high-return investment in understanding and decision-making, potentially offering a multi-aspect business intelligence solution.
In this talk, we explore how to extract real value from logging by mapping it to three universal dimensions of learning:
-Things we know that we know
-Things we know that we do not know
-Things we do not know that we do not know
What if your engineering team could move at the speed of AI – without sacrificing quality, security, or control? Enter the Agentile Team: a lean, AI-powered, and platform-enabled evolution of Agile, designed to amplify human creativity.
In this session, we’ll explore how Agentile teams codify architectural tradeoffs (security, capacity, cost) into specs, templates, and guardrails, enabling rapid experimentation and smarter decision-making from day one. Drawing on real-world experiments and platform engineering principles, Suzanne will share practical strategies for developers and platform engineers to thrive in this new era of agentic, high-velocity software delivery.
Most developers use AI coding tools like a search engine: type a question, hope for a good answer. But the teams getting real results are engineering the context: configuring system prompts, structuring project knowledge, and managing the AI’s working memory like a first-class architectural concern.
This workshop is built around a real coding project. Participants will receive a small application to build – and throughout the session, they’ll apply each technique as they learn it, experiencing firsthand how context engineering transforms their AI’s output quality.
We’ll cover:
– **The context window** – what tokens are, what counts as context, and why your AI gets worse mid-conversation (context rot)
– **Basic prompting techniques** – prompt engineering is no longer as important, but can still improve your results
– **System prompts as project configuration** – set up instruction files (CLAUDE.md, copilot-instructions.md) that make every AI interaction project-aware
– **Grounding techniques** – anchor AI output in real data instead of letting it guess, using file uploads, documentation references, and structured retrieval
Each concept is immediately put into practice on the project – so by the end, participants will have both a working application and a repeatable playbook for getting better results from any AI coding tool.
Targeted to software engineers, bring a laptop with an AI coding assistant installed.
Nobody’s talking about this: MCP isn’t just another way to build chatbots. It’s the bridge we’ve been missing between AI reasoning and real-time data systems.
Teams build AI applications that work great in demos but fall apart with production data. Your agents analyze historical reports but can’t tell what’s happening in your Kafka streams. They’re blind to schema changes and disconnected from events that matter to your business.
Instead of treating streaming platforms like black boxes, you expose them directly to your agents via MCP protocol. Suddenly, your AI doesn’t just read about data—it lives inside your data flows.
Learn what becomes possible when you stop thinking about AI as an external service and start treating it as part of your streaming architecture. We’ll build systems where agents subscribe to real-time events, reason about evolving schemas, and make decisions that ripple through your data platform.
