Abstract

I've been building production AI agents with LangGraph - sometimes getting it right, often learning the hard way. This talk is about what I actually shipped, including a distributed multi-agent setup at PagerDuty. I'll break down the real tradeoffs between LLM-driven and deterministic orchestration, what broke, and how I'd approach it differently now.

Bio

Viktor is a Senior Software Engineer at PagerDuty, where he builds AI systems and scalable backend infrastructure. With a full-stack background ranging from fast-paced product teams to major enterprises like Sky and NBCUniversal, he specializes in Python, AWS deployment, and AI application development using frameworks like LangGraph.

Abstract

In AI and LLM projects, we often talk about big ideas. But in my experience, the real difficulty is not always having the idea. It is knowing what to do next. When an idea feels too big, too abstract, or too ambitious, it is easy to postpone it, overbuild it, or create a demo that looks interesting but does not solve a real problem. In this talk, I would like to share a simple way of thinking that has helped me in real automation, integration, and LLM projects: instead of asking only “Will this idea work?”, I try to ask “What is the next step I can take to validate it?” That next step can be a small prototype, a manual test, a workflow with an LLM, a conversation with users, or even the conclusion that AI is not necessary yet. I will share practical lessons from building AI workflows, assistants, and enterprise automations, including some doubts, mistakes, and decisions that usually happen before an idea becomes a real product. The goal is not to present a perfect formula, but to open a conversation with the community about how we can move from AI enthusiasm to useful experiments, without losing focus or control.

Bio

Fabricio De los Santos is an AI Solutions & Integration Architect with more than 25 years of experience in enterprise software, system integrations, automation, and applied AI across LATAM and Europe. He is co-founder of DELFA and leads the development of ROMA360, an AI-enabled automation platform focused on workflows, integrations, and assistants. His work combines practical LLM integrations, RAG, APIs, FastAPI, n8n, low-code experience, and real production projects. He is also an author and writes about AI, technology, and human decision-making.

Abstract

We've learned to expect privacy from the tools we use. Signal, Proton, and others have shown that companies can build real products while proving they never see your data. But what about AI? For users handing over their most personal questions, and for companies sitting on sensitive data, third-party APIs are something to avoid. The most useful AI is often the AI they can't safely use. Is running everything locally the only answer? We'll start there, looking at how far local and edge AI have come and where it breaks down. But even with cost and scale out of the way, imagine a company that has a good value-add and wants to ship AI features while preserving privacy. Following in Signal and Proton's steps, they'd have to expose all their hard work: fine-tuned model weights, agentic systems, and more. Current solutions for processing private data remotely exist, but they're not part of your everyday dev feed. We'll work through the techniques that close the gap: from computing on encrypted data to hardware-isolated enclaves, weighing how deep your trust needs to go, the tools you can reach for today, and what's coming next.

Bio

Afonso Oliveira has spent 14 years figuring out how to make AI useful. He first applied this in research, earning a PhD with honors from Instituto Superior Técnico on neural networks for 5G resource allocation, then brought that rigor into product development. At Siemens, he turned scattered AI projects into a unified product that cut proof-of-concept delivery from a week to under an hour. At Mindmymind, he built cognitive profiling and self-reflection systems where ML, games, and LLMs meet mental health. Today he writes and consults on AI engineering through OliveGradient, while advising on and implementing AI solutions for a growing SME. Somewhere along the way he also co-founded a startup, won a few international hackathons, and wrote a chapter in a book on AI published by ISCTE.

Abstract

Building the first AI agent is easy. The real challenge begins when dozens of teams start building dozens of agents. As organizations scale their AI investments, teams naturally optimize for speed and autonomy. They choose different frameworks, define their own tool interfaces, implement memory differently, create custom evaluation approaches, and establish their own operational practices. Each decision makes sense locally. The problems emerge when the organization wants these agents to collaborate. Capabilities developed by multiple teams become difficult to combine. Integrations become expensive. Evaluations become incomparable. Observability becomes fragmented. Innovation slows as every new initiative requires custom glue between incompatible systems. In this talk, I will discuss why agentic development is reaching its platform moment. Drawing on lessons from scaling AI products and platforms in large organizations, I will explore the standards, interfaces, and architectural principles that enable teams to innovate independently while remaining interoperable. We will cover common patterns for tool integration, memory, evaluation, observability, and governance, as well as the organizational challenges that emerge as AI adoption grows. The future challenge is not building more agents. It is enabling agents built by different teams to work together.

Bio

Tetiana Torovets is a data science and AI leader with more than 10 years of experience building and scaling machine learning products. Her background spans search, personalization, recommendation systems, experimentation platforms, and conversational AI across both consumer and marketplace businesses.

Abstract

Helene Guillaume sold Wild.AI to NYSE listed group in October 2025. Her company opened the path of AI for biohacking. As a world pioneer in AI, she is now advising governments on AI, writing a book, and sharing with her community the actual tools and ways to use AI day to day.

Bio

Helene Guillaume built and sold Wild.AI to NYSE:ZEPP, the pioneering longevity AI company. She's an athlete and a writer. When she's not on a surf board, she's advising founders or company boards.

Abstract

Conversational AI has quietly become the most intimate form of human–computer interaction we have ever built. From character.ai companions racking up billions of messages, to AI tutors that hold a child's attention longer than any app before them, talking to a machine no longer feels like science fiction - it feels routine. But behind every smooth conversation lies a stack of design decisions: how do you keep latency under the threshold where it stops feeling human? How do you let users interrupt mid-sentence? Do you bet on a single end-to-end speech model, or stitch together ASR, an LLM, and TTS like classic plumbing? In this talk we'll tour the landscape of conversational AI — its breakout applications, its risks (especially in companion and edtech use cases), and the two architectural schools fighting for the future. Then we'll roll up our sleeves and build one. Live, on a laptop, for free, walking through every block of a modern voice agent. You'll leave with a block diagram in your head, a list of swappable components, and a ready-to-code spec you can hand to Claude Code tonight to ship your own AI companion before bedtime.

Bio

Xavier Anguera is SVP of AI for Language Learning at Babbel and a recognised expert in Speech AI, Deep Learning, and Large Language Models. With more than 20 years of experience, over 100 publications, and multiple patents, he has led the development of AI systems at production scale across industry and research.

Abstract

Prompt engineering has traditionally been manual trial-and-error. When AI agents degrade in production, teams make gut-feel changes and hope for the best. How do you know if a prompt change actually improves things, or just shifts failures elsewhere? We built an automated system that transforms production failures into quantifiable fixes. It starts by fetching examples of suboptimal sessions from our observability platform. The system analyzes these failures and clusters them into actionable patterns. For each pattern, it proposes specific prompt changes with before-and-after text. Then comes the critical step: validation. Only improvements that maintain or boost performance against our golden dataset make it into an automated pull request. Finally, the golden dataset itself is updated with representative samples of assessed failure. The result? We analyze all failures weekly and ship validated fixes—all without manual debugging. This approach transforms vague ""my AI isn't working"" complaints into concrete, measurable improvements backed by production data.

Bio

Scott Bastek is a Data Science professional at PagerDuty with a background from MIT. Based in Lisbon, he focuses on applying data science and modern AI techniques to solve practical engineering and operational challenges.

Abstract

Almost ten years ago I was a data scientist who memorized APIs, read papers to find a gap, and scoped every project to what I could personally code. Today I write less code than ever and ship more software than ever — across research, data work, building a company, navigating immigration, and helping my family. This is five short before/after stories with one shared shape: where my thinking used to start, where it starts now, and the skill I quietly stopped exercising in exchange. No architecture diagrams — instead, the human version of a problem many of us are solving in our systems: when producing work gets cheap and verifying it gets expensive, do we upgrade our judgment as fast as our tools? I'll use real examples — from RAG-based legal assistants to building imigrei.pt and Prep Docs to help immigrants face Portuguese bureaucracy—and close with an uncomfortable question. You'll leave having relocated at least one bottleneck in your own workflow, and a little unsure whether you're still good at the part that now matters most.

Bio

Marcelo Ventura is a data scientist and AI engineer based in Lisbon. Over nearly a decade he has built ML and data systems across Brazil, Singapore, New York, California, and Central Europe — from loan-forecasting neural nets and cashflow automation to RAG-based GenAI assistants and legal copilots at fintechs and startups. He holds an MSc in Applied Economics and has published research in statistical modeling and econometrics. A Brazilian immigrant himself, he is now Data & AI Engineer at imigrei.pt, where he builds tools — including Prep Docs — that help immigrants navigate Portuguese bureaucracy, and hosts the Imigrantes de Sucesso podcast.

Abstract

Voice bots work perfectly well when running a demo on Vapi or Retell, but not in actual production deployment. At DestiLabs, we completed 5 recent projects each teaching us something different about this technology and we're willing to share those insights with you. If you'd like to know more about - Why voice-to-voice models fail in phone calls

• What challenges suddenly appear when your project must be deployed on four continents
• How your name can be (and will be) mispronounced by the voice bot

Bio

I'm the cofounder of DestiLabs, an AI/ML agency that specializes in building something no one has ever done so far. Through last 3 years we've built:

• Lip sync model before Sora was even born
• Room segmentation solution that still beats Segment Anything 3
• Voice platform that can run calls cheaper and faster than Vapi
• Image to 3D pipeline that creates 3D coins from 2D images

Abstract

Most "I built an app with AI" stories stop at the prototype. This talk is about the harder part: taking an agentic coding tool from a working demo to software you can put real users on and maintain by yourself. I'll use app.smarthostinghub.com — a production app I designed, built, and run single-handedly — as the running case study, with a quick live look at how it's structured. The core idea: once the model can write the code, the bottleneck moves to context and verification. So I'll focus on just three things that make solo agentic development hold up in production:

• Set the context first: a written spec, data model, and a CLAUDE.md / docs layer that act as the agent's memory — because any rule you don't write down, it will silently break.
• Ship in small, reviewable slices: every feature end-to-end (data → API → UI → test) in chunks small enough to actually review, with types and tests as the gate that stops bad code from merging.
• Keep humans on the risky seams: auth, billing, and database migrations stay under explicit review, never on autopilot. Then the honest part: what broke, and what I'd do differently now. Attendees leave with a short checklist they can apply to their own projects the next day.

Bio

Aleksandr Kamenev is the founder and CEO of NerdHeadz (nerdheadz.com), a Wyoming-registered AI and custom software development agency; he's based in Lisbon. Over ~12 years in software he's moved from enterprise development to leading a distributed team that has shipped 60+ projects, and today he spends much of his time building and running production applications single-handedly with agentic coding tools. Trained as a radio engineer, he bootstrapped NerdHeadz past seven figures in revenue without outside funding.

Abstract

Most people talk about AI tools. I want to talk about the thing that makes them actually work i.e. Memory. Not the model, not the prompt, the persistent knowledge layer underneath it all. I'll walk through how high-performers are building their second brain: a structured, queryable layer of their own thinking that compounds over time. I'll cover what actually needs to go in (and what doesn't), how to design it so an LLM can reason over it, and what this looks like in practice for founders, consultants, and operators running on context-switching overload.

Bio

AI practitioner with 14+ years in the field, 2x published author (Packt), Techstars alumni, and founder of Timeln. I've delivered production ready AI systems across BFSI, Telecom, O&G, Healthcare, and ad tech.

Abstract

AI agents don't search like humans. They reason between every search call, and unlike humans, they're not lazy about spelling it out: what worked, what didn't, what to look for next. That chain of thought is rich retrieval signal that today's search engines throw away. Late interaction retrieval models can exploit that signal. Unlike single vector embeddings, which compress an entire document into one point in space, late interaction keeps a vector per token, capturing far more nuance with far smaller models. The tradeoff is more compute at search time, but efficient indexing techniques have largely solved that. And because these models are comparably small, they're fast and cheap to fine tune for your domain. That matters especially when you work with proprietary or specialized data where off the shelf models fall short: a generic retriever has never seen your internal documents, your terminology, your relevance criteria. Fine tuning bridges that gap. That's where frontier LLMs change the game. Training a retriever for your domain used to mean hiring annotators. Today, frontier models judge relevance at human evaluator quality and smaller, more affordable models can handle many of the same tasks at a fraction of the cost. I'll walk through a reproducible recipe: pool candidates with cheap retrievers, rerank them with LLM judges, extract training pairs. Then fine tune a late interaction model in minutes on a single rented GPU.

Bio

Throughout her career, she has led teams responsible for delivering AI and data products at scale, from ranking and recommendation systems to agentic AI platforms. She is particularly interested in the intersection of technology, organizational design, and platform thinking—helping companies move from isolated AI experiments to interoperable, production-grade AI ecosystems.

Abstract

Bio

João Freitas is General Manager and Engineering Lead for AI at Pager Duty. João leads PagerDuty AI initiatives and he is the main representative of the Lisbon office, being responsible for its growth, expansion, and culture. With about 20 years of experience in Software Development, Machine Learning, and as a People Manager, he was previously CTO at a startup in the area of Artificial Intelligence and has taken several roles at Microsoft in the areas of Speech Technologies and Data Engineering. João also holds a PhD in the areas of speech technology and human-computer interaction, filed several patents, and published over 40 articles in peer-reviewed international conferences and journals. He is also a regular speaker at AI conferences and the author and co-author of book chapters and one book.

Abstract

Training gave us powerful models, but inference is where the real production battle begins. Agentic workflows are burning money by the second, GPU economics are tightening each quarter, and a single unpredictable latency tail or silent kernel failure can turn a working system into an outage. Reliability thus is no longer just about uptime.This keynote is a call to SREs, GPU engineers, and LLMOps practitioners. I’ll share some of the hard lessons, emerging patterns, and new rules shaping the next decade of reliable AI systems

Bio

Abi Aryan is the founder of Abide AI, educator and author of the O'Reilly book LLMOps and an upcoming book in GPU Engineering. With over 10 years of experience productionizing AI/ML systems and prior research experience at UCLA under Dr. Judea Pearl, she has been focusing on making AI systems reliable, observable, and sustainable at scale.