Schedule

Agentic systems fail in predictable ways. A second refund on the same order. A payout sent to the support desk instead of the buyer. An order status of “probably shipped.” These aren’t random glitches — they’re symptoms of one missing layer.

LLMs reason probabilistically over domains they only partially understand, and no amount of prompt engineering fully closes that gap. This talk argues that the missing layer is an explicit ontology: a formal, shared map of a domain’s concepts, relationships, and constraints, sitting outside the model as enforceable logic.

The result is a hybrid neurosymbolic architecture — probabilistic reasoning inside, logical guardrails outside. Drawing on a pattern that has quietly powered enterprise systems for over a decade, the session shows how lightweight ontology constructs surround an agentic system with rules it cannot violate — turning brittle, unpredictable agents into ones you can actually trust in production.

AI didn’t suddenly become emotional when it learned how to have a conversation. But now that AI can remember us, it can begin to feel less like a tool and more like a relationship. Whether we mean to or not, we are designing emotional dynamics into our products.

As AI shifts from something we use to something we interact with, designers need new ways to understand what happens between people and technology. Drawing on a decade of research into the secret psychology of digital life, Pamela will introduce a practical framework for spotting these dynamics and their second-order effects, from algorithmic identity to delegated intimacy. Attendees will leave with a practical way to anticipate how AI systems shape feelings and to design for human agency and well-being.

AI can generate more code than ever, but more code does not automatically mean more value. In this talk, Emilie Schario explores how AI is shifting the role of the engineer from producing every artifact to designing the systems that reliably turn intent into outcomes. She’ll cover how judgment, verification, constraints, recovery, and human oversight become more important as agent-driven development scales.

As AI rapidly reshapes the way we build products, lead teams, and make decisions, many organisations are becoming faster, but not necessarily wiser. In the rush toward optimisation, automation, and scale, something more fundamental is quietly being tested: our ability to remain human inside increasingly machine-shaped systems.

This talk introduces the Luminary Framework, a leadership and cultural model developed through real-world experience leading high-performing engineering organisations through complexity, growth, and transformation. Rather than treating AI as purely a technical challenge, the framework explores the human capabilities that become more valuable as intelligent systems become more powerful.

The session will explore three core dimensions: Narrative Shift, Emotional Literacy, and Systemic Clarity, and why these are becoming critical leadership competencies in the AI era. It will examine how teams lose coherence when leaders optimise for output over meaning, why psychological safety becomes harder in highly accelerated environments, and how organisations risk amplifying bias, fragmentation, and disconnection when human depth is removed from decision-making.

Every forecast is a bet on which data to trust, and that bet gets harder when the polls themselves become campaign weapons. This talk explains how a hierarchical Bayesian model forecasts elections: it tracks how opinion drifts, learns each pollster’s systematic bias, and turns conflicting surveys into probabilities with honest uncertainty. The case study is Hungary’s 2026 election, where SzazKilencvenKilenc (199), an independent poll aggregator, used it to forecast all 199 seats in parliament. Its engine, Kronikas, is now open source and already powering forecasts for elections internationally. We close with where polling is heading, and why the same approach works anywhere business data sources disagree.

Before an AI agent touches a real system, it should prove itself somewhere safe. I built a sandbox for that: a simulated factory where an agent handles customer orders, and everything it does is measured against what the human order desk did. At first the agent only watches and decides in the background, without acting. When it agrees with the humans often enough, it earns the next level: suggesting, then proposing plans that a person approves, then acting on its own where actions can be undone. Actions that cannot be undone, like confirming an order to a customer, always stay with a person. When people correct the agent, it remembers the correction, turns repeated ones into rules, and drops back a level until it has proven itself again. The sandbox follows a set of company rules that the agent is never shown, so I can check exactly which ones it learned, and halfway through the run the rules change. I compare the results with an agent that never learns, and the talk includes a live demo, the numbers, and the open-source code.

Using AI to generate software can solve our problems when it comes to automation, but it introduces new challenges that needs to be handled: Security, Compliance, Oversight, Control, and a good experience. Essentially, if we’re not careful, we will just be swapping one set of problems for a new set of problems.

In this talk, Janne will show the audience how to get started building with AI, and then switch to talking about how to also address each of these new AI-introduced challenges through governance, so we don’t just build, but build responsibly, even as non-developers.

Every talk this year is about making AI agents do more — write more code, take more actions, run more autonomously. Yet many enterprise AI pilots still fail to deliver production value, and the usual suspects — model choice, prompts, infrastructure — are often not the real culprit.

AI everywhere. Companies hustling not to stay behind. However, things are not as easy as we are usually told if we want to leverage the power of AI successfully.

In this session, we will discuss 7 pitfalls that are often lost in the omnipresent AI clamor: Dangers of anthropomorphization, leaky abstractions, review fatigue, productivity traps, AI vampires, sovereignty issues, and more. Not all are easy to fix, but we need to find answers on our journey towards AI. Let’s explore together!

AI is changing not only what we design, but how we design it. New models, tools, workflows, and “this changes everything” moments arrive faster than we can adapt.

But not every new release changes the job. And being good at the newest tool isn’t the same as becoming a better designer.
In this talk, I’ll share three practical ways to work through that uncertainty: how to tell when AI has actually changed the way we design, which skills are worth investing in, and how to turn experimentation into real experience rather than just keeping up with the latest tools.

Through concrete examples from my own work at Meta, we’ll look at what is genuinely changing in design practice – and what becomes even more valuable as AI gets better.

AI can generate code, design data pipelines, create tests, and even propose entire data models in seconds. But generating a technically correct solution is not the same as building the right solution.

Behind every useful data product lies an understanding of business concepts, processes, relationships, and – most importantly -, the questions the data is expected to answer.

In this session, we will explore why data modelling and business understanding become more important, not less, as AI takes over more of the implementation work.

We will look at the role of data engineers and designers in providing the context, semantics, and structure that AI cannot simply infer from a prompt.

Because in the age of “AI-generated everything”, the real competitive advantage may no longer be knowing how to build, but knowing what should be built, and why.

A year ago, PFF had a ChatGPT subscription. That was it. Today we’ve achieved 7.5x developer output and built more in the last six months than we did in the previous two years with a team one-fifth its former size.
This talk is about how we got there, and everything that happened along the way; the stuff that worked, the stuff that didn’t, and the stuff nobody warned us about.
I’ll cover the obvious and the non-obvious. Why hackathons aren’t the answer. Why executive AI committees are a sign your leadership is detached from the problem, not engaged with it. Why bolting AI onto your existing processes is the most expensive way to achieve nothing. And why things started really well for us, then got hard, and what we did about it.
I’ll share the practical checklists we built through trial and error and how to actually transform your org before your competitors figure it out. Not theory. Not a vendor pitch. Just what we learned from doing it for real, getting it wrong a few times, and ending up somewhere genuinely different.
If your company’s AI strategy is still a subscription and a Slack channel, this talk is for you.

Most conversations about production AI agents focus on the agent itself — the prompts, the orchestration, the framework. But the moment you put an agent in front of real enterprise data, a different problem dominates: the data layer wasn’t designed for this consumer. Data lakes were built for analysts and dashboards. Transactional systems were built for applications. Neither was built for a non-deterministic, token-hungry, latency-sensitive reasoning loop that may issue thousands of unpredictable queries per minute.

This talk explores the often-neglected superpower of product design: vision. Not just strategy decks or north star flows, but emotionally charged, speculative, team-energising visionary ideas and thinking, the kind that lifts teams out of the backlog and helps companies imagine what’s next.

Through a cinematic, storytelling-driven format, I’ll use familiar movie moments to illustrate why visionary design matters, how to start practicing it, and how to push through resistance when no one’s asking for it. This talk will be honest, fun, and deeply grounded in personal and team experiences.

AI is no longer just a technological novelty, but one of the most critical business drivers of digital transformation. This presentation explores how AI can become a true competitive advantage when treated not as isolated projects, but as a capability integrated across the entire organization. We will examine the key factors for successful AI implementation, from the role of data assets and platforms to establishing governance and operating models. Through real-world examples, it demonstrates how AI supports knowledge management, software development, and decision-making. It also highlights the organizational, security, and risk management prerequisites necessary for sustainable AI adoption. The session offers practical insights into deliberately building and scaling AI capabilities. Our core message is that AI is not an end in itself, but a key catalyst for efficiency, innovation, and long-term business competitiveness.

Agents are now consumers of your data stack, and unlike analysts they never pause to ask whether a metric is accurate or still relevant. How do you make your data platform robust? What are the best tools out there for both BI and AI analytics, and are they sufficient to prevent semantic rotting?

If you’re a data engineer, data platform engineer, or data architect, you know how difficult it is to build a truly self-serve data platform. In fact, data platforms are often organisational bottlenecks, preventing data publication, consumption, and application of data, and by extension, AI.

This means you and your team spends time on low-value work: moving data between systems, managing access policies, and handling endless support tickets.

Instead, the data platform should be an *enabler*, providing self-service tooling that shifts ownership and responsibility to data producers while freeing your team to unlock data’s true value.

In this talk, Andrew explains why the self-service data platform is the foundational layer that abstracts complexity and provides standardised, easy-to-use capabilities. The key architectural feature is the data contract, which provides the context needed to build the platform easily while maximising user value.

By the end of this talk you will:

– Understand how to shift ownership left without pushing complexity onto busy teams
– Learn how to automate data management tasks, reducing cognitive load across the organisation
– See how data contracts enable automation, transforming them from documentation into live platform features
– Explore how to embed data governance as code rather than governance by people

Why do teams stacked with talent still fall apart under pressure, while teams with modest resources somehow execute like clockwork? Most answers to that question are borrowed from motivational speakers and management fads, not from science. This talk is different. Drawing on original psychometric research, psychologist Joseph Pelrine unpacks a diagnostic framework that explains, precisely and testably, where high performance actually comes from inside a team, and where it quietly breaks down. If you lead, build, or depend on a team that has to get it right when it matters, this session will change how you look at every team you’re part of. Come with your assumptions. Leave with a different way of seeing your own team’s ceiling, and how far below it you might currently be operating.

We put students in front of a webcam to protect the integrity of remote standardized tests. Then our production pipeline started flagging kids for phone use while they drank from a water bottle. On the hard cases, two times out of three.

This talk is about how to build near-real-time detection on vision LLMs. I’ll walk the pipeline end to end: segmenting live video, running VLM passes over short windows, and turning raw model output into signals you can trust.

Every lesson here was measured against a human-annotated data set. The video format you send moves accuracy more than the model you pick: one input change cut false positives by 27 points. Telling the model to “be careful” buys almost nothing, because vision-model language priors override the pixels, a failure two 2026 papers quantify. Two frontier models were unusable for us.

And evaluation, not generation, is the hard part. When two annotators watched the same video and disagreed on what counted as cheating, we handed the same footage to ten people, just to force the argument out into the open. That pushed us to write an objective grading rubric for our annotation team, then to rebuild the scoring logic around difficulty and ambiguity instead of a flat pass or fail.

You leave with a repeatable way to build and evaluate vision detection pipelines, and a sharper distrust of your own metrics.

Our north star: a data expert at every collaborator’s side. A PM, a CSM, a GTM analyst, a marketer — anyone — should be able to get an analyst’s answer whenever they need one.

Two modes:
Reactive — ask in natural language, get the answer a senior analyst would have given, with the same caveats and context.
Proactive — the analyst shows up before you ask: the Monday read on last week’s launch, the “so what?” behind a metric move, the draft of your weekly update. This is less about answering faster and more about helping the whole team decide faster.

The model is the easy part. The hard part is the foundation underneath it: a governed semantic layer of data products and semantic views, exposed through a persona-aware harness that only lets the agent see what the user’s role is allowed to see, with skills that make it reason — and eventually act — like an analyst.

It works: on our first pilot use case, this foundation lifted end-to-end accuracy from ~40% to >90% on the eval set.

As AI hype collides with workplace realities, design leaders face a threefold challenge: managing executive expectations fueled by LinkedIn headlines and pithy X posts, calming teams worried about being replaced, and figuring out how to integrate AI into both new products and legacy systems.

In this talk, Chris Avore shares lessons from leading design organizations through this transition—from using AI to power product innovation to leveraging it for team enablement, like automating operational tasks to create time to design better outcomes. Attendees will leave with practical strategies to set realistic expectations, ease team anxiety, and drive meaningful AI adoption in their own organizations.

AI is like a Swiss army knife: good at everything to some degree, focused on nothing. That makes it hard to derive a use case from — and it isn’t a new situation.

The nineties imagined the internet as a library and a parliament. We built a mall and a stadium. I walk through two timelines — what we imagined, and what actually got built — then draw the same diagram for AI. The two shapes are strikingly similar, and that similarity tells us something about what’s coming.

On top of that conclusion sits a method: one that shows which blade of the AI Swiss army knife to focus on for your own product.

In 2025, Fin (formerly Intercom) took on an ambitious goal: double the throughput of its engineering team, not by building fancy demos, but by using AI agents to get real features, built on a large existing SaaS codebase, into the hands of paying customers. In this talk Brian will go past the hype into the practicalities of scaling agentic engineering at the organisational level: the layers that actually matter, such as skills and codebase preparation, provenance, the SDLC process around agents, designing feedback and verification loops, and building evals that improve the system itself. He’ll share what has and hasn’t worked on the way beyond 2x. This talk is aimed at engineering leaders and staff+ engineers in established organisations trying to get real results out of working with AI agents.

We’re in the middle of a shift: companies are buying AI agents, vendors are selling them, but the rules are still unclear. Pricing models range from tokens to subscriptions to outcomes—each with tradeoffs—and questions around security, onboarding, and real value remain unresolved. In this talk, Eduards brings structure to the chaos by comparing AI agents to software subscriptions and human employees. Each follows a different model, with distinct pricing, security, and user experience assumptions. You’ll gain a clearer framework for deciding when and how to deploy AI agents effectively in your organization.

Tech és leadership podcast, blog, hírlevél, közösség. Ez itt az Anonim Vezetők Klubja. Hetente megjelenő epizódunkban a leadership és szoftverfejlesztési világ aktuális híreiről, trendjeiről és persze AI agymenéseiről beszélgetünk.

At GitGuardian, 80 to 90% of our code is now written by AI. We’re building towards what we call a “dark factory”: systems that take a well-written spec and ship a working feature with very little human typing.

GitGuardian is a 7-year-old, post-Series C cybersecurity startup with a 70-person engineering org that’s scaling fast, and we’re SOC 2 Type II certified. Customers trust us to protect their secrets, so there’s very little room for mistakes. We have to reach our goal of getting to a software dark factory while holding the same quality bar as before, not quietly lowering it to make the numbers look good. Our journey has been like changing the engine of a car while it’s still driving!

This is the practical version of that story. Getting here took far more than handing everyone Cursor and Claude Code. I’ll walk through the building blocks we’ve had to put in place to make agentic engineering safe at our scale, from the infrastructure to run fleets of background agents to keeping a lid on spiralling token costs, and I’ll be honest about what worked, what broke, and what it cost us.

I’ll also talk about the human side, because that’s been the hardest part. What it does to how people learn, the stress it creates, the questions about what engineering careers look like from here, and the real work of driving adoption and meeting resistance head-on.

If your team is going down this path, you’ll leave with the how of what we did and not just the why — the technical scaffolding, the changes to how we work, and the lessons we’re still learning about the people side.

One happy surprise worth a mention: going faster hasn’t meant fewer engineers. We’re hiring heavily.

AI has transformed how engineers work, but most hiring processes haven’t kept up. Candidates now enter interviews with AI tools that can generate code and refine system designs in seconds, yet many companies evaluate engineers as if AI didn’t exist. At Elementor, I redesigned our hiring process to focus on how engineers think, collaborate with AI, and apply judgment when AI-generated solutions fall short. In this talk, I’ll share practical strategies for assessing coding skills, system design, and behavioral signals in the AI era. Whether you’re hiring, interviewing, or leading engineering teams, you’ll leave with a framework for evaluating engineering talent today.

Conventional wisdom says beauty is in the eye of the beholder… so, did the elders lie, or is there no objectively good design?

Well, when you encounter a beautifully crafted product, you just feel it. This talk pauses and explores that moment, asking how to reproduce it in your own work through a slow, diligent process. But slow is a hard sell in a business world that already moves at a ruthlessly breakneck speed – one that new-fangled AI tools promise to make even faster. The question now is: what unique advantages does a methodical approach even offer? How do you defend slow work?

This session is a love letter to product design as an art form that rewards patience and rigor, even as AI promises to shortcut the hard work. It’s for designers ready to treat taste and craft not as some subjective mystery or innate talent, but as legitimate skills – ones that new tech can enable for better, bolder, more meaningful design.

Az új technológiák előnyeit azonnal látjuk: növekvő hatékonyság, gyorsabb munka, nagyobb kényelem. Az árnyoldal megjelenésére azonban éveket, vagy akár évtizedeket kell várnunk – de biztosan megérkezik. Menedzserként, vezetőként nekünk kell kezelni ezeket a nehézségeket. Hogyan (és miért) maradjon emberközpontú a szervezetünk az AI-eszközök elterjedésével? Milyen új vezetői viselkedéseket vár tőlünk a csapatunk és a szervezetünk? Ebben a podcast beszélgetésben körüljárjuk az AI legnagyobb veszélyeit a vezetői munkánkra, és hatásos megoldásokat adunk rájuk.

AI will not replace data engineers tomorrow. But it is already changing how we work — including, occasionally, helping us do stupid things with more confidence.

Used badly, AI will make us weaker engineers. We outsource judgment, trust the output too quickly, miss the details, and slowly lose the habit of critical thinking. Used well, it can make us sharper. AI can help with the parts our brains are weaker at: holding lots of context, exploring many options, drafting code, explaining errors, and moving faster.

This talk is inspired by a real example: Claude Code building a dbt project on DuckDB from API data. It created models, tests, docs, incremental loads, and freshness checks. Impressive. Until you realise it missed API pagination, dropped useful fields, and made questionable modeling choices. The pipeline looked like it worked, but it did not.

AI is great at speeding up the mechanical parts of data engineering: writing boilerplate, debugging dbt errors, generating SQL, suggesting tests, and exploring unfamiliar code. But it does not remove the need for judgment.

Someone still needs to ask: Is the data complete? Is the model correct? Are the assumptions safe? Would I trust this dashboard in a real business decision?

So this is not a talk about “AI instead of data engineers.”, it is about data engineers using AI well. Not to delegate thinking, but to help us magnify our skills.

Andy Cotgreave will explore the tension between the rapid advancements in AI data analysis tools and the persistent need for human craft and context in deriving insights.

He’ll argue that while AI can automate basic analytical tasks, there are significant barriers such as the nuances of chart selection, and the semantic gap.

Andy will describe how these problems appear and provide a framework you can use for assessing new tools’ ability to do data analysis. He’ll look at how to bridge the gap between business language and raw data, proving that data storytelling remains indispensable.

Bullet points:
– The Importance of the human craft in the age of AI
– A practical framework to assess the value of AI-driven analytics tools
– Understanding the places AI analytics can replace or sit alongside “traditional” dashboards

In my experience building products at companies like Google, Spotify, and Microsoft, I’ve learned one powerful truth: the most loved products aren’t just useful—they’re emotionally resonant. They spark joy, create meaningful connections, and leave lasting impressions. That’s what I call Product Delight.

In this talk, I’ll introduce the Delight Framework I’ve developed to help product teams go beyond functional features and create moments of deep delight. I’ll share how to uncover emotional motivators, design for both functional and emotional needs, and move beyond the “nice-to-have” mindset—making delight a core product strategy.

Craftsmanship is the fundamentals of quality. This talk presents a practical framework—Fit, Flow, Fidelity—that helps product teams deliver digital experiences that work for real people across diverse bodies, minds, and contexts.

⁜ Fit combines Intuitiveness and Predictability to make interactions clear and consistent across modalities.
⁜ Flow combines Efficiency and Optimization to conserve time and energy by reducing friction.
⁜ Fidelity combines Polish and Reliability to signal care and ensure access endures through updates and sessions.

Through Microsoft examples, we bring each facet to life with clear, relatable scenarios—like how returning focus after closing a dialog shows predictability, or how precise caption timing reflects polish. Rather than listing checklists, the talk focuses on stories and practical illustrations that make Fit, Flow, and Fidelity easy to grasp, demonstrating how these fundamentals of craft turn ordinary interactions into experiences that feel intentional and inclusive for disabled users.

How do we future-proof not only our careers but also our products and business models in the age of AI? Mirela explores the shift from simply using LLM to speed up PM work to building AI-first products where and when it matters. And three case studies.

What happens when a successful monolith needs to evolve into a platform capable of serving users across continents?

This talk follows the transformation from a region-bound, synchronously connected architecture toward an event-driven platform built around domain boundaries, business events, sagas, and globally distributed read models.

We will explore an architecture in which writes remain centralized while reads are served closer to users through regional projections. Along the way, we will examine the practical consequences of this model: eventual consistency, delayed and out-of-order events, retries, idempotency, data freshness, partial failures, and recovery.

The technical architecture is only part of the challenge. Moving beyond the monolith also requires dozens of teams to change how they design services, publish events, model data, and reason about failure.

Rather than presenting a perfect final architecture, this is a practical account of the decisions, trade-offs, and lessons involved in building an event-driven platform across continents.

Everyone wants to turn on AI features . Then Legal walks into the room and says no. The question we will answer is WHY?

Enterprise AI comes with questions about data, permissions, security, compliance, auditability, human oversight, and trust.

This talk looks at the hidden work behind making AI safe to enable at scale, connects it to the European AI Act, and turns those requirements into practical UX principles for responsible AI products.

Because great enterprise AI isn’t just powerful. It has to be trustworthy enough to actually turn on.

Large language models can reason, but without memory they remain largely stateless, forcing users to repeat context and limiting an agent’s ability to personalize its behavior over time. In this talk, I will share how we built LinkedIn’s Cognitive Memory Agent (CMA), a horizontal cognitive memory platform that enables stateful, context-aware AI agents at scale. I will explain how CMA combines conversational, episodic, semantic, and procedural memory to help agents learn from interactions, understand user preferences, and adapt to how recruiters work.

Using LinkedIn’s Hiring Assistant as a real-world example, I will walk through the architecture behind memory ingestion, hierarchical knowledge representation, and reasoning-based retrieval across multiple memory layers. I will also discuss the production challenges that emerge when memory becomes a first-class component, including latency, evaluation, privacy, access control, stale or conflicting information, and maintaining user trust.

The hiring market has flipped. After years of frenzied growth, tech is now defined by mass layoffs, cautious headcount, hiring freezes, and hundreds of applicants for every role. Hiring cycles have grown longer: more interviews, more barriers, lower confidence.

The problem isn’t the market, it’s that most hiring managers treat recruiting as marketing. Transparent recruiting fixes both sides of the equation by giving candidates the clarity and equity they need to perform their best, while giving hiring managers a faster, more productive hiring process.

With global examples from companies big and small, you’ll see how HubSpot hired 16 highly qualified content designers in just 5 months by focusing on a simple idea: people perform their best when they know what’s expected of them.

Audience takeaways:
– A simple strategy and set of tactics you that transforms your recruiting practices, especially in a slow market where every hire counts
– Compelling ways to set expectations for candidates before they even talk with you, reducing noise in your pipeline and building a strong employer brand
– How to rewrite your job posts to speak directly to your target audience — their goals, motivations, and frustrations — so your roles stand out in a sea of sameness
– How to help your best candidates succeed faster with enablement materials and radically transparent recruiting tools

GHOST is a high-precision GPS-based trajectory tracking platform built from scratch to support large-scale automated data collection. Its development demonstrates a successful, fully AI-driven approach combining agentic coding, spec-driven development, and coordinated AI agent teams. Central to this approach is “feedback-guided agentic team simulation,” an original method in which agents’ system prompts implement a feedback mechanism inspired by proportional–integral (PI) control to maintain alignment with the specification and reduce hallucinations. Throughout development, average cost per request was continuously tracked as a key indicator of AI cost efficiency. The project illustrates how explicit specifications, structured agent collaboration, and continuous measurement can turn AI-assisted development into a disciplined engineering process.

Assistants made coding faster without making delivery faster. I’ll cover what changes when agents take on whole units of work across the SDLC, where the handoff holds up, and where it breaks.

Lessons from enterprise rollouts: how to avoid a different tool per team, establish context that people and agents actually use, and lay the environment engineering foundation so agents can build, test, and verify their own output (so you don’t babysit them).

It’s going to be fun!

AI in analytics engineering doesn’t start with autonomy. It
starts with structure. At Bolt, we are building a layered system of skills, workflows, and validation tools that progressively extends what engineers can delegate to AI without losing control of production
quality. Tomasz and Pedro walk through the real initiatives
powering this at Bolt from the Analytics Engineering angle.

These days, everyone is talking about Agentic AI and how to create entirely new apps with just a few prompts. We are living the AI hype. I’ve stopped counting how many apps have proudly launched AI assistants and other features nobody ever uses.
In this talk, I’ll walk you through how we at Wise scale our product with AI. I will cover our journey from discovery to experimentation and delivery, showing how we identify the biggest opportunities and manage the risks around them. You will see a number of real examples where we apply this technology to build money without borders for more than 15 million customers.

AI is turning Product Managers into builders, making engineers dramatically faster, and enabling people far beyond product teams to create software. But while output is exploding, customer outcomes and P&L aren’t automatically following. Based on firsthand conversations with product leaders, I’ll explore what meaningful AI adoption looks like beyond simply producing more, faster. The real opportunity is to collapse the distance between an idea and learning, and get dramatically better at finding out what works. And we’ll explore what all of this means for product organisations, Product Managers, and the skills worth building now.

As an online grocery store, we often wonder what happens in the exact moments people make a decision. What makes you spontaneously add something to the shopping cart? Why do you decide to buy ingredients for three dinners instead of four? To get closer to what happens in these moments, we combined diary studies with interviews. Participants recorded short video diaries after each shopping trip. Later, during the interviews, we could revisit specific moments and trips they had shared with us, using them to better understand not just what happened, but why.

This presentation is about why we needed this specific method, what made the diary study actually work (as a team, we all had a few unsuccessful diary attempts under our belts), and the things we definitely hope to do better next time.

Ask an AI a hard business question and you get an answer that sounds fluent and certain, and that nobody can check. It might even be right. You have no way of knowing, and no company can sign off on a decision on that basis.

This talk follows one business question from the moment it is asked to the moment someone is willing to put their name on the answer, building up the engineering that has to sit around the model: getting to the right data, checking generated queries before they run, holding the output to business rules the model knows nothing about, making every number traceable back to where it came from, and treating approval as part of the workflow. Each layer shows up for the same reason it does in real projects, because the previous one failed.

That changes the question. Not whether AI can be trusted, but how much confidence we have in this one result, and on what grounds. Measured confidence then sets how much review is needed and where autonomy stops.

You will leave with a working picture of how these systems are built, a way to match autonomy to the cost of being wrong, and language for discussing AI risk in numbers instead of adjectives. With a live demo, run end to end.

The model is not the system. Reliability comes from what you build around it, and from what happens when it turns out to be wrong.

You’ve spent the last two days learning how to build better AI systems. Give me 45 minutes to show you how attackers break them.
This is not another talk about how to build a secure AI agent. Instead, we’ll look at AI applications through the eyes of an attacker: what we look for, how we turn seemingly harmless behavior into attack primitives, and how we chain them into real impact.
We’ll go beyond the familiar “ignore previous instructions” examples and explore how modern AI applications can be attacked without direct user input, how agents can be socially engineered into performing legitimate actions for malicious purposes, and why agent guardrails should never be treated as security boundaries.
Through practical demonstrations and real-world offensive security research, including CVE-2026-25724, we’ll see what happens when AI-native attacks meet traditional application vulnerabilities, permissions, tools, and trust boundaries.
The goal is not to give you another security checklist. It is to give you the attacker’s mindset, so the next time you design an AI system, you can look at it not only as its builder, but also as the person trying to break it.

Football has always trusted the eye test, experience and instinct, but those are no longer enough on their own. From scouting a left-back in Scandinavia to analysing why a team struggles to progress the ball, data is now involved in almost every football decision. Recruitment teams can search thousands of players in seconds, analysts can break down matches in incredible detail, and coaches can access insights that were almost impossible to produce a decade ago. AI is adding another layer, helping football departments find patterns, filter information and turn complex data into something useful. More information, however, does not automatically mean better decisions. Modern football is still played by humans, there are just a lot more algorithms watching now.

On paper a voice agent is three boxes in a row: speech-to-text, an LLM, text-to-speech. Build one and you find out how much sits between those boxes. The agent cuts people off mid-sentence, or leaves a two-second gap that makes it sound broken, or answers a question the caller was still in the middle of asking. Getting a machine to hold a conversation that feels natural turns out to be a problem of turn detection, utterance boundaries, and milliseconds.

This talk goes through where those milliseconds actually go and how you get them back. We’ll walk the latency budget of a live agent end to end, look at the architectural decisions that buy speed, and be honest about what each one costs you in quality. We’ll also compare the cascaded speech-to-text > LLM > text-to-speech pipeline against speech-to-speech models: what each approach is good at, where each falls down, and how to decide which one your use case needs.

Along the way we’ll look at real agents built for very different jobs, and why designs that solve the same problem end up looking nothing alike. Examples come from ElevenLabs agents, but the constraints are the same whatever stack you’re on.

What I used to write about a product was the tip of an iceberg – fifteen percent of what I knew, standing on an abundance of hard-won experience and muscle memory I could not have explained if asked.

Writing with AI inverted that: what I produce is now deeper than what is in my head, and I can no longer stand behind every sentence in it. In product, documents are not records of decisions – writing them is how we decide. AI broke that proxy. What do we trust instead? I’ll show you where it helped, where it quietly hurt, and what I do differently now.

Mi történik, amikor mindenki kap egy digitális ügyvédbojtárt, és a hivatalokhoz ötször annyi beadvány érkezik? Demokratizálódik az ügyintézés, vagy csak a chatbotok kezdik el egymást vegzálni? Az AI mint adminisztratív mankó, panaszgép és a jövő végtelen várólistájának fő beszállítója ebben az adásban kibeszélésre kerül.

We tend to treat time as the ultimate neutral baseline—a shared, objective metric ticked out in seconds, sprints, and release cycles. But the reality is that time is deeply subjective, woven into the cultural DNA of how we structure our lives, build trust, and deliver work. In one culture, an agreed-upon deadline is a rigid contract; in another, it is an opening bid in a relational, evolving process. When teams operate across borders, these fundamental differences in perception quietly govern everything from how agreements are reached to how feedback lands. What looks like decisive momentum to one leader reads as reckless aggression to their team; what feels like thorough alignment to one engineer looks like paralysis to another.

In the fast-moving world of frontier technology and AI, these invisible fault lines rarely stay confined to geography. They multiply at the intersection of disciplines. AI researchers, product managers, UX designers, and executive stakeholders often speak fundamentally different languages, each operating on their own internal clocks, value systems, and definitions of success. When friction occurs, leaders often mistake systemic cultural or domain misalignments for competence or execution issues. The result is stalled roadmaps, frayed trust, and squandered potential.

This session cuts through the myth of the “one-size-fits-all” leadership playbook. Drawing on field-tested frameworks of cultural and cognitive dimensions, we will unpack how time, authority, and communication actually function inside distributed, cross-functional teams. You will learn to recognize the subtle cues that signal disconnect before it turns into deadlock, bridge the semantic gap between deep tech and human-centered design, and transform latent friction into a powerful driver of momentum. Leading global tech teams isn’t about flattening differences, it’s about learning to hear the distinct rhythms your teams run on, and harmonizing them to drive impactful outcomes.

In this talk, we’ll uncover the structural, cultural, and execution patterns that separate high-performing data organizations from mediocre ones. Drawing on real-world experiences across small tech startups, legacy corporations, and Big Tech giants like Meta, I’ll share what actually drives engineering impact – and what non-Big Tech teams can start adopting tomorrow.

Feeding massive amounts of data into an LLM does not magically create an agent that understands your environment. Passive retrieval, like standard RAG, often breaks down when dealing with complex, time-sensitive enterprise workflows. In this talk, we’ll open the hood on active context synthesis. We will explore how to structure data pipelines using engineering patterns like Lambda Architecture to build scalable, resilient AI memory that truly understands state, time, and scale.

AI innovation is everywhere—but turning it into profitable growth is much harder.

As software evolves from helping people do their jobs to increasingly performing tasks autonomously, companies need to rethink not only what they build, but how they package, price and monetize it. Traditional seat-based SaaS models are coming under pressure, while new approaches based on usage, credits, agents and outcomes are emerging.

Which AI innovations should companies monetize—and how? When should AI be included, packaged as a premium capability, or charged based on consumption or outcomes? And how can companies capture the value they create while keeping pricing simple and predictable for customers?

This session explores how software leaders can turn AI innovation into sustainable revenue—from identifying customer value to choosing the right monetization model and scaling it successfully.

The release of Claude Opus 4.5 in November 2025 marked a turning point for the design industry. The cheapest way to imagine a new interface used to be to paint it first (and build it later). AI agents are now so capable at coding that it’s becoming cheaper to explore new ideas by building them than by painting pixels on a canvas. With that, the design process, the loop of exploring possibilities and refining what works, is moving from pixels to code.

This shift is transforming how we work, and it’s going to get uncomfortable for a lot of us. But as the cost of trying ideas falls, there’s more room for design to be distinct and expressive. We can explore interactions, experiences, and expressions of a brand that once cost too much to pursue. We can build our own tooling and throw it away when it’s done its job. The era of design centered on pushing pixels is coming to an end—but design itself is becoming broader and deeper than ever.

Eduard Ruzga spent years as an engineer at Prezi before starting Desktop Commander, an open-source tool that lets AI work directly on your computer. His first attempt flopped in 2023. The second found users, then funding. But when the AI platforms underneath you ship something new every month, is anything you’ve built actually defensible? We talk about searching for product-market fit in a market that won’t sit still, when to pivot and when to hold, and what an engineer has to unlearn to become a founder.