Everyone is building with AI, but few are deploying it successfully. Join Husayn Kassai, Founder & CEO of Ollo, for a practical look at why so many AI initiatives stall: unclear decision-making, weak ownership, and bots bolted onto workflows that are already full of friction.

Husayn introduces the AI "harness": the organisational layer of judgment, governance, context and tooling that turns what models can do into real business impact. You'll learn how to define what good looks like, who should own AI-assisted decisions, and how to give systems the access and context they need.

Attendees will leave with a framework for choosing between generalist tools, specialist products and custom workflows. They'll also see why the ability to deploy AI must be built in-house. Successful AI isn't model-led; it's organisation-led.

Everyone is building with AI, but few are deploying it successfully. Join Husayn Kassai, Founder & CEO of Ollo, for a practical look at why so many AI initiatives stall: unclear decision-making, weak ownership, and bots bolted onto workflows that are already full of friction.

Husayn introduces the AI "harness": the organisational layer of judgment, governance, context and tooling that turns what models can do into real business impact. You'll learn how to define what good looks like, who should own AI-assisted decisions, and how to give systems the access and context they need.

Attendees will leave with a framework for choosing between generalist tools, specialist products and custom workflows. They'll also see why the ability to deploy AI must be built in-house. Successful AI isn't model-led; it's organisation-led.

Every engineer using Claude Code or Copilot generates a session full of hard-won context: gotchas, conventions, decisions. Today, that knowledge evaporates the moment the session ends, and the next engineer rediscovers it the hard way.

This talk shares an early but promising system for capturing that context and turning it into shared team memory. Sessions are logged centrally, then processed into durable, tagged knowledge fed back into future sessions.

Filipe will cover what's working, what's still limited, and the harder problems ahead: keeping extracted knowledge accurate rather than confidently wrong, who owns retiring stale entries, and stopping shared memory becoming another thing to maintain. A practitioner's account from an engineering manager building this now, not a finished product pitch.

Key takeaways

• A working pattern for capturing team knowledge from AI coding sessions

• Why "most frequent" metrics hide the moments that matter most

• The real bottleneck: extraction quality, not storage

More about Filipe: https://alpomar.dev

Agentic AI moves the human decision earlier and compresses the time to make it, raising the bar for what counts as judgement rather than reflex. The question now is whether the person still in the loop can see clearly, hold their position under speed and pressure, and exercise authority they can prove. This session sets out what agentic AI requires of human judgement now: decisions tested under time compression, evidenced when they matter, and owned by a named person who can show they held. Humans still decide. The proof required to show that judgement holds is what has changed. Key takeaways:

• Judgement counts as authority only when it has been tested under real pressure and time compression.

• The human role in agentic systems moves earlier: judgement now happens before execution.

• What has changed is the proof required: judgement must be shown to hold, not assumed because a human was present.

In this session, I will discuss how different responsible AI techniques can be leveraged to evaluate the performance, reliability, and limitations of Large Language Models (LLMs). By systematically varying prompts, we uncover model behavior across various tasks, including reasoning, summarization, and code generation. We’ll discuss benchmarking methods, prompt sensitivity, and how to design effective evaluation frameworks.

The Neural Forest digital art project introduces a pioneering "phygital" approach that materialises digital concepts into tangible, premium video sculptures. While advanced generative AI tools expand the limits of imagination, their visual art outputs typically remain trapped as fleeting pixels on a screen. This project bridges that virtual-physical divide, transforming fluid digital files into high-value, physical art assets.

The production pipeline seamlessly moves from initial concept to digital fabrication. In a collaboration between human and machine, 3D mesh modelling is used, continuing through industrial 3D printing, and high-gloss chrome plating. And raw organic structures are transformed into reflective chrome-finished sculptures. These physical forms securely cradle LCD screens running custom video loop artworks made from a unique combination of photography, AI, and datamoshing.

Looking ahead, Neural Forest heralds the transition towards a metamorphic engine that bridges digital concepts and tangible reality. It points to a future of pure creation, where machines will materialise ideas into diverse physical forms, ranging from synthetic compounds, and eventually, the human form itself. This presentation demonstrates a sophisticated pipeline where AI and nature converge, offering an exclusive look at the future of high-end, tangible digital art.

How do you evaluate an AI system when no suitable benchmark exists and the data you need is difficult to obtain?

This talk follows the construction of a semi-synthetic benchmark for detecting errors in US medical bills. Starting with public healthcare data, we reconstructed clinical encounters, incorporated hospital pricing, and introduced controlled billing errors. Each step required decisions about realism, coverage, and what “good performance” should mean.

Metric design proved just as important as the dataset; catching ten small billing errors while missing one expensive error can reward a model without helping the patient. We designed metrics to better capture patient financial utility, examined performance across models, and identified limitations that still require expert validation.

Takeaways:

• Build evaluation data through grounded synthesis, when real examples are not available.

• Assess benchmarks for coverage, alignment, realism, and room for improvement.

• Choose metrics that reflect user outcomes, including the consequences of false positives.

• Recognize the pros and cons of synthetic eval data.

Soon, billions of AI agents may negotiate, buy, sell, hire, code, advise and transact on behalf of people and companies. But before agents can become economic actors, we need to solve a foundational problem: identity.

How will one agent know whether another is legitimate? Who authorised it? What reputation does it carry? What can it access, delegate or commit to?

This talk introduces Know Your Agent (KYA) - a trust and identity layer for the emerging machine economy. It explores how identity, credentials, permissions, provenance and reputation could enable autonomous agents to safely interact with humans, organisations and each other.

The future of AI may not be about knowing your customer. It may be about knowing their agent.