XConf APAC | Singapore

To be announced

I spent 26 months building a distributed database in Rust with heavy AI help. Three times, AI nearly broke it. Code that compiled, passed tests, benchmarked well, but would have corrupted data in production. Each failure looked like success.

The consensus says AI takes the coding while humans move up to design and testing. It's backwards. AI belongs on the disposable edges. Humans belong in the middle, holding the understanding no test can verify from outside.

This talk is those three near-misses, the research behind them, and Human Replay: the method that keeps the understanding yours.

AI chatbots now handle our banking, healthcare, and support but how do we know they're safe? Unlike traditional software, the same question can get a helpful answer one moment and a wrong, unsafe, or manipulated one the next, and a human tester chatting for an hour barely scratches the surface.

This talk introduces AI SimTest: an engine that uses AI to test AI. It generates hundreds of simulated users, including adversarial ones, runs real conversations against any bot, and judges every response for accuracy, safety, and compliance. We'll show how "AI testing AI" catches failures before real people do.

Multi-agent pipelines fail differently from single agents. One agent gets it wrong and you see it. A pipeline compounds wrong assumptions invisibly — each agent confidently consuming the uncertain output of the last.

This is the contract problem: when one agent's output is another's input, how do you encode what it doesn't know?

We built and deployed this in production. Two agents, one analysing, one transforming, passing a typed contract between them. Every field carries a confidence level. A gating mechanism stops the pipeline when certainty falls below threshold.

Agents don't fail — they guess. Build your contracts like that's true.

AI is changing software delivery, but the real breakthrough is not just making individual developers faster. It is redesigning delivery around AI-native squads.

In this talk, Jessie Xia and Siyuan Li will share how small, high-transparency teams of 3–6 people can deliver business outcomes faster by embedding AI across every role — product engineer, quality manager, and domain expertise. Through real delivery examples, they will show how team structure, ceremonies, and daily workflows change when AI becomes a shared team capability.

The talk will also look ahead to harness-driven agentic delivery, where humans guide intent and quality while autonomous agents assist with execution. Mastering the AI-native squad model today is the foundation for the future of software delivery.

Production incidents are costly, chaotic, and deeply manual. Engineers sift through logs, metrics, configs, and code under pressure, often with no clear starting point. Agentic AI changes that.

In this talk, Vijay presents a live demo of First Responder, an agentic system that investigates production incidents end-to-end. Starting from an alert, automatic triaging and closes the loop with automated fix generation and pull request creation, with engineers staying in control throughout.

Architecture Decision Records often die the moment they are written: approved, archived, and quickly ignored as teams and systems evolve. In the age of AI, that could change.

In this talk, two solution architects show how ADRs can become living, versioned specifications that guide implementation, shape AI-assisted workflows, and support continuous governance across many teams. We explore how documents become machine-consumable context, how architectural intent can be traced into code and review loops, and how AI can help reduce drift. The deeper shift is socio-technical: designing not just systems, but the environment that keeps them aligned.

As AI agents rush into every product roadmap, it’s become dangerously easy to optimise for what AI can do, rather than what people actually need. Many teams are shipping AI-first features that are impressive in demos but fragile in the real world: eroding trust, amplifying bias, or quietly adding cognitive load.

In this talk, I’ll share a practical approach to keeping humans at the centre of AI-first product design. We’ll look at how to frame problems in human terms before choosing models, how to design for consent, control and transparency, and how to use guardrails and hypothesis-led experimentation to balance desirability with safety, accountability and economic impact.

AI factories promise massive compute, but GPUs alone do not create a productive AI platform. The challenge is to make thousands of accelerators, networks, power systems, cooling infrastructure, and facilities operate as one reliable, efficient system.

Drawing on first-hand experience building AI datacenter infrastructure, this talk explores the engineering platform behind the AI factory: Kubernetes, topology-aware orchestration, inference serving, developer workflows, high-resolution telemetry, resilient data pipelines, and power/grid-aware scheduling. We examine the architectural principles and trade-offs required to turn scarce, expensive infrastructure into production AI that is operable, scalable, and trustworthy.

To be announced