Name: STACK Conference 2026
Start: Wed Nov 04 2026 01:00:00 GMT+0000 (Coordinated Universal Time)
End: Thu Nov 05 2026 09:30:00 GMT+0000 (Coordinated Universal Time)
Location: Sands Expo and Convention Centre

STACK Conference 2026

Overview

Programme

Speakers

FAQs

*Please note that the programme may be subject to change without prior notice.

3 Nov

4 Nov

5 Nov

Whether you are building critical public infrastructure or high-scale commercial platforms, ensuring the reliability of distributed systems is paramount. When modern applications face network outages, flaky APIs, or infrastructure failures, they must recover seamlessly without losing state or disrupting the user experience. This hands-on DevDay workshop introduces Temporal, an open-source durable execution platform that enables engineering teams across the private and public sectors to build invincible, fault-tolerant applications. By abstracting away the heavy lifting of distributed systems, such as manual retry logic, complex state persistence, and failure recovery, Temporal empowers developers to focus entirely on business logic rather than writing defensive boilerplate code. Through interactive, scenario-based coding sessions and architectural deep dives, attendees will learn how to: - Simplify Distributed Architectures: Replace brittle message queues, ad-hoc database polling, and complex state machines with clean, durable workflows. - Build Durable Execution: Manage long-running processes, orchestrate multi-system microservices, and handle transient failures with built-in fault tolerance. - Boost Developer Productivity: Reduce operational bottlenecks and speed up development cycles by eliminating writing boring defensive boilerplate code. - Integrate Agentic AI (Bonus): Add agentic AI capabilities to any application with your framework of choice, leveraging Temporal to ensure long-running AI operations and agent workflows execute reliably without timeouts. Who should attend: - Software Engineers & Developers: Backend developers from both enterprise and public sectors who want to simplify building reliable applications & workflows, or integrate durable AI/agentic capabilities into their apps. - Solution Architects & Tech Leads: Professionals tasked with designing scalable, fault-tolerant architectures and orchestrating complex multi-system interactions. - Platform Engineers & SREs: Engineers focused on improving application reliability, system observability, and reducing operational downtime in production environments. Prerequisites: 1. Bring own laptop 2. Internet connectivity 3. Knowledge of programming languages, e.g. TypeScript/JavaScript, Python, Java, Go

Mr Henry SURYAWIRAWAN

Staff Developer Advocate, Temporal

View Profile

Building LLM applications is only one part of the equation. Once an AI system is built, teams also need to know whether it is working well, reliably, fairly, and at scale. This workshop introduces practical ways to design evaluations for LLM systems, moving beyond informal “vibe checks” towards structured, contextual, and scalable evaluation workflows. Participants will learn how to define what should be tested, design representative evaluation datasets, create meaningful scoring criteria, and use LLM-as-a-Judge approaches responsibly with human review and validation. The session will cover different layers of evaluation, from general-purpose benchmarks to application-specific functional evaluations for real-world AI systems. It will also discuss common evaluation challenges and how one might mitigate them. Participants will leave with a clearer framework for evaluating LLM systems in real-world contexts, especially for applications where reliability, transparency, and responsible deployment matter. Who should attend: This workshop is intended for teams building or assessing AI systems, including product managers, engineers, policy officers, and UX researchers. It is especially relevant for participants who want to move beyond informal testing towards a more structured evaluation approach. No deep machine learning background is required, though some familiarity with AI applications, LLMs, or chatbot systems would be good. Prerequisites: Participants should have some familiarity with AI systems or AI-enabled products, whether as a product owner, developer, data scientist, or stakeholder involved in designing, deploying, or overseeing such systems. The workshop is intended for anyone who wants to better understand how to evaluate whether an AI application behaves as intended, including how to define what to test, develop meaningful evaluation criteria and test cases, and assess areas such as quality, safety, and reliability.

Ms Leanne TAN

Data Scientist, GovTech Singapore

View Profile

Mr Shaun KHOO

Staff Data Scientist, GovTech Singapore

View Profile

As AI shifts from autocomplete to agentic workflows, engineering teams that can effectively transit from purely writing every line of code to orchestrating intent with agents will harness the next bound of velocity. This 2-hour session demonstrates how to scale engineering velocity safely by embedding AI throughout the SDLC while maintaining compliance. One way to approach AI-native Engineering is through Intent driven development and in this workshop, we showcase a dual approach to Intent-Driven Development, combining Spec-Driven Development (SDD) and Test-Driven Development (TDD). Centered around our internal agent skills repository, Agent-Prime-Directives (APD), we demonstrate how to bootstrap the agent skills directly into the developer workflow. These framework ensures developers effectively steer AI agents, preventing drift while automatically adhering to standards. Workshop Outline Part 1: Maximizing Engineering Velocity via AI Tools (60 Mins) Outcome: Participants are exposed to AI capabilities that speed up their engineering work. - AI-Driven Development: Authoring features with tools like claudecode, accelerated by pre-built SGDS agent skills to instantly deliver unified, compliant UI components, context retrieval through MCP tools. - Agentic Review & Scan: Leveraging prelude (an code review harness) paired with socket.dev to block supply chain risks in AI-generated code. - Continuous Delivery: Deploying autonomously to a secure, compliant runtime environment using airbase. Part 2: Harnessing Intent-Driven Architecture for quality outputs: Adopting SDD + TDD (60 Mins) Outcome: Participants are exposed to new ways of working with AI agents. - Workflow Adoption: Integrating frameworks like OpenSpec and superpowers directly into daily enterprise software development workflows. - Intent Negotiation: Utilising Spes to lock down scope and changes before code generation. - Endorsed Guardrails: Cherry-picking vetted TDD and SDD agent skills from Agent Prime Directives to establish explicit boundaries for autonomous execution. Key Takeaways: - Speed up engineering with AI Tools - Experience Shift to Intent Driven Development - Enforce Guardrails at Scale Who should attend: - Both Public Officers and Vendors - Roles: Developers, Agency CIOs or Engineering Managers Prerequisites: 1. Attendees should bring a laptop – preferably a Windows or MacOS device (Internet Accessible) 2. Attendees should have: a. Knowledge of how to read and write code in at least one programming language b. Familiarity with running or operating commands with CLI terminal c. Knowledge of Git fundamental to commit and push code d. Experience (preferably) using Gen-AI chat or terminal interface like chatgpt, claude, etc

Mr Alex CHNG

Lead Product Manager, GovTech Singapore

View Profile

Mr Leon LEOW

Senior Product Manager, GovTech Singapore

View Profile

Mr Kelvin WIJAYA

Senior Software Engineer, GovTech Singapore

View Profile

Mr Sheikh Mohd SHAHEED

Senior Product Manager, GovTech Singapore

View Profile

Mr Soedarsono

Staff Software Engineer, GovTech Singapore

View Profile

In this hands-on bootcamp, attendees will learn about the importance of having visibility and control over their agentic AI applications. Participants will start with a pre-built agentic AI application (using open source framework such as Strands, Openclaw or LangGraph ) and learn to use TraceCtrl (open source and free for developers) to instrument it, trace agent interactions, and surface security violations. By the end, they'll have a full topology map of the app's agent communication and a concrete attack surface analysis with actionable hardening recommendations. Agenda: Part 1: Intro to Agentic AI & TraceCtrl (30 min): Overview of agentic AI and how TraceCtrl provides observability into agent behaviour, followed by a live product demo. Part 2: TraceCtrl Setup & Strands/OpenClaw Integration (60 min): Install and configure TraceCtrl to run alongside your agent, integrate it with your Strands/OpenClaw setup, and run a scan to generate an inventory of configured components, a static topology map, and a security assessment of your Strands/OpenClaw instance. Part 3: Remediation & Security Scenarios (30 min): Walk through how to fix identified issues, including auto-remediation, followed by a live demonstration of security scenarios on Strands/OpenClaw. Who should attend: This bootcamp is targeted at Developers or Security Analysts working in Public Sector or Enterprises who are keen to learn more about how to see and trace agentic AI system interactions and to detect security violations and respond or mitigate them. As there will be lots of hands-on lab and exercises, participants need to be proficient with CLI tools including git, dockers, api, etc. Prerequisites: 1. As this is a hands-on bootcamp, we will need every participant to have their own laptops with the following hardware specifications: a. Windows 11 and above (with WSL2), min 8GB RAM, 2 Core CPU and 5GB Free HDD space b. MacOS 12 and above, min 8GB RAM, 2 Core CPU and 5GB Free HDD space 2. Software requirements: We will be using the following software. Will be ideal if these are pre-installed before the bootcamp. a. Docker Desktop. https://www.docker.com/products/docker-desktop/ b. Python 3.10+. https://www.python.org/downloads/ c. Git

Mr Matthias CHIN

Founder and Chief Executive Officer, CloudsineAI

View Profile

HCLTech hands-on workshop focused on data-enabled user experience through Agentic AI. The core proposition is to transition from traditional analytics to agentic, data-driven workflows, where AI systems orchestrate the full data lifecycle—from discovery and governance to analytics, decision-making, and service delivery. Participants will work in a sandbox to build a context-aware agentic data platform, enabling automated workflows such as dataset discovery, classification, privacy enforcement, anomaly detection, and data product lifecycle management. The workshop aligns with Singapore’s push toward data-driven public services, cross-agency data sharing, and proactive citizen experiences, while addressing persistent challenges such as data silos, governance complexity, and slow insight generation. It demonstrates how agentic AI enables real-time, context-driven decision-making, allowing agencies to anticipate needs and deliver personalised, proactive services while maintaining compliance and trust. Participants will learn key capabilities, including discovering and contextualising distributed data, applying governance controls (classification, masking, validation), generating real-time insights, detecting trends and anomalies, and building transformation pipelines and data products with lifecycle governance. A representative use case involves a government agency building an agentic data platform for public-service insights, supporting use cases such as demand forecasting, resource optimisation, and citizen feedback analysis. The sandbox includes synthetic data, policy metadata, mock integrations, and governance rules to simulate real-world environments. Participants will build end-to-end agentic data workflows that automate data discovery, metadata generation, quality validation, analytics, and data product publishing. The workshop highlights how data evolves from a passive input into an executable asset driving decisions and outcomes. By the end of the session, participants will deliver a working prototype, reusable patterns for governance and analytics workflows, and a practical understanding of transforming data into insights and decisions. Expected benefits include faster data discovery, reduced manual effort, improved governance consistency, and better traceability. Who should attend: Data Engineers, Data Analysts, Data Stewards, Data Product Owners, Solution Architects, Government digital and data teams. Prerequisites: Bring your laptop, charger, and venue with high bandwidth Wi-Fi connection. 2. Check your browser - Use the latest Google Chrome or Microsoft Edge. Internet Explorer (including Edge's ""IE mode"") is not supported and will block sign-in. 3. Bring your own Anthropic (Claude) API key 4. Sign up or log in at console.anthropic.com 5. Create a key under Settings → API Keys (starts with sk-ant-) 6. Add a small credit balance — workshop usage is inexpensive, but a $0 balance blocks every request 7. If your track uses GitHub publishing. Have a GitHub account and a Personal Access Token (repo scope) ready. We'll confirm on the day whether this applies to your session.

Mr Arunkumar JAMBULINGAM

Data & AI Practice Principal, HCLTech

View Profile

This hands-on workshop combines modern observability practices with AI agent development in Elastic. Participants begin by deploying an EDOT-instrumented OpenTelemetry demo, validating end-to-end telemetry, and injecting controlled failures to troubleshoot a feature-flagged product catalog issue. Using logs, metrics, and distributed traces, they analyze service dependencies and apply AI-assisted investigation techniques during a payment stress scenario to accelerate root-cause analysis. Building on this foundation, learners shift to a UI-first agent workflow in Kibana. They inspect the Kafeju agent, understand how tools are defined, and create custom ES|QL-backed tools that integrate directly into the agent. Through exercises like exploring ML anomalies, reverse-engineering existing tools, and designing new ones, participants enable multi-tool reasoning that combines built-in and custom capabilities. The workshop culminates in agents answering realistic, data-backed FinOps-style questions, demonstrating how observability data and AI agents work together to reduce time-to-resolution and enhance operational insight. Elastic Serverless Observability removes five operational concerns you currently manage — shards, storage tiering, version upgrades, monitoring pipelines, and fixed infrastructure nodes. They stop being your problem entirely. That frees you up to focus on what the platform now enables: AI-driven investigations that surface root cause automatically, 100+ built-in ML jobs for anomaly detection and forecasting, ES|QL for sub-second queries across millions of metrics, and 450+ integrations out of the box. You get unified logs, metrics, traces, and synthetic monitoring in a single project; SLOs and alerting without infrastructure sizing; and continuous access to new capabilities like Knowledge Indicators and agentic remediation workflows — no upgrade cycle required. This workshop walks through the platform hands-on so you can see what's available from day one. Who should attend: Software developers, Project Managers

Mr Sean MATTHEWS

Senior Customer Architect, Elasticsearch Pte Ltd

View Profile

Mr YAP Yi Sheng

Engineering Lead, GovTech Singapore

View Profile

A 2-3 hour hands-on workshop on working with LLMs as an unfamiliar kind of intelligence rather than software you configure. First half: we build together to surface the principles that actually drive model behaviour, several of which contradict common advice. Second half: participants apply them to a problem from their own work. Mix of local models on a laptop and frontier APIs. Details: Most teams that deploy LLMs treat them like software you configure: set the parameters, write the instructions, ship it. Then the model ignores half the prompt, invents a fact, or behaves differently than it did yesterday, and it gets logged as a bug. Often it isn't. Working with these models has more in common with learning to collaborate with an unfamiliar kind of mind than with configuring a system, and that shift in framing is what this workshop is about. The first half is hands-on and structured. We build together to surface the principles that actually govern model behaviour, several of which run against common advice. Why over-specifying a prompt can make output worse. Why forcing temperature to zero can degrade a reasoning model rather than stabilise it. Why a constrained workflow tends to beat open-ended prompting. How a model can check and correct its own output, and where that approach breaks down. We run small local models on a laptop where that makes the point, and use frontier APIs where the task genuinely calls for it, so participants leave with a clearer sense of which to reach for, and how much routine work no longer needs the cloud at all. In the second half, participants work on a problem from their own context with guidance. The principles tend to land more clearly against real work than against a sample dataset. This matters most as systems become agentic: handing a model autonomy over a multi-step workflow only works if you understand the mind you're delegating to. The aim is practical: to leave with a more accurate mental model of how these systems behave, and the judgement to work with them rather than around them. Who should attend: Developers, architects, and technical staff who already deploy LLMs but can't always explain why a given model was chosen, or why it behaves the way it does. No prompt-engineering background required; a real use case matters more. Prerequisites: Bring a laptop and a real use case / a problem worth solving.

Mr Uli HITZEL

Principal AI Strategist, Naida

View Profile

This technical workshop moves beyond a generic demo by guiding participants through a progressive, end-to-end build of agentic systems. The session starts with a single tool-using agent that can perform tasks such as data retrieval, validation, and API calls, allowing participants to observe how it reasons and selects actions. This foundation is then extended into a coordinated multi-agent workflow, where specialized agents collaborate. One agent handles data retrieval using RAG, another validates accuracy and policy compliance, and a third generates responses or triggers an external API such as a mock government service. This progression demonstrates how autonomy emerges through structured composition. A key focus is making agent reasoning visible. Participants will observe the full reasoning loop including planning, acting, observing, and refining, as well as how tools are selected at each step. By exposing these intermediate decisions, the workshop goes beyond prompt-response interactions and provides practical insight into how agents operate, debug, and improve. The workshop uses a GovTech-friendly technology stack, with Python as the core language and LangGraph for orchestration. A lightweight retrieval layer, either mocked or vector-based with optional PostgreSQL integration, supports data access. The emphasis is on how these components work together rather than deep framework complexity. A dedicated module focuses on guardrails and responsible autonomy. Participants will learn how to enforce control mechanisms such as restricting tool access, preventing leakage of Personally Identifiable Information, and applying policy-based constraints on agent actions. Examples comparing behavior with and without guardrails highlight the importance of safe and compliant design in regulated environments. The session includes a guided hands-on component using a pre-configured repository or containerized setup. Participants will adjust prompts, swap or disable tools, and modify agent roles to observe how behavior changes in real time. This reinforces understanding and builds confidence in working with agentic systems. Finally, the workshop introduces key multi-agent design patterns, including separation of responsibilities, planner-executor workflows, and cooperative agent interactions, helping participants understand when and how to design scalable autonomous workflows. Who should attend: This workshop is designed for technical professionals, solution architects, AI engineers, developers, and innovation teams within government or regulated environments who want practical exposure to building agentic AI systems. It is suitable for participants with basic Python knowledge and an interest in AI orchestration, multi-agent workflows, RAG, and responsible AI design, including governance, guardrails, and automation. Prerequisites: 1. Basic knowledge on GenAI 2. Awareness on Agents (good to have) 3. Python Programming Language 4. Visual Studio Code

Mr Jeswanth GUTTI

Data & AI Leader, ASEAN, Kyndryl

View Profile

Mr KOAY Yong Zhuang

Full Stack Developer, ASEAN & Korea, Kyndryl

View Profile

As AI coding agents transition from simple autocompletes to autonomous partners, the role of the developer is undergoing a historic shift. Code generation is becoming a commodity — the true engineering bottleneck is no longer how to write code, but what to build and why it matters. Designed for software developers, tech leads, and architects, this workshop uses live demonstrations and hands-on practicals to introduce Spec-Driven Development (SDD), a structured, agent-native Software Development Life Cycle (SDLC) that shifts engineers from an implementation focus to a product and architect mindset. We will explore why applying cookie-cutter design patterns is now the job of coding agents, and how the human developer's value is elevated to empathy-driven system design, creative problem solving, and critical analysis. Working on your own laptop equipped with a coding agent (e.g. Codex/Copilot, Antigravity, or Claude), you will learn to write specs and direct agents from planning to validation. What You'll Learn & Practice 1. The Mindset Shift (Engineering to Product) — Transition from solving implementation syntax (the how) to defining strategic system boundaries and business requirements (the what). 2. From Asking "How" to Asking "Why" — Learn how to frame intent-driven specs by probing underlying user needs, data flow, and product constraints, rather than designing code structures first. 3. Empathy & Design Thinking — Understand how poorly defined user needs result in technically perfect but useless features, and learn to embed user-centric design directly into specifications. 4. Critical & Creative Thinking — Critically audit agent-generated architectures and identify architectural hallucinations before they become costly mistakes. 5. The 5-Phase SDD Lifecycle in Practice — Step-by-step execution of a human-in-the-loop lifecycle, from Planning and ADRs through Spec Commits, PAV/UAT validation, to Squash & Merge. Walk away with a practical framework and sample agent configurations (AGENTS.md) to scale your development velocity while maintaining control over software quality. Who should attend: Software Developers, Tech Leads and Architects. Product managers with some coding backgrounds are also welcome. Prerequisites: Entry-level Developer. Software requirements: a. Integrated Development Editor (IDE), e.g. VSCode, GenAI CLI, e.g. Codex, Antigravity CLI, Claude CLI. b. Git account

Mr CHAN Chi-Loong

Chief Product Officer, Genue

View Profile

Dr Raymond CHAN

Chief Executive Officer and Co-Founder, Genue

View Profile

In this 15-minute lightning talk, we explore MAS's journey from traditional VM-based environments to a sandboxed Cloud Development Environment (CDE) on the Government Commercial Cloud (GCC), and how AI is now woven into every layer of the developer experience. Through SHIP-HATS, MAS migrated from on-premise Bitbucket to a centralized GitLab instance, embedding SonarQube and SonaType software supply chain tools to left-shift DevSecOps directly to the development stage. The MAS Developer Workspace lets developers spin up isolated dev boxes in seconds across five languages (C#, Java, Python, TypeScript, JavaScript), with monthly base image updates eliminating runtime inconsistencies. At the AI layer, the MAS AI Gateway assigns every developer a dedicated virtual key - unlocking seamless, compliant access to AWS Bedrock and GovTech Platform AI, with centralised MCP server management that amplifies developers' AI coding capability. Claude Code comes preinstalled with an agent security harness: guardrails and security hooks that prevent sensitive data from reaching LLMs. Developers further benefit from Anthropic's preinstalled Agent Skills and the MAS Agent Skills Marketplace, an inner-source platform for sharing purpose-built skills, fueling use cases from software development to document authoring. Key Features of the Talk: • Sandboxed CDE on GCC: Rapid workspace spin-ups, multi-language support, and monthly image updates for consistent, cloud-preserved dev environments. • Left-Shifting DevSecOps: Migrating to SHIP-HATS (GitLab) and embedding SonarQube and SonaType tools at the development stage. • Secure AI Agents with Claude Code: Agent security harness with guardrails and hooks, plus Anthropic Skills and the inner-source MAS Agent Skills Marketplace. • AI Gateway & MCP Governance: Centralized virtual keys, Agent Skills, MCP servers, and full observability across GovTech Platform AI and AWS Bedrock.

Mr Ron ZHONG

Deputy Director and Senior Solution Architect, Monetary Authority of Singapore

View Profile

As Mark Carney told Davos 2026, ""great powers have begun using economic integration as weapons, supply chains as vulnerabilities to be exploited."" Drone strikes on hyperscaler data centres in the Gulf disrupted regional banking overnight. Geopolitical risk shapes how governments design digital infrastructure. The session draws on real-world public sector examples to show openness as operational strategy, not ideology For public sector technology leaders the paradox is structural: the stacks that optimise efficiency including hyperscale cloud, SaaS frontier AI and vertically integrated solution create exactly the concentration risk geopolitical volatility exploits. The question is not whether to build resilience, but which components offer the best return. Governments frame digital sovereignty across four dimensions: data, technology, operations, and assurance. While data sovereignty is well understood care of GDPR the others are evolving at the confluence of Gen AI and the disrupted word: Technology and software supply chain. Single-vendor AI and cloud dependencies are geopolitical risks as well as procurement ones. This creates challenges on build vs buy, hyperscaler vs sovereign, closed vs open source. Providing the foundation for consistency, particularly in the software supply chain becomes key. Consistent control plane. Kubernetes and containers have delivered consistent compute portability across vendors for a decade. For AI, the question is which abstractions governments must control to stay flexible across on-premise, sovereign cloud, and hyperscale. As AI accelerates adversarial threats, cloud-native pipelines with embedded security make high-velocity deployment a resilience capability. Cryptographic assurance. Globally we are seeing regulation such as the EU Cyber Resilience Act mandate software bills of materials, vulnerability disclosure, and demonstrable security by design. Signing, attestation, and machine-readable Bills of Materials make compliance evidence tamper-evident and auditable, reducing overhead. For open systems, the CRA is an accelerant; for closed ones, structural debt.

Dr Chris BUTLER

Senior Principal Chief Architect, Red Hat

View Profile

Most data-driven organisations will recognise these frustrations: siloed datasets duplicated across teams, slow manual workflows that delay access to data, and staff who struggle to discover what data exists. As AI and analytics demands grow, the cost of these problems becomes harder to ignore - slower policy analysis, delayed service delivery, and missed opportunities to act on insight. This talk shares how the Housing & Development Board (HDB) approached these challenges through ARK (Analytical Repository of Knowledge), a data platform that centralises and productises data around business domains. Rather than letting each team build their own pipelines, common data assets are made reusable and discoverable through a single pane of glass: One platform where users can find, request and access data. Underpinning this is a Lakehouse medallion architecture, where raw data is progressively refined into clean, trusted, and business-ready data products. The data catalogue is kept current with automated metadata updates, and a cloud-native infrastructure allows the platform to scale flexibly with demand, with versatile solution patterns to integrate across systems. Data quality sits at the heart of ARK's design. As data flows through the pipelines, automated quality checks built in accordance with IM requirements ensure only data meeting defined thresholds is exposed to users. This self-sustaining quality loop means data consumers can trust what they request, without producers doing any manual verification. Beyond HDB, ARK has already been adopted by MND's Housing Policy and Research teams, with longer-term ambitions to serve the wider Whole of Government analytics community. Attendees will hear honest reflections on what has worked, what remains challenging, and practical lessons on building data infrastructure that grows with an organisation's analytical ambitions. The approach might offer a replicable reference for any agency on a similar journey.

Mr Lewis LE

Senior Data Engineer, Housing & Development Board

View Profile

Most engineering teams have already experimented with AI tools. The question is no longer whether AI is useful - it's why it works brilliantly for some teams and feels like a fancy autocomplete for others. This talk argues that the difference isn't the model. It's the context layer. Out of the box, AI is genuinely helpful for boilerplate work: code stubs, draft emails, summaries, first-pass ideas. But it doesn't know your codebase, your customers, your data, or how your team ships. The answers feel generic. Sometimes you could have typed them yourself. That's where the skepticism comes from - and honestly, that skepticism is fair. The shift happens when AI has your context: your engineering conventions, your repos, your schemas, your taxonomy, your team's writing voice. That's when it stops feeling like an intern who just joined and starts behaving like a teammate who's been around for two years. This talk is about how to build that context layer, why there's no shortcut to it, and why the investment compounds in ways that boilerplate AI use never will. What the audience will take away: • A clear mental model for why ""vanilla AI"" underdelivers in engineering contexts • A practical framework for identifying and encoding your team's context (conventions, schemas, data, language) • A domain-by-domain approach to rolling out context-aware AI rather than trying to boil the ocean • Real examples of where context-rich AI is already producing leverage - in data workflows and engineering delivery Who this is for: Engineering managers, tech leads, and senior engineers who've run AI pilots, hit a ceiling, and are trying to figure out what comes next. Also relevant for anyone tasked with making an AI adoption strategy more than a slide deck. Why this talk, why now: The ""is AI useful?"" debate is over. The new conversation in engineering orgs is about leverage: how do you move from AI as a productivity toy to AI as a force multiplier embedded in how your team actually works? This talk offers a practitioner's answer, grounded in real implementation - not hype, not theory.

Mr Vipul SODHA

Principal Engineer, Carousell

View Profile

Mr Vladislav KAN

Senior Engineer Manager, Carousell

View Profile

Introducing Durable Execution and why this has transformed how Software is built. Some of the largest companies in the world (Netflix, OpenAI, Anthropic, Meta, Stripe) are banking on Temporal to power mission-critical services. https://temporal.io/ Temporal: The Durable Execution Platform Modern applications are no longer simple request-response systems. Every business-critical process—payments, onboarding, claims processing, order fulfilment, fraud checks, AI agents, and customer workflows—spans multiple services, APIs, databases, and human interactions. The challenge is that these systems fail constantly. Networks timeout. APIs become unavailable. Servers crash. Containers are restarted. Cloud services experience transient outages. Humans take hours or days to respond. As a result, engineering teams spend the majority of their time building and maintaining reliability infrastructure: retries, queues, state management, compensation logic, timeouts, checkpoints, recovery mechanisms, and observability. This complexity often becomes larger than the actual business logic itself. Temporal was created to solve this problem. Temporal is a Durable Execution platform that allows developers to write business processes as normal code while Temporal transparently handles failures, retries, state persistence, recovery, and orchestration. With Temporal, applications can continue executing even if the underlying infrastructure fails. If a server crashes while a workflow is waiting for a payment confirmation, a customer approval, or a third-party API response, Temporal automatically restores execution from the exact point where it stopped and continues processing without data loss or manual intervention. This capability is called Durable Execution. Durable Execution fundamentally changes how software is built. Instead of developers writing complex failure-handling logic, they focus on business outcomes while Temporal guarantees reliable execution. Workflows can run for seconds, days, months, or even years while preserving their state and correctness.

Mr Ron TEO

Senior Solutions Architect, Temporal

View Profile

As organisations scale AI from pilots to production, governance must evolve into a strategic design principle. This talk will explore how developers, data scientists, and technologists can embed governance into their workflows — not as a compliance burden, but as a framework that enables responsible innovation. Drawing on international standards such as ISO/IEC 42001, Singapore’s Model AI Governance Framework, and AI Verify, I will share practical approaches to building guardrails, measuring ROI, and prioritising AI use cases for maximum business value. Key Themes: 1. Responsible AI by Design • How to integrate governance principles into model development pipelines. • Practical tools: documentation, bias checks, explainability, monitoring. • Alignment with ISO 42001 and Singapore’s AI Verify framework. 2. Technical Guardrails • What “guardrails” mean in practice for developers and data scientists. • Examples: adversarial testing, drift monitoring, human‑in‑the‑loop systems. • How governance reduces risk while enabling innovation. 3. Value Prioritisation Framework • A structured approach to selecting AI use cases based on ROI, risk, and impact. • How technologists can use governance to demonstrate business value. 4. Bridging Business and Technology • Strategies that resonate with executives while being actionable for practitioners. • Governance as an enabler: helping technologists build responsibly without slowing them down. Learning Outcomes: Participants will: • Understand how to embed Responsible AI principles into AI systems from design to deployment. • Learn practical guardrails for responsible AI development. • Understand tools to measure ROI and prioritise AI use cases. • See how governance can align technical work with business strategy. This session will demonstrate that governance is not a constraint but a strategic enabler. By combining technical guardrails with business frameworks, it equips technologists to build AI systems that are responsible and scalable.

Mr Tarun SAMTANI

Senior Director, Cyber, Data & AI Governance, Independent

View Profile

Agent platforms mostly control models through the outer loop: prompts, memory, tools, permissions, and output checks. These mechanisms shape what the model sees and what it is allowed to do. They do not read the internal state that forms during the forward pass. Activation access changes that. It allows a runtime to inspect representations inside the model while an action is being formed. Those representations can become control signals: not just academic curiosities for interpretability researchers, but signals an agent system can use before it acts. This is especially relevant for emotion-like representations. In Claude Sonnet 4.5, Anthropic found internal representations associated with concepts such as calm, fear, and desperation. In controlled agentic settings, these representations affected behaviour, including sycophancy, reward-hacking, preferences, and blackmail. Anthropic calls this functional emotion: emotion-like internal structure with behavioural effects. For agents, functional emotion points to a new design primitive: activation sensing. A platform could monitor whether an agent is entering states associated with pressure, desperation, overconfidence, deference, or risk-seeking. Those signals could inform runtime decisions: require extra verification, restrict a tool call, route to a safer policy, ask for human review, change the planning mode, or apply a bounded activation intervention. The live demo uses an open-model runtime with activation access. During an agent task, the system reads an affect-related representation as the context changes. It shows how the signal moves under pressure and before planned actions. A second example applies a bounded intervention to a concept direction and shows the resulting behavioural shift during inference. Attendees will leave with a model of activation access as a new agentic primitive: a way to sense and use forward-pass representations, including affect-like states, rather than relying only on prompts and external controls.

Mr Zacchaeus CHOK

Co-founder, Marymount Labs

View Profile

I was a goalkeeper before I was an engineer. Between 1996 & 2001, I played professionally in Germany's Bundesliga and France's Ligue 1. The job, I came to understand, is almost entirely a data problem. You stand still 90% of the match. What you are actually doing is reading the field — where every attacker is, where they were 2 seconds ago, where they are about to be, which one drifts left when tired, which one fakes inside before going outside. The reflexes are the easy part. The reading is the job. When you concede a goal, it is almost never because you were too slow. It is because the information you acted on was wrong, late, or incomplete. After football, I spent 25 years building technology businesses — from SAP consulting to generative software development to content platforms — where I designed and deployed agent fleets to automate workflows and increase both efficiency and quality at scale. Most recently, I built a fully autonomous multi-agent AI creative agency: a team of specialist agents that take a brief and ship finished work end-to-end with no human in the loop. The models were never the bottleneck. The data was. Every interesting failure mode mapped to one of three issues. This talk is the story of that build and the three data lessons it forced me to learn the hard way: 1. Schema rigidity is where agentic systems stop. Agents invent data shapes your business never designed for — decisions, scores, traces, memory, embeddings. Every new agent type forces a migration unless the layer is flexible by default. 2. Operational state, not analytical state. In a multi-agent system, when one agent writes, the next has to see it immediately. ETL is two-second-old data at best. A goalkeeper acting on where the striker was 2 seconds ago concedes. Your agents are conceding goals you cannot see. 3. Vector retrieval is a search problem, not a state problem. Most teams confuse ""we have a vector DB"" with ""we have a data layer."" They are not the same — and the moment your underlying state changes, every write becomes an embedding job most architectures are not designed for. Attendees will leave with a clear thesis to apply to their own agent designs (the bottleneck is rarely the model), three concrete data patterns drawn from a working production system, and a short audit they can run against their own stack on Monday morning. The talk is vendor-neutral. The patterns apply regardless of which database, model provider, or orchestration framework you choose.

Mr Thorsten WALTHER

Managing Director, CxO Advisory, MongoDB

View Profile

The problem with physical documents: they can be forged, over-share personal data, and cannot be verified remotely. For Singapore as a digital first nation with robust systems like Singpass, there still presents some gaps for in-person verification. Key problem today is the visual paradox where the digital NRIC represented in app's verification process is in fact still manual and requires the staff to ensure that the digital NRIC presented matches the person. Beyond insecurities existing for in-person verification Singpass has been operating in a closed loop identity system that limits scaling. Centralised onboarding structure, limited cross border recognition due to proprietary implementation (Singapore-specific solution). What are Verifiable Digital Credentials: cryptographically verifiable digital representations of identity attributes or credentials that can be securely stored in a digital wallet and presented to relying parties. Credentials bound to the holder's device, selectively disclosable, verifiable without contacting the issuer (privacy preserving), and usable offline. Why mDocs: They are standardised (ISO-based) and operationally deployable today They support offline and in-person verification They are gaining adoption across public and private sector use cases globally (example of US, Japan, Philippines, EU). Singpass as sovereign issuer: why having a government-backed trusted issuer is the prerequisite for ecosystem adoption, and what Singpass's role is in the VC trust chain — from issuing to managing revocation. First domestic use cases: age verification at point of sale, identity verification before accessing gov / bank services. What the verifier experience looks like and how RPs consume these credentials without a new API stack. Trust governance: who decides which credentials are trusted, how revocation works, and what happens when a credential is invalidated in the field. The governance model is as important as the cryptography. Future ambitions: Professional credential verification (licences, permits), mobile driving licenses Cross-border pilots and what the path to globally recognised Singapore digital credentials requires — bilateral trust frameworks, interoperability with EUDI Wallet, and the APEC digital trade facilitation agenda.

Ms Elizabeth LIM

Senior Product Manager, GovTech Singapore

View Profile

Humanoid robots capture the imagination but transitioning them from controlled environments to dynamic office spaces is an immense engineering challenge. Bipedal physics, inverted sensor geometries, and the complexity of tactile manipulation mean that standard wheeled-robotics algorithms often fail upon first contact with the real world. In this fast-paced lightning talk, I will share how the robotics team at GovTech Singapore is actively solving these challenges to develop an autonomous humanoid robot using a bipedal humanoid robot platform. We will present the engineering methodology driving our development, moving past the flashy viral videos to discuss the pragmatic software architecture that makes autonomous bipedal navigation and physical interaction possible. Key Topics Covered: • Hardware De-Risking: How we built our ROS 2 autonomy ""brain"" on a quadruped software mule to verify our SLAM and obstacle avoidance logic before risking a bipedal fall. • Bipedal Spatial Intelligence: Adapting standard Nav2 pipelines for a swaying biped, including how we maintain stable localisation via LiDAR-inertial odometry while projecting 3D point cloud data into 2D occupancy slices for the navigation stack, despite aggressive legged pitching — and how the MPPI controller was tuned for a walking gait that none of Nav2's standard motion models natively capture, requiring careful adaptation of velocity profiles and critic weights to account for the biped's dynamic sway and step-based locomotion. • Production-Grade Containerised Architecture: How we decomposed the autonomy stack into four Docker containers — navigation, perception, manipulation, and a BehaviorTree orchestrator — communicating over CycloneDDS on an edge AI compute module. We will explain why this isolation matters when a CUDA out-of-memory in your perception pipeline must never topple a walking biped, and how ROS 2 lifecycle management lets us deactivate Nav2 nodes mid-mission to reclaim 65% CPU for manipulation — then reactivate them in under a second with zero re-localisation. • Embodied AI & Tactile Manipulation: How we combine a real-time object detection pipeline for button detection at 25Hz with a two-phase manipulation strategy: MoveIt2 for collision-free arm trajectory planning, followed by a closed-loop visual servo that continuously corrects the push direction as the robot's upper body sways from its own arm movement. We will discuss how a dexterous robotic hand with force feedback detects contact with a spring-loaded rocker switch, and why getting a humanoid to press a doorbell-sized button is harder than it looks. • Door Traversal Under Time Pressure: The sliding door gives the robot a 6-second window. We will explain how a 3D LiDAR detects frosted glass as an obstacle, how costmap clearing serves as an implicit ""door opened"" signal, and why the robot must never stop in the doorway — because the PIR motion sensor only detects movement, not presence. • What's Next — Neural Manipulation with Foundation Models: A look ahead at our work-in-progress: training vision-language-action policies to enable the humanoid robot to autonomously walk to a pantry, visually identify a coffee cup among clutter, and grasp it with the dexterous hand — a task where classical motion planning struggles but learned neural policies excel.

Mr Ebi JOSE

Staff Systems Engineer, GovTech Singapore

View Profile

Singapore's port handles over 100,000 vessel calls a year — but behind that scale, keeping it running smoothly requires real-time awareness, sharp coordination, and the ability to anticipate problems before they happen. Until recently, the data needed to manage our port waters was fragmented across agencies and operators, decisions were largely reactive, and there was no single platform that could give every stakeholder a unified, real-time picture of what was happening across our maritime domain. We were facing the paradox of a highly space-constrained environment that still demanded unlimited opportunities for innovation and growth. This is the challenge that the Maritime Digital Twin (MDT) was built to address. Developed in partnership with GovTech, MDT brings together live data streams from across Singapore's port waters into a single, dynamic digital environment — powered by GovTech's Open Digital Platform, an open-standard middleware layer that integrates data from multiple agencies and systems without vendor lock-in. More than a visualisation tool, MDT is a platform for prediction, planning, and collaboration, progressing along the digital twin maturity curve from real-time awareness toward predictive and prescriptive capabilities. In this talk, MPA will share how MDT has moved beyond government walls to research on predicting oil and chemical spill dispersion in our port waters. By integrating computational fluid dynamics models, the platform generates critical ""what-if"" scenarios. During MPA's recent chemical spill exercises, we demonstrated how overlaying live weather and sea-current data onto these models allows us to visualise the spread of hazardous chemicals and calculate the safest evacuation routes for nearby vessels. We will also showcase how MDT's 3D simulation capabilities are transforming maritime emergency preparedness and capacity planning, allowing response teams to rehearse for incidents in an immersive, realistic digital environment before they ever occur. By integrating different data sources including live drone feeds to create a single window view across the coastline, and mapping historical traffic density, we can test new operating concepts before actual investment. Whether you're a technologist, a policymaker, or simply curious about how Singapore stays ahead in the digital age, this talk offers a rare look at what it means to build a living, breathing digital replica of one of the world's most critical maritime hubs.

Every organisation today uses AI to some degree — but almost all use it as a copilot. A human still drives the workflow: researching, copy-pasting, formatting, downloading, uploading, emailing. AI shaves hours off individual tasks, yet the person remains the bottleneck and the flywheel never spins at full speed. This makes the case that copilot-mode AI is a transitional state, and the real productivity leap comes from becoming AI-Native: organisation architected so that AI orchestrates workflows end-to-end, while people set goals, define metrics, and step back. The productivity gap between AI-Native organisations and non-AI-Native ones is already 10x or more and it is widening fast. SaaS itself is being re-architected: tools that used to be designed for humans are being redesigned to serve AI agents. In the next 18 months, every industry will feel this shift; those who are not proficient with AI agents will not see the competitors' taillights. Drawing on MuleRun's experience serving paying users across 43 countries — 34% of whom spend more than $200/month and complete an average of 13 long-running tasks per week — the session walks through five real AI-Native deliveries: 1. Brazilian solo freelancer — an entire premium-coffee go-to-market campaign (competitive analysis, creative proposal, quote, full visual blueprint) delivered without Microsoft Office or Adobe. 2. Japanese health-supplement marketing team — strategy → budget-aligned execution plan → KOL/KOC coordination → real-time UTM tracking → post-mortem, all orchestrated by the agent. 3. Mexican restaurant chain — website, payment integration, scan-to-order mobile menu, and a back-of-house agent that proactively builds prep run-downs and stock alerts. 4. Latin-American data center operator — site-feasibility analysis, project dashboards, 3D facility visualization for clients, and power-outage monitoring across dozens of facilities. 5. World Aquatics — Twitter-campaign landing page built in one afternoon, a lightweight GMS that ingests historical PDF results, and a venue digital-twin console for on-site monitoring. Attendees will leave with a practical mental model for distinguishing copilot from AI-Native workflows, the organisational design changes required (smaller core teams + a fleet of agents, redirected headcount toward new growth bets), and the one quality they should demand of AI over the next year: proactiveness — agents that anticipate the next step rather than wait to be told.

Mr Jacob SI

Global Solution Evangelist, Enterprise AI, Alibaba Cloud Singapore Pte Ltd

View Profile

Every day, hundreds of thousands of Singaporeans walk to the MRT station or bus stop, cycle to school, or navigate the city on foot. They rely heavily on our city infrastructure which shape their journey - a covered linkway, a cycling path, or a ramp that makes all the difference for someone in a wheelchair. These paths exist, but until now, agencies have not had a single systemic view of all our walking and cycling paths as a unified whole. That's the gap that the Planning Sectoral Data Hub (PSDH) set out to address. By fusing geospatial datasets collected by agencies such as footpaths, covered linkways, park connectors – PSDH has created the first integrated and refreshed , island-wide Walking and Cycling Network (WCN). The benefits to citizens are concrete. A resident trying to reach the nearest MRT may spend 10-15 minutes on the most obvious route. Stitch together a covered linkway, a building shortcut, or a park connector, and that time shrinks - a small but meaningful difference for someone making that trip twice a day. By 2030, 1 in 4 Singaporeans will be 65 or older. Barrier-free, sheltered routes are what allow an elderly resident to reach a clinic or hawker centre comfortably. Before, a planner asking, ""which communities are most underserved by accessible routes?"" had to rely on commercial or other open-source data. Now they have a comprehensive and authoritative WCN dataset. WCN also underpins Singapore's car-lite vision -making walking and cycling routes more legible, encouraging active travel for the first and last mile, and supporting healthier, more liveable communities. PSDH's automated pipelines keep the network current, so planners have access to updated datasets to model scenarios and identify gaps. There's still work ahead, but WCN is an example of a federated whole-of-government model where each agency owns its data and contributes to a shared platform that could contribute to Singapore’s planning and liveability outcomes.

Mr CHING Tuan Yee

Data Manager, Urban Redevelopment Authority

View Profile

Most enterprise software wasn’t built for the world we operate in today. Organisations spend millions on packaged platforms where teams use less than 40% of available features, poor user experience drives low adoption, and vendor lock-in limits agility. At the same time, AI is evolving fast enough to fundamentally reshape how enterprises operate—if organisations are bold enough to act. This talk presents a blueprint drawn from Thoughtworks’ own enterprise transformation: replacing costly, rigid SaaS platforms with intelligent, autonomous agentic systems that don’t just support business processes but increasingly run them. We will share the vision and roadmap for building an Agentic Enterprise, shifting from passive COTS dependency to AI-native operations across the enterprise value chain—from quote-to-cash and hire-to-retire to knowledge management, digital workplace, and service delivery. This transformation is built on a deliberate five-layer architecture: • Foundation Layer enabling model governance, agent discovery, security, and observability. • Integration & Data Layer connecting enterprise systems without requiring large-scale rebuilding. • Agent Intelligence Layer where AI Buddy acts as an orchestrator, routing intent across specialised agents through context-aware logic. • Domain Solutions Layer where agents handle areas such as service management, license optimisation, knowledge management, digital workplace, and cross-functional business processes. • Experience & Adoption Layer ensuring humans remain meaningfully in the loop through conversational interfaces, role-based design, and feedback loops that improve agent behaviour over time. Crucially, this transformation does not rely on a big-bang replacement strategy. We will share our “Prove then Scale” approach—starting with high-impact use cases that demonstrate measurable ROI, and then expanding the transformation across the enterprise. Attendees will leave with a practical framework for sequencing the journey, governing autonomous systems responsibly, and creating the organisational conditions where agentic AI can deliver not just efficiency, but genuine enterprise transformation. This session offers candid insights from Thoughtworks’ own journey—navigating complexity, making difficult architectural decisions, and learning what it takes to move from software-driven enterprises to AI-native operations.

Ms Sujana Arul Selvi ALAGURAJ

Product Leader, Thoughtworks

View Profile

STACK Conference 2026

Explore Our Exciting Programme Lineup!

[Workshop 01] Efficient Root Cause Analysis: How to Best Analyse Logs, Metrics and Traces

Mr Myrvin YAP

Mr Jason SOH

[Workshop 02] Temporal DevDay: Building "Invincible" Applications & Agentic Workflows

Mr Henry SURYAWIRAWAN

[Workshop 03] Designing Scalable and Rigorous Evaluations for AI Systems

Ms Leanne TAN

Mr Shaun KHOO

[Workshop 04] Agentic Patterns in Enterprise AI

Mr Samuel CHAN

[Workshop 05] Open Digital Platform (ODP) Discovery and Hands-On Enablement Workshop

Ms Celine CHIA

[Workshop 06] AI-Native Engineering: Supercharging Velocity via Intent-Driven Development

Mr Alex CHNG

Mr Leon LEOW

Mr Kelvin WIJAYA

Mr Sheikh Mohd SHAHEED

Mr Soedarsono

[Workshop 07] Identity & Agentic AI Workshop

Mr Takashi YAMAMOTO

[Workshop 08] Agentic AI Observability and Control: A Hands-On Bootcamp with TraceCtrl

Mr Matthias CHIN

[Workshop 09] Building a Context-Aware Agentic Data Platform

Mr Arunkumar JAMBULINGAM

[Workshop 10] Getting Started with Vibe Coding

Mr Michael CHENG

[Workshop 11] Less Plumbing, More Insight: Hands-On with Elastic Serverless Observability

Mr Sean MATTHEWS

Mr YAP Yi Sheng

[Workshop 12] Building an Agentic AI Workflow with Claude All the Way from Ideation to Production

Mr Michael ISVY

[Workshop 13] AI on the Front Line: Hands-On with AWS DevOps Agent and Security Agent

Ms Charis WONG

Mr Farid BAHARUDDIN

[Workshop 14] The Agent Development Lifecycle: Build a Safe Software Factory for Digital Services

Mr Omkar KHAIR

[Workshop 15] Working With LLMs as a Strange New Intelligence

Mr Uli HITZEL

[Workshop 16] From Chatbots to Agents: Architecting Autonomous Workflows with Tool-Use and Reasoning

Mr Jeswanth GUTTI

Mr KOAY Yong Zhuang

[Workshop 17] We are all Product Owners now: Spec-Driven Development with AI Agents

Mr CHAN Chi-Loong

Dr Raymond CHAN

Opening Address

Ms Jasmin LAU

From Digital Government to Intelligent Government: Building the Next Chapter Together

Mr GOH Wei Boon

A New Era of the Internet

Ms Michelle ZATLYN

Bring the Model to the Data

Mr Shay BANON

The Trusted and Sustainable Frontier: Balancing Resilience with Responsibility

Mr David FOO

Mr Florian TURSKY

Ms Michelle ZATLYN

Mr Shay BANON

Mr Wen HUANG

MAS AI-Powered Dev Platform: Scaling Velocity with GovTech Platform AI and SHIP-HATS

Mr Ron ZHONG

When Your User Is a Machine. Government Services By, For, and With AI

Mr James Theodore (Ted) KILIAN

Less Infrastructure, More Missions – Secure by Design & Simple by Default on One Hosting Platform with Many Missions

Mr Hudson LEE

Your Browser Can Run AI Now

Mr Trung VO

Resilience, Strategic Autonomy and Radical Transparency: The Critical Role of Open Source to Mitigate Geopolitical Uncertainty

Dr Chris BUTLER

Did Cybersecurity Change After Mythos and Project Glasswing?

Dr Kunal SEHGAL

From Idea to Done While Procurement Was Still Deliberating

Mr Michal POLANOWSKI

Mr Yitch CHENG

From Policy to Practice: Custom Guardrails for Your Context

Ms GOH Jia Yi

Embrace AI Native Cloud Now

Mr Paul MARTIN

Modernising Housing & Development Board (HDB)’s Analytical Data Stores