Agent Wars, Episode 3: Revenge of the Traces
On observability as the layer agentic AI deployments may miss, and it's importance in deciding whether you can trust anything else you build
How we build AI for public good
A technical blog for AI science and engineering
Leading AI Engineering in Government
Reflections on building reusable capabilities, empowering teams and staying relevant in a fast-moving field
Strengthening Chatbot Safeguards with Red-Teaming Agents
How we used agentic AI to proactively test and strengthen chatbot safeguards for chat.gov.sg
Decision Models for Guardrails: Exploring Jev and Kev for Moderation
A closer look at the new paradigm of frontier models, and how we explored Jev and adapted its open-source interpretation - Kev, for Singapore-context moderation.
Behind VEILED Expressions: A Singapore-context benchmark for mental-health AI evaluation
We built VEILED-SH to evaluate indirect self-harm/suicide language in the Singapore context, shaped by local language, privacy safeguards and practitioner insight.
Building a Serverless GenAI Platform on AWS
Balancing real-time performance, model cost and structured multi-agent workflows
Building a Smart and Safe Whole-of-Government Chatbot for Singaporeans
How we build custom policy guardrails and balance safety with helpfulness for chat.gov.sg
Stewarding What Was Entrusted
No Forward-Deployed Team really starts from zero, every achievement begins with a foundation laid by someone before us. Looking back on my journey, I realise what I inherited matters just as much as what I've accomplished.
When Agent Skills Evolve: The Need for a Layer of Control
What we learnt from building Rhizomatic, a human-in-the-loop proof of concept for reviewing, versioning and rolling back changes to reusable agent skills.
Guardrails in the Wild: Closing the Retraining Loop for LionGuard 2
Operationalising MLOps with production feedback to continuously evaluate, retrain, and improve our guardrails.
What is a Mixture-of-Experts model, really?
An experiment reveals how different layers shape what a model understands, retrieves and ultimately generates.
One Year in Multimodal AI at GovTech
“I came here expecting to build models for people who already knew what they wanted. I discovered that the harder work is figuring out what should be built, who it is for, and whether it solves a real problem.”