Sunil
Sivadas
Ph.D.

Testing makes the invisible visible

At the IMDA AIVF AI Assurance Sandbox Showcase at ATxSummit, Dr Martin Saerbeck and I shared our case study on building a reusable assurance framework around a live AI career guidance chatbot, one that serves graduates, career switchers, and retrenched workers across Singapore. Our approach began with the people the system serves and the harms they face: tone with vulnerable users, safeguards for minors, parity across gender and age.

Wan Sie Lee, Executive Director of the AI Verify Foundation, framed the whole conversation in one line. Prof. Russell added the horizon: as systems approach superhuman levels, real-world testing reaches its limit. Mathematical, analytical guarantees will be needed. The systems deployed today are a different case, they already shape real decisions. Careers. Money. Health.

Testing makes the invisible visible. For the systems shaping real decisions today, that is how we earn trust.

Visualising Agentic AI as a Triangle, not a straight line

Two of the most-cited studies pull in opposite directions: METR's "agents are doubling in capability" and MIT's "95% of GenAI projects fail." Looking at only one leaves you either overestimating what you can deploy, or underestimating what's possible. The missing lens is the tasks corner, GDP-Eval-style evidence showing how frontier models already perform on the work we actually pay people to do.

Capabilities tell you what's possible. Tasks tell you where value appears first. Adoption tells you how hard it is to ship.

The Power of Possibilities with MERaLiON

Panellist at "Imagine AI – The Power of Possibilities with MERaLiON" at SWITCH 2025. MERaLiON is Southeast Asia's empathetic multimodal LLM, designed for the region's diverse languages, cultures, and communication styles. We explored how multimodal LLMs reduce solution complexity, cut response latency, and how hyper-localising AI is the foundation for meaningful adoption and trust, not a side concern.

At NCS, we evaluate MERaLiON not just in isolation but as part of a larger AI stack, integrating with internal workflows, tuning for performance, measuring deployment cost and resilience.

AI is changing tasks before it changes jobs

Most companies are investing in the wrong place. The gap between adoption and impact is the biggest risk. In his Budget 2026 speech, PM Wong announced a National AI Council he'll personally chair and named the real challenge: many firms say they are using AI, but end-to-end transformation requires organising data, rebuilding systems, redesigning processes and jobs, and retraining workers.

That gap, between adoption and impact, is what I addressed in my AI at Work masterclass. 200+ use cases surfaced through NCS Group's SunshineAI hackathons, and a knowledge-sharing culture that turns individual discoveries into organisational learning.