Checkmate SG
Contributing to Singapore's national misinformation detection system — as a safety analyst and, increasingly, on the product side. Over 5,000 users and growing.
// the problem
Misinformation spreads faster than it can be manually verified. In Singapore's multilingual environment, the challenge compounds — content circulates across English, Mandarin, Malay, and Tamil simultaneously. The system that catches it has to be part-AI, part-human, and trustworthy enough that thousands of active users rely on it.
// why this matters
Misinformation doesn't just spread — it entrenches. A false belief accepted as true changes how people vote, who they trust, and what they share. In Singapore's multilingual environment, the window to intercept a false narrative is narrow. The platform is about closing that window.
// how it works
As a volunteer safety analyst, I triage and investigate user-reported content — suspicious messages, screenshots, images — classifying for misinformation, scams, and harmful claims. I evaluate the AI model's verdict for each submission, voting to confirm or correct it, and contribute structured observations that feed back into improving the system's detection logic over time.
Beyond the analyst role, I've proposed and am scoping four product initiatives for the team:
A website and public content strategy — SEO and GEO audit of checkmate.sg, a "Scam Trends" public page that turns the platform's existing data into a regularly updated, shareable resource, and a content calendar tied to proactive scam alerts.
An AI evaluation framework — a labeled dataset drawn from checker-voted submissions and a set of core performance metrics (accuracy vs checker consensus, false positive rate, language accuracy), so model and prompt changes can be scored offline before deployment rather than tested in production.
An internal data dashboard — a Looker Studio dashboard connected to BigQuery, giving the whole team live visibility into checker activity, AI accuracy rates, scam category trends, and volunteer engagement.
A proactive alerts PRD — owning the product spec for a WhatsApp-based feature that warns users about trending scam patterns before they're asked, moving Checkmate from reactive to proactive.
// trade-offs i made
Full automation would be faster. It would also be wrong often enough to erode the trust that makes the platform valuable. The hybrid model — AI classifies at scale, human analysts provide the judgment layer — is slower and resource-intensive, but right for a product where a false positive or false negative has real consequences for real people.
// what i learned
The eval framework matters more than any individual feature. You can't safely improve a model you can't measure offline. Building the scoring infrastructure first is the high-leverage move — everything else becomes faster and safer once it exists.
// what's next
The eval framework is the highest-leverage next step. You can't safely iterate on a model you can't measure offline. Everything else — alerts, better coverage, new languages — becomes faster and safer once you have a scoring system you trust.
// trust & safety
Trust is the entire product. A misinformation detection tool that gets things wrong — or that users can't trust to get things right — is worse than no tool at all. Every design decision at Checkmate reflects this.
The hybrid architecture — AI classifies at scale, human analysts provide the judgment layer — is a deliberate safety choice. Full automation would be faster. It would also be wrong often enough to erode the trust that makes the platform valuable. The human layer exists not because the AI can't be trusted at all, but because the stakes of a false positive (flagging something true as misinformation) or a false negative (missing a harmful claim) are high enough to require human accountability in the loop.
The three-state null model in the horizon scanning tool — distinguishing hallucinated confidence, structural failure, and true null — is a specific response to the trust problem. A tool that silently fails, or that fabricates a clean bill of health, trains its users to ignore it. Making the null state auditable is a foundational reliability commitment.
The AI evaluation framework I'm building for the team is the long-term trust investment: a way to score model and prompt changes offline before they reach users, so reliability improves systematically rather than by trial and error in production.