Banking & Payments Technology Leader · AI Strategy, Agentic Systems & Governance

For 20+ years I've built technology inside high-stakes financial systems — credit cards, commercial payments, tokenization, clearing, disputes. Systems where a mistake moves real money, so every action needs identity, authorization, evidence, and audit.

Now I'm applying that discipline to a new question:

How do we put AI and technology in front of students and classrooms without giving up privacy, judgement, or trust?

Payments taught me that trust is engineered, not assumed. AI and educational tools are no different.

An open, verifiable pathway engine for Singapore's education system — covering PSLE $\to$ Secondary, O-Level / SEC $\to$ Post-Secondary (JC / MI / Poly), and A-Level $\to$ University (70-point UAS).

• Rules as configuration, not hard-coded logic: Policy changes (2027 O-Level withdrawal, 2028 L1R4 $\le$ 16) update via versioned YAML, not code rewrites.
• Dignified backward guidance: Guarantees at least 3 distinct routes (including non-direct pathways) rather than a single high-stakes cutoff.
• Evidence vs Fit separation: Historical entry ranges are never blended into subjective fit scores. Fit is 0–100, fully transparent, and explainable.
• Privacy by construction: 100% on-device execution, zero telemetry, zero analytics, with single-file desktop binaries for Windows, macOS, and Linux.

MIT · 305 automated tests · 37 golden cross-engine fixtures · 815 cited data facts with automated freshness gates

An open-source AI marking assistant for Singapore exam essays — PSLE composition and A-Level General Paper.

• Rubric-grounded feedback: Evaluates against published band descriptors, underlines errors with concrete fixes, and provides tracked-changes rewrites.
• Private by default: Runs completely local (Ollama / local models), ensuring student essays never leave the machine.
• Teacher-first: Gives educators their evenings back while providing students with patient, personalized practice.

MIT · Local models to frontier LLMs · Ships with its own evaluation datasets

• A prompt is not permission: Production systems need clear policy, boundaries, evidence, and auditability.
• Evals over demos: Every project ships with its own test suites and golden evaluation fixtures. If you can't measure it, you're just marketing it.
• Local-first & privacy by construction: Student data and personal decisions should stay on the user's machine.
• Augment the human, never replace judgement: Technology handles the tireless verification and synthesis; human educators and families make the human decisions.

📍 Singapore · Views my own, not my employer's.