PORTFOLIO
From business problems
to working systems.
Duo · Full-stack engineer / design and delivery of AI applicationsFive projects across education technology, customer operations, content production and field service. Each was scoped, architected, built and tested end to end by me, and is now in use, under store review or used in tender presentations.
Featured project
Study Companion · iPad application
In production useA closed-loop learning system for Singapore Primary 5 science, maths, English and Chinese. The student answers on the iPad, the parent reviews each marking decision, and the system schedules the next review for every topic while building a record of error causes and mastery.
Forgetting curve: the effect of a single attempt, quantified
Scheduling uses FSRS, the spaced-repetition algorithm Anki ships by default, rather than a hand-tuned interval ladder. The dashed line projects memory decay without this attempt, the solid line projects it afterwards, and the vertical markers show the review date before and after.
Example: “pollination” reappeared after 19 days and scored half marks, so the next review moved from 2 days out to 10. The parent can override this in one step — retest tomorrow, or reset after re-teaching.
- Mastery is derived only by deterministic code from confirmed marking; the language model never writes it directly.
- Lost marks are attributed to four causes — missing concept, incomplete answer, exam technique, execution slip — each mapped to its own review strategy.
- The shared question bank is read-only for every user; an edit creates a private copy, leaving the original and its grading history intact.
Screenshots were taken in the iPad simulator against a separate demo dataset; student details, papers and answers are fictitious.
Other projects
Student & customer operations platform · tender prototype
15 business modules and 6 role-based workspaces, in English and Chinese. The AI inbox identifies and files customer information while the conversation is under way; in a complaint, once sentiment crosses a threshold, all sales prompting stops and the case passes to a manager. Open questions are answered live by a language model, and every reply cites the knowledge entries and procedure it relied on, so it can be audited.
Workflows are fixed as orchestrated SOPs built from seven node types — trigger, knowledge, system action, decision, person, wait and message; all six gym scenarios can be run step by step, with configuration and review-to-publish shown end to end. A companion Python analytics engine, run on about 250,000 rows of simulated raw data from 7 branches, validates per-channel conversion, retention alerts, campaign-link attribution and multi-source data intake.
PSLE NotebookLM Helper · Chrome extension
Works with Google NotebookLM to turn study material and a student’s actual answers into review priorities, achievable daily tasks and matching practice.
Content Radar · research and publishing platform
Collects Singapore education policy from primary sources, de-duplicates and verifies it, then drives topic selection, illustrated content and publishing. This site’s public posts and pathway planner are produced by the platform.
Maintenance work-order system
Covers the full cycle from intake and dispatch to on-site work and sign-off, with multi-criteria filters, a per-order timeline and a statistics dashboard.
Request an interactive demo
Interactive demos of the tender prototype and the work-order system are available, with a dedicated link issued to each visitor. Please get in touch by email or social media.