DevSecOps and platform engineer, Singapore
I run the CI/CD platform behind 300 repositories and 50+ production pipelines on AWS and Kubernetes, and I build the AI agents that work inside it.
Since April 2024 at NCS I've owned the path to production for a regulated enterprise estate: the CI templates, the security gates and the releases that more than 50 pipelines depend on.
I'm also studying for a Master of Technology in Software Engineering at NUS-ISS. At home I run more than 30 self-hosted agents on a NAS, because the fastest way I learn a tool is to run it in production for myself.
I own the path to production for a 300-repository, 50+ pipeline estate: reusable GitLab CI templates, Terraform-managed AWS and zero-downtime EKS/ECS releases.
SonarQube, Fortify, OWASP Dependency-Check, Nexus IQ and Trivy run as pipeline gates, not review steps.
I instrumented all four DORA metrics. The data behind them drove a ~35% cut in mean time to deploy.
I built an air-gapped self-healing CI agent: an LLM diagnoses a failed job and opens a draft fix MR, which a person must approve.
Platform and DevOps Engineer, NCS GroupApr 2024 to now
Platform engineering for regulated enterprise accounts.
An LLM diagnoses a failed GitLab job behind an 80% confidence gate, then opens a draft merge request with the fix. It never merges by itself. It runs air-gapped on CPU-only runners with no external API calls.
Read the case studyI am using AI to automate recurring platform operations, starting with AMI refreshes, so patched machine images roll out without manual steps.
A Node.js and React dashboard that shows live DORA metrics for 50+ pipelines to the teams that own them, with AI-assisted failure analysis and a dependency risk radar.
A Backstage front door over Terraform and Terragrunt modules, run by GitLab CI. Claude picks templates only from the approved module library. Presented to NCS leadership.
Five public projects, each with the code on GitHub. Select a diagram to open it full size.
I can take a model to production with a release gate that training and CI both enforce.
The pipeline for a clinical triage model. One release gate (red-flag recall ≥ 0.95) is shared by training and CI, with an MLflow registry, DVC lineage, fairness and drift monitoring, canary promotion and drift-triggered retraining.
MLflow · DVC · Evidently · Fairlearn · GitLab CI
Read the code ↗
I write infrastructure as code with security and policy checks that block a bad change before it is applied.
Terraform and Terragrunt for the same system on AWS ECS Fargate. The GitLab pipeline runs tflint, Checkov, gitleaks and OPA policy gates, uses a push-only OIDC role, needs a manual apply and destroys itself on a schedule.
Terraform · Terragrunt · AWS ECS Fargate · OPA
Read the code ↗
I can ship a full AI system end to end: 16 Terraform modules and an 85-job DevSecOps pipeline.
The full system, built as an NUS-ISS team project: multi-agent triage (FastAPI and Next.js) on AWS ECS Fargate, shipped through a GitLab DevSecOps/MLSecOps pipeline with LLM red-teaming.
FastAPI · Next.js · Terraform · GitLab CI
Read the code ↗
I use AI to cut security triage noise, and I make it show evidence instead of trusting it.
An agent that makes every Semgrep and OWASP Dependency-Check finding prove itself with file and line evidence before anyone spends time triaging it.
Semgrep · Dependency-Check · Nemotron on Nebius
Read the code ↗
I build analysis that explains an outcome: which offensive and defensive plays decided the game.
Breaks down each team's offence and defence from the play-by-play (shot types, schemes, named sets and coverages) to work out why a team won or lost that game. It also rates all 30 teams, simulates the season 10,000 times and reviews every missed prediction.
Python · Elo · Monte Carlo · Claude CLI · Telegram
Read the code ↗Scheduled pipelines and Telegram bots that run in Docker on my home NAS.
Watches Garmin data for at-rest heart-rate episodes, runs safety checks and sends alerts, LLM coaching and doctor reports to the family on Telegram.
Python · Garmin Connect · Telegram · Docker
An agentic photography coach. Claude Code subagents critique each photo, cite master photographers, suggest Fujifilm recipes and track progress in an Obsidian vault.
Claude Code subagents · Telegram · Obsidian
Watches the Singapore property market. Official data and listings feed a code-only finance engine (upfront and monthly cost, 10-year bear, base and bull cases). The LLM only writes the words.
Python · official data · Telegram
Scrapes COE, used-car and EV prices every day, runs a cost-of-ownership engine and posts a report to Telegram.
Python · Playwright · Telegram
A weekly pipeline that scrapes, de-duplicates and ranks Singapore dining venues for each occasion, then publishes to Obsidian and Telegram.
Python · Docker · Obsidian · Telegram
A family grocery price bot for Singapore: a daily best-value report on Telegram and a web page with product photos.
Python · Telegram · web page
Finds one recipe a day by web search. Each recipe must pass a strict gate (time, cost, live source link, no repeats) before it is posted to Telegram.
Claude CLI · validation gate · Telegram
A private Telegram gym coach. Claude Code writes the weekly plans, every plan passes a shoulder-injury check, and Garmin recovery data shapes the day.
Claude Code · Garmin · Telegram
A Singapore Pools TOTO bot: the latest result, what the next draws will do and where the next big prize is, sent to Telegram and an Obsidian vault.
Python · Docker · Telegram · Obsidian
A local-first World Cup assistant: match predictions anchored on a Poisson model and explained by an LLM, with live scores, standings and news through a bilingual Telegram bot.
Poisson model · Claude CLI / Ollama · Telegram
A Singapore luxury watch market bot: reference market values, deals after every cost, paper trading and a daily lesson. It fails safe when a source's terms are unconfirmed.
Python · Telegram
An hourly Telegram bot that teaches the Fujifilm X100VI: film recipes, composition, and Singapore light and weather.
Python · TelegramSelf-hosted quiz bots I built to study for my own exams. Each runs in one container.
Practice questions for SAP-C03. On a wrong answer Claude draws the architecture with AWS icons. A readiness report arrives every week.
Telegram · Claude tutor · Docker
A study bot for the CCAR-P exam with a Claude tutor, official-icon architecture diagrams and a weekly readiness report.
Telegram · Claude tutor · Docker
A study bot and question bank for SAA-C03 built on open-licensed community questions, with a Claude tutor and AWS-icon diagrams.
Claude tutor · Docker
Practice questions, Claude tutor diagrams, study cards and a weekly readiness report for AIP-C01.
Telegram · Claude tutor · Docker
A task-based CKA bot for bash beginners. You type the commands, Claude grades them and explains every flag.
Telegram · Claude grader · Docker
Practice questions, flashcard study cards, Claude tutor diagrams and a weekly readiness report for the HashiCorp Terraform Associate (004) exam.
Telegram · Claude tutor · Docker
Practice questions, pipeline diagrams drawn by the Claude tutor and a weekly readiness report for the GitLab Certified CI/CD Associate exam.
Telegram · Claude tutor · DockerThe tools I use most, and the certifications and study behind them. The full list is in my resume.
I shoot in colour, mostly on a Fujifilm X100VI. These are from trips to Sichuan, Chongqing and Taiwan, and from the 2026 airshow.
