Independent thinking / practical systems

Making AI useful.

Based in Singapore
Working across systems

I build products to solve everyday problems. I write about agentic coding, AI agents, and knowledge that works.

Explore the writing

Selected writing

Notes from building.
Research worth sharing.

Full archive ↗
01

An Interactive Singapore Map, Built in One Conversation

A self-contained interactive map of Singapore - official planning boundaries, the full MRT network with an offline journey planner, statistics choropleths, georeferenced historical maps and walking tours - assembled from open data into one HTML file with an AI agent, no API keys. What the build produced, how it ran, and the token economics.

Read article
02

The Devin Issue Pilot

Five days of turning Devin into an issue-driven teammate: seven PRs merged after midnight, a merge rule rewritten four times, a second Devin that reviews before any PR exists, and a weekly quota burned to 86% on a mode nobody checked. The timeline, the key learnings, and how it works now.

Read article
03

How Devin Fusion Ran a Five-Day Company-Brain Cutover

A 688M-token Devin Fusion session - Claude Fable 5.1 leading, SWE-2 as sidekick - cut over our company brain in five days. Why the job was so token-heavy, and why the saving over a single frontier model is mostly input, not output. With a two-minute narrated video.

Read article
04

The Last Project You Build Alone

Why a second brain should be custom to its owner, independent of any one AI harness, and able to give every future agent durable context.

Read article
05

Two Brains, One Operator

Why personal and corporate agent memory must split, how each half is governed - server-enforced write fences for the company brain, a permissive universal adapter for the personal one - and how Qwen 3.8 Max and Qwen 3.8 Flash Next on two DGX Sparks planned the architecture and audited 1,837 pages of company knowledge.

Read paper
06

From Company Knowledge to Verified Changes

How Success Brain, shared abstractions, pattern linting, CodeGraph, and built-in verifiers work together to guide agentic engineering.

Read article

Areas of interest

Code. Knowledge.
People. Possibility.