The product decisions behind dynamic results, focused tools, evals, and one agent loop.

99.4% of Claude Code tokens are input tokens. These 5 open-source repos cut tool-layer waste by 60–98%, saving 25–45% per session.

The terminal-first setup I actually work in day to day: tmux for sessions, Neovim for editing, and Claude Code wired into both.

What separates a portfolio that gets read from one that gets skipped — framing projects around the decision they informed, not the model you trained.

A data-driven look at the trade-offs — price, commute, and long-term value — behind a decision most Singaporeans make on instinct.

SRE from a developer's angle: SLOs, error budgets, and why treating operations as a software problem changes how you build.

How to spread inference traffic across ML services without melting a node — a practical approach to load balancing model workloads.

End-to-end: packaging a Faster-R-CNN object-detection model and getting it serving reliably on Kubernetes.