I build cloud systems and write about them.

I wanted to understand what actually sits behind a frontier model, so I built an interactive picture of the one rack it would take to serve one, and worked out the numbers as best I could.

With data center costs all over the news, I tried to work out what the rack behind one copy of a frontier model costs to buy and run, part by part, down to a rough cost per million tokens.

The costliest part of a forgotten cloud resource isn't the bill. It's the hunt for who owns it.

Finding AWS waste is the easy part. Acting on what you found is where teams get stuck.

More cloud data doesn't help without guidance, which is why opinionated recommendations beat dashboards.

Built a low-latency voice agent using Pipecat and AWS Nova Sonic. Great voice quality, painful observability. Here’s what worked, what didn’t, and what’s next.

I built Git AI to automate commit messages and PR descriptions using AI — no subscriptions, no bloat. It’s an open-source tool that saves time and keeps your Git workflow efficient and affordable.

I built an AI chatbot using FastMCP, odoo-rpc-client, and Claude 4 to easily chat with my Odoo ERP data. This is how I did it.

Timesheet management sucks. I built cal2ts to fix it. This AI-powered web app turns calendar events into timesheet entries with one click, letting our Cloud303 engineers focus on building great software instead of fighting with Odoo.

I looked at a few different ways to build my blog, but ended up going with AWS Amplify, Contentful, and Next.js. Here’s why this setup made the most sense for what I needed.