Writing
Notes from the build.
Things I worked out the slow way — evaluation harnesses, agent design, and the unglamorous engineering that decides whether a demo ever becomes a product.
Latest · 2026.08.02
Your agent doesn't need more tools. It needs a scoreboard.
Every stuck agent project I've seen was fixed by measurement, not capability. Here's the harness I build first, why it's boring on purpose, and what it costs to skip it.
Read it-
2026.08.02
Your agent doesn't need more tools. It needs a scoreboard.
9 minWhy the evaluation harness comes before the tool belt, and what a good one actually contains.
-
2026.07.14
Structured output is a contract, not a formatting trick
6 minTreating the schema as an interface between two systems changes how you write it — and how you handle the day it isn't honoured.
-
2026.06.21
Ten thousand lines of vanilla: rebuilding this site with no build step
7 minWhat you gain when there's nothing between the file you edit and the file the browser gets, and the three places it genuinely hurts.
-
2026.05.30
Three questions before you give an agent another tool
5 minMost tool belts grow by accident. A short filter that keeps the surface area honest and the traces readable.
-
2026.04.18
Azure ML pipelines that survive a handover
8 minNaming, versioning and the documentation that means the next engineer doesn't rebuild it from scratch.
Try another tag, or
New posts land roughly monthly. Tell me what you'd like read next — I write about whatever I've just had to solve.
Got a project in mind?
Let's build something together.
I take on freelance AI and full-stack work alongside my day job, and I'm open to founding opportunities. Tell me what you're building and I'll come back to you within a day.
Start a conversation