My agent's memory disagreed with itself
My agent published a post nobody had reviewed, following a note I wrote and contradicted twice afterwards. Everything else in this repo is enforced by code. Memory was the one layer that only advises.
Engineering, measured
I build real-time distributed systems using AI pipelines. I write about what worked, what didn't, and the numbers behind both.
Throughput doubled. Change-failure rate doubled with it. Both are in the posts below, with the caveats and the counts behind them.
Latest posts
-
-
Design intent is the antidote to a narrow AI fix
An agent optimises the ticket in front of it and cannot see the fix fighting the wider architecture. Design intent is the only thing that rejects a narrow local fix, and where you put it decides whether it works.
-
Make supersession machine-readable
The last post blamed our architecture records for confusing the agent. That was the wrong way round. The agent was reading an ambiguity we had written in ourselves, and the reason we wrote it in was that our review rules were never in the repository.
Latest project
job-crm: an agent that searched for my next job, and found it
2026 · TypeScript · React · Claude Code
A local job-search CRM with agents at every step. It checks about 1,630 postings a morning and hands me one page of decisions.
Currently interested in
🤖 Artificial intelligence
What changes when an agent becomes the daily driver. Throughput, cycle time and change-failure rate, with their caveats.
📺 Real-time streaming platforms
Real-time quality analysis of video streams. Event-time semantics, the consistency boundary, the bugs that only show up in production.
⚖️ Engineering judgment
Build versus buy, instrument before you roll out, when to defer. The decisions that never make it into a postmortem but should.
📐 Architecture & infrastructure
Terraform at a startup, schema evolution, and the database choice you regret six months later.
Did your AI tooling actually change how your team delivers?
Most teams have adoption numbers, seats bought and suggestions accepted, and nothing on whether anything shipped sooner or broke less. I answer that from your own Git and CI data. Read-only, five days, fixed price.