writing
rssWriting.
Essays about the things I build, mostly about what breaks when AI systems meet real data.
- 5 min read
A Timer Is Not a Measurement
I made reads 59% faster, then found that most of my latency column was the cost of calling the clock. The two stopwatch calls cost more than the operation they were timing.
read → - 5 min read
The most dangerous AI error is the one nobody has to notice
AI does not become safe because it sounds confident. In a medical note, it has to show what it knows, what it cannot verify, and when a clinician needs to decide.
read → - 7 min read
A 200 Is Not a Read
An AI agent will tell you it read a page it never actually got. A 200 means the bytes arrived, not that you came away with the page. Here is the gate that closes that gap.
read → - 4 min read
Most of your critical vulnerabilities cannot be reached
A scanner tells you a vulnerable function exists in your dependencies. It does not tell you whether your code can actually reach it. ReachGate answers that second question with a deterministic verdict you can verify yourself.
read → - 5 min read
What Microsoft's Agent Governance Toolkit doesn't cover: data supply chain trust
AGT answers whether an agent is authorized to call a tool. There is a second question it does not answer: whether the agent should act given the current state of the data behind the decision.
read →