Home

Writing

Writing.

Write-ups of things I built and measured. The data and code behind them are linked from each one.

  1. Field study

    119 open-source repos, scored for coding agents

    I scored 119 widely used open-source repos with crag. In July 2026, 41 percent had no instruction file for coding agents, and all of them sat in the bottom two grades.

    Read it
  2. Benchmark, part two

    When self-hosting a model pays for itself

    Break-even math from real token logs. Cloud calls cost $1.24 to $4.92 per thousand; owning a 24 GB GPU costs about $60 a month and pays off from about 12,000 calls a month.

    Read it
  3. Benchmark

    Self-hosted and cloud vision models, 1,200 calls compared

    Six vision model setups on the same scenes. A self-hosted 30B model scored 6.05/9 against 6.60 for the best cloud model and answered in 0.77s against 3.6 to 7.8s.

    Read it
  4. Client work

    MetricHost, a game-server hosting platform

    How MetricHost works: Spring Boot services and Kafka on Kubernetes in the EU and US, idle game servers that sleep and wake in under a second, and self-provisioning machines.

    Read it