Blog
- When LLMs Are Forced to Be Wrong: A Case Study with Sonnet and Opus
A practical case study showing how LLMs fail under ambiguity and missing signals when forced to produce an answer, even with structured constraints and careful prompting.
- Claude Code Started Calling Haiku After a Version Update
A Claude Code update introduced unexpected Haiku calls alongside Opus. This post documents the discovery and version comparison.
- Why My Claude Prompt Cache Hit Rate Was Stuck at ~50%
Git workspace context silently changed my Claude inputs, causing prompt caching to fail and doubling evaluation costs.
- Visitor Pattern Strikes Back in Java
Why the Visitor pattern is becoming relevant again in modern Java with sealed classes and pattern matching.
- Hello, World
Starting a personal engineering blog and why writing matters.