Engineering Notes
Short, opinionated ideas on building reliable software in an AI-native world.
- testingtooling
A Test Platform Should Not Become Another Dependency
A quality platform succeeds when teams can use it independently, not when it becomes a gatekeeper for every change.
- llmevaluation
Three Rules for LLM-as-a-Judge
A model can help evaluate an AI system, but it cannot turn vague criteria into trustworthy evidence.
- agentic-aievaluation
When I Would Not Use an Agent
An agent is a poor default for deterministic, high-risk, or poorly observable work.
- mcpopenapi
Why MCP Does Not Replace APIs
MCP makes APIs easier for agents to use; it does not replace the contracts and operational disciplines behind them.