Field notes
Production failures leave better notes than successful demos. This archive records failure boundaries, evaluation decisions and engineering repairs that are safe to discuss publicly.
2026
Three months of silence in an IoT pipeline
A disabled trigger, an apparently healthy dashboard and the observability changes that followed.
7 min
Agent release gates
Six checks for tool use, evidence, permissions and recovery before an agent reaches live users.
6 min
A failure map for RAG systems
Retrieval, evidence and decision failures separated before anyone rewrites the prompt.
5 min
MLOps at the failure boundary
A review method that traces reproducibility, deployment, monitoring and recovery from the broken workflow.
5 min