Reliability / 6 minute read
Evaluation gates as release conditions
An eval that only produces a dashboard number changes nothing. Treated as a release condition, it becomes the thing that keeps a bad output away from a user.
Read the articleWriting
Notes on making AI systems dependable.
Short essays on the parts of this work that decide whether a system survives contact with production: gates, failover and operational memory.
Architecture / 5 minute read
Designing for provider failover
Treating model vendors as interchangeable capacity behind one interface turns a vendor incident into a routing decision instead of an outage.
Read the articleOperations / 5 minute read
Operational memory for agent systems
Every incident is either a lesson written into the system or knowledge that leaves with whoever handled it. Only one of those compounds.
Read the article