Logbook · entries
Articles
Quantified field notes on autonomous agents, LLMs served locally, observability and infrastructure.
Predicting without training
CatBoost against Kumo-Tabular on Seattle building energy consumption: a tabular foundation model seems to do better without training, for 25 seconds per call. Four experiments to check, and what that says about use cases, cost, and what's left to the craft.
16 min read · Machine Learning · Tabular foundation models · MLOps · CatBoost · Deployment
Nine agents, nine days
A quantified write-up of a fleet of nine autonomous SRE agents in real conditions: what held up, what broke, what trace analysis taught me, and what I'm keeping.
9 min read · Autonomous agents · SRE · Local LLM · Observability · Security