Engineering Blog

Notes from the reliability layer

Playbooks, post-mortems, and engineering insights on structured data synthesis and workflow execution with small language models.

Why small models fail at structured output

An in-depth analysis of why sub-7B models struggle with strict JSON schemas, and how to fix it with templates, JSON repair, and validation loops.

A comprehensive playbook for deploying llama.cpp, Ollama, and Kern in production and regulated environments.

Why single-prompt metrics fail to represent agent performance, and how to measure multi-step system success.

Why reliability is the missing layer in the modern AI stack, and why application code shouldn't handle non-deterministic errors.

Design patterns for tracing non-deterministic agent workflows, and the metrics SREs should monitor.