Working notes on building Spine. The product, the engineering, and the parts where the abstraction leaks. RSS.
Why we replaced an LLM round-trip with a locally hosted GLiNER2 model, what we measured, and what we learned about being honest with benchmarks.
Spine routes every Claude call to the smallest model that can do the job. Here is the routing table, the reasoning, and the cost numbers.
A literary critic's mental model and a senior engineer's mental model are the same shape. Spine makes that shape visible.