Notes from the build

A staged multi-model pipeline: 96% accuracy, 4x faster, new models swapped in an afternoon

A contract-intelligence SaaS rebuilt its single-model pipeline as four model roles behind a provider-agnostic routing layer. Extraction accuracy 82% → 96%, p95 latency 9s → 2.1s, model spend…