SmallerModels

Case studies

Smaller models work on real workloads.

We search for the requests that a cheaper, device-capable model can handle, freeze the routing policy, and evaluate it on unseen traffic.

Published case studies

Customer-service drafting

Open-ended text generation

Draft policy-compliant support replies from ticket context using a frozen three-branch router.

Read the case study →
coverage
23.1% 95% error bound: ± 0.8 pp
precision
95.31% 95% error bound: ± 0.9 pp

Restaurant Text → JSON

Structured extraction

Convert natural-language restaurant searches into exact JSON with a frozen small-model router.

Read the case study →
coverage
17.1% 95% error bound: ± 0.8 pp
precision
95.6% 95% error bound: ± 1.1 pp